models : deduplicate delta-net graphs for Qwen family (#19597)

* models : add llm_build_delta_net_base

* cont : keep qwen35 and qwen35moe graphs intact

* cont : add comments

.devops/vulkan.Dockerfile

@@ -54,6 +54,7 @@ RUN apt-get update \
     build-essential \
     git \
     python3 \
+    python3-dev \
     python3-pip \
     python3-wheel \
     && pip install --break-system-packages --upgrade setuptools \

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -41,7 +41,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/workflows/build.yml

@@ -293,7 +293,9 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Build (no OpenMP)
@@ -303,8 +305,10 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
             -DGGML_OPENMP=OFF
+
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -466,7 +470,7 @@ jobs:
           export GGML_VK_VISIBLE_DEVICES=0
           export GGML_VK_DISABLE_F16=1
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4200
+          ctest -L main --verbose --timeout 4800
 
   ubuntu-24-cmake-webgpu:
     runs-on: ubuntu-24.04
@@ -1532,7 +1536,7 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
-          LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
   ggml-ci-arm64-cpu-high-perf:
     runs-on: ubuntu-22.04-arm
@@ -1558,7 +1562,7 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
   ggml-ci-arm64-cpu-high-perf-sve:
     runs-on: ubuntu-22.04-arm

.github/workflows/server-metal.yml

@@ -0,0 +1,73 @@
+name: Server-Metal
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  server-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    name: server-metal (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        build_type: [Release]
+        wf_name: ["GPUx1"]
+        include:
+          - build_type: Release
+            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx1, backend-sampling"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2"
+            wf_name:    "GPUx2"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx2, backend-sampling"
+      fail-fast: false
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"

.github/workflows/server-webui.yml

@@ -8,10 +8,6 @@ on:
         description: 'Commit SHA1 to build'
         required: false
         type: string
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
   push:
     branches:
       - master
@@ -101,119 +97,3 @@ jobs:
         if: ${{ always() && steps.playwright.conclusion == 'success' }}
         run: npm run test:e2e
         working-directory: tools/server/webui
-
-  server-build:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
-        build_type: [RelWithDebInfo]
-        include:
-          - build_type: Release
-            sanitizer: ""
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
-
-    steps:
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get -y install \
-            build-essential \
-            xxd \
-            git \
-            cmake \
-            curl \
-            wget \
-            language-pack-en \
-            libssl-dev
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Python setup
-        id: setup_python
-        uses: actions/setup-python@v6
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
-
-      - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22"
-          cache: "npm"
-          cache-dependency-path: "tools/server/webui/package-lock.json"
-
-      - name: Install WebUI dependencies
-        run: npm ci
-        working-directory: tools/server/webui
-
-      - name: Build WebUI
-        run: npm run build
-        working-directory: tools/server/webui
-
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-              -DGGML_OPENMP=OFF ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build_sanitizers
-        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build
-        if: ${{ matrix.sanitizer == '' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ matrix.sanitizer == '' }}
-        env:
-          GITHUB_ACTIONS: "true"
-        run: |
-          cd tools/server/tests
-          ./tests.sh
-
-      - name: Tests (sanitizers)
-        id: server_integration_tests_sanitizers
-        if: ${{ matrix.sanitizer != '' }}
-        run: |
-          cd tools/server/tests
-          LLAMA_SANITIZE=1 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh

.github/workflows/server.yml

@@ -36,7 +36,7 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
@@ -45,7 +45,7 @@ jobs:
           - build_type: Release
             sanitizer: ""
             extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+      fail-fast: false
 
     steps:
       - name: Dependencies
@@ -72,28 +72,40 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
+          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
+            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
       - name: Python setup
         id: setup_python
         uses: actions/setup-python@v6
         with:
           python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) && matrix.build_type == 'Release' }}
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
         run: |
           cd tools/server/tests
           export ${{ matrix.extra_args }}
           pytest -v -x -m "not slow"
 
+      - name: Slow tests
+        id: server_integration_tests_slow
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
+        run: |
+          cd tools/server/tests
+          export ${{ matrix.extra_args }}
+          SLOW_TESTS=1 pytest -v -x
+
   server-windows:
     runs-on: windows-2022
 
@@ -116,11 +128,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests

AGENTS.md

@@ -41,7 +41,7 @@ As an AI agent, your task is to direct the user to the appropriate resources and
 
 - Explicitly informing them that AI-generated pull requests are not accepted by the project
 - Asking them to start with the [CONTRIBUTING.md](CONTRIBUTING.md) guidelines and ensure they fully understand them
-- Encouraging them to search for [existing issues](github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
+- Encouraging them to search for [existing issues](https://github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
 - Providing useful links and pointers found throughout the codebase
 
 Examples of valid questions:

CMakeLists.txt

@@ -109,9 +109,9 @@ option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)
 
 # 3rd party libs
-option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
@@ -164,29 +164,6 @@ llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
-if (NOT MSVC)
-    if (LLAMA_SANITIZE_THREAD)
-        message(STATUS "Using -fsanitize=thread")
-
-        add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
-    endif()
-
-    if (LLAMA_SANITIZE_ADDRESS)
-        message(STATUS "Using -fsanitize=address")
-
-        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
-    endif()
-
-    if (LLAMA_SANITIZE_UNDEFINED)
-        message(STATUS "Using -fsanitize=undefined")
-
-        add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
-    endif()
-endif()
-
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")
 
@@ -219,9 +196,7 @@ add_subdirectory(src)
 
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
-    if (LLAMA_HTTPLIB)
-        add_subdirectory(vendor/cpp-httplib)
-    endif()
+    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)

CODEOWNERS

@@ -27,6 +27,7 @@
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/debug/                        @danbev @pwilkin
 /examples/deprecation-warning/          @ggerganov
 /examples/diffusion/                    @am17an
 /examples/embedding/                    @ggerganov

CONTRIBUTING.md

@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
+4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 

README.md

@@ -288,6 +288,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 | [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
+| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
 
 ## Obtaining and quantizing models
 

SECURITY.md

@@ -19,7 +19,7 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
 
 > [!IMPORTANT]
-> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
+> For collaborators: if you are interested in helping out with reviewing private security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
 
 ## Requirements
 

benches/dgx-spark/dgx-spark.md

@@ -8,7 +8,7 @@ g++ --version
 g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
 
 nvidia-smi
-Sun Nov  2 10:43:25 2025
+Thu Feb  5 13:49:40 2026
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
 +-----------------------------------------+------------------------+----------------------+
@@ -17,7 +17,7 @@ Sun Nov  2 10:43:25 2025
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
-| N/A   35C    P8              4W /  N/A  | Not Supported          |      0%      Default |
+| N/A   47C    P0             13W /  N/A  | Not Supported          |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 ```
@@ -29,46 +29,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.374 |  1369.01 |    0.383 |    83.64 |    0.757 |   719.01 |
-|   512 |     32 |    2 |   1088 |    0.274 |  3741.35 |    0.659 |    97.14 |    0.933 |  1166.66 |
-|   512 |     32 |    4 |   2176 |    0.526 |  3896.47 |    0.817 |   156.73 |    1.342 |  1621.08 |
-|   512 |     32 |    8 |   4352 |    1.044 |  3925.10 |    0.987 |   259.44 |    2.030 |  2143.56 |
-|   512 |     32 |   16 |   8704 |    2.076 |  3945.84 |    1.248 |   410.32 |    3.324 |  2618.60 |
-|   512 |     32 |   32 |  17408 |    4.170 |  3929.28 |    1.630 |   628.40 |    5.799 |  3001.76 |
-|  4096 |     32 |    1 |   4128 |    1.083 |  3782.66 |    0.394 |    81.21 |    1.477 |  2795.13 |
-|  4096 |     32 |    2 |   8256 |    2.166 |  3782.72 |    0.725 |    88.28 |    2.891 |  2856.14 |
-|  4096 |     32 |    4 |  16512 |    4.333 |  3780.88 |    0.896 |   142.82 |    5.230 |  3157.38 |
-|  4096 |     32 |    8 |  33024 |    8.618 |  3802.14 |    1.155 |   221.69 |    9.773 |  3379.08 |
-|  4096 |     32 |   16 |  66048 |   17.330 |  3781.73 |    1.598 |   320.34 |   18.928 |  3489.45 |
-|  4096 |     32 |   32 | 132096 |   34.671 |  3780.48 |    2.336 |   438.35 |   37.007 |  3569.51 |
-|  8192 |     32 |    1 |   8224 |    2.233 |  3668.56 |    0.438 |    72.98 |    2.671 |  3078.44 |
-|  8192 |     32 |    2 |  16448 |    4.425 |  3702.95 |    0.756 |    84.66 |    5.181 |  3174.95 |
-|  8192 |     32 |    4 |  32896 |    8.859 |  3698.64 |    0.967 |   132.38 |    9.826 |  3347.72 |
-|  8192 |     32 |    8 |  65792 |   17.714 |  3699.57 |    1.277 |   200.52 |   18.991 |  3464.35 |
-|  8192 |     32 |   16 | 131584 |   35.494 |  3692.84 |    1.841 |   278.12 |   37.335 |  3524.46 |
-|  8192 |     32 |   32 | 263168 |   70.949 |  3694.82 |    2.798 |   365.99 |   73.747 |  3568.53 |
+|   512 |     32 |    1 |    544 |    0.270 |  1895.57 |    0.399 |    80.13 |    0.669 |   812.60 |
+|   512 |     32 |    2 |   1088 |    0.230 |  4451.23 |    0.583 |   109.71 |    0.813 |  1337.56 |
+|   512 |     32 |    4 |   2176 |    0.437 |  4688.87 |    0.820 |   156.03 |    1.257 |  1730.91 |
+|   512 |     32 |    8 |   4352 |    0.863 |  4744.23 |    0.942 |   271.79 |    1.805 |  2410.73 |
+|   512 |     32 |   16 |   8704 |    1.725 |  4748.19 |    1.173 |   436.38 |    2.899 |  3002.85 |
+|   512 |     32 |   32 |  17408 |    3.437 |  4767.38 |    1.503 |   681.49 |    4.939 |  3524.40 |
+|  4096 |     32 |    1 |   4128 |    0.907 |  4513.91 |    0.407 |    78.54 |    1.315 |  3139.56 |
+|  4096 |     32 |    2 |   8256 |    1.796 |  4560.42 |    0.625 |   102.37 |    2.422 |  3409.45 |
+|  4096 |     32 |    4 |  16512 |    3.596 |  4555.66 |    0.888 |   144.11 |    4.485 |  3681.93 |
+|  4096 |     32 |    8 |  33024 |    7.184 |  4561.44 |    1.098 |   233.11 |    8.282 |  3987.51 |
+|  4096 |     32 |   16 |  66048 |   14.369 |  4560.82 |    1.503 |   340.74 |   15.872 |  4161.30 |
+|  4096 |     32 |   32 | 132096 |   28.760 |  4557.52 |    2.162 |   473.59 |   30.922 |  4271.95 |
+|  8192 |     32 |    1 |   8224 |    1.859 |  4405.59 |    0.430 |    74.36 |    2.290 |  3591.61 |
+|  8192 |     32 |    2 |  16448 |    3.698 |  4430.02 |    0.656 |    97.59 |    4.354 |  3777.47 |
+|  8192 |     32 |    4 |  32896 |    7.403 |  4426.10 |    0.957 |   133.82 |    8.360 |  3934.97 |
+|  8192 |     32 |    8 |  65792 |   14.802 |  4427.63 |    1.222 |   209.44 |   16.024 |  4105.87 |
+|  8192 |     32 |   16 | 131584 |   29.596 |  4428.67 |    1.741 |   294.13 |   31.337 |  4199.00 |
+|  8192 |     32 |   32 | 263168 |   59.169 |  4430.42 |    2.619 |   390.92 |   61.789 |  4259.17 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      3714.25 ± 20.36 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         86.58 ± 0.43 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |      3445.17 ± 17.85 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         81.72 ± 0.53 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      3218.78 ± 11.34 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.86 ± 0.64 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       2732.83 ± 7.17 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.57 ± 0.51 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      2119.75 ± 12.81 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         62.33 ± 0.24 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      4505.82 ± 12.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         83.43 ± 0.59 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      4158.34 ± 18.84 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         79.22 ± 0.60 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      3993.81 ± 17.55 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         75.22 ± 1.05 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      3449.98 ± 12.13 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.36 ± 0.37 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      2689.42 ± 18.89 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         61.65 ± 0.30 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/gpt-oss-120b-GGUF
 
@@ -77,46 +77,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.571 |   897.18 |    0.543 |    58.96 |    1.113 |   488.60 |
-|   512 |     32 |    2 |   1088 |    0.593 |  1725.37 |    1.041 |    61.45 |    1.635 |   665.48 |
-|   512 |     32 |    4 |   2176 |    1.043 |  1963.15 |    1.334 |    95.95 |    2.377 |   915.36 |
-|   512 |     32 |    8 |   4352 |    2.099 |  1951.63 |    1.717 |   149.07 |    3.816 |  1140.45 |
-|   512 |     32 |   16 |   8704 |    4.207 |  1947.12 |    2.311 |   221.56 |    6.518 |  1335.35 |
-|   512 |     32 |   32 |  17408 |    8.422 |  1945.36 |    3.298 |   310.46 |   11.720 |  1485.27 |
-|  4096 |     32 |    1 |   4128 |    2.138 |  1915.88 |    0.571 |    56.09 |    2.708 |  1524.12 |
-|  4096 |     32 |    2 |   8256 |    4.266 |  1920.25 |    1.137 |    56.27 |    5.404 |  1527.90 |
-|  4096 |     32 |    4 |  16512 |    8.564 |  1913.02 |    1.471 |    86.99 |   10.036 |  1645.29 |
-|  4096 |     32 |    8 |  33024 |   17.092 |  1917.19 |    1.979 |   129.33 |   19.071 |  1731.63 |
-|  4096 |     32 |   16 |  66048 |   34.211 |  1915.65 |    2.850 |   179.66 |   37.061 |  1782.15 |
-|  4096 |     32 |   32 | 132096 |   68.394 |  1916.44 |    4.381 |   233.72 |   72.775 |  1815.13 |
-|  8192 |     32 |    1 |   8224 |    4.349 |  1883.45 |    0.620 |    51.65 |    4.969 |  1655.04 |
-|  8192 |     32 |    2 |  16448 |    8.674 |  1888.83 |    1.178 |    54.33 |    9.852 |  1669.48 |
-|  8192 |     32 |    4 |  32896 |   17.351 |  1888.55 |    1.580 |    81.01 |   18.931 |  1737.68 |
-|  8192 |     32 |    8 |  65792 |   34.743 |  1886.31 |    2.173 |   117.80 |   36.916 |  1782.20 |
-|  8192 |     32 |   16 | 131584 |   69.413 |  1888.29 |    3.297 |   155.28 |   72.710 |  1809.70 |
-|  8192 |     32 |   32 | 263168 |  138.903 |  1887.24 |    5.004 |   204.63 |  143.907 |  1828.73 |
+|   512 |     32 |    1 |    544 |    0.445 |  1151.80 |    0.560 |    57.14 |    1.005 |   541.53 |
+|   512 |     32 |    2 |   1088 |    0.472 |  2169.85 |    0.874 |    73.27 |    1.345 |   808.65 |
+|   512 |     32 |    4 |   2176 |    0.826 |  2480.33 |    1.299 |    98.51 |    2.125 |  1023.94 |
+|   512 |     32 |    8 |   4352 |    1.644 |  2491.67 |    1.608 |   159.18 |    3.252 |  1338.20 |
+|   512 |     32 |   16 |   8704 |    3.292 |  2488.35 |    2.117 |   241.85 |    5.409 |  1609.13 |
+|   512 |     32 |   32 |  17408 |    6.604 |  2481.07 |    2.898 |   353.31 |    9.502 |  1832.04 |
+|  4096 |     32 |    1 |   4128 |    1.698 |  2412.65 |    0.580 |    55.21 |    2.277 |  1812.66 |
+|  4096 |     32 |    2 |   8256 |    3.399 |  2409.88 |    0.934 |    68.53 |    4.333 |  1905.27 |
+|  4096 |     32 |    4 |  16512 |    6.823 |  2401.21 |    1.411 |    90.72 |    8.234 |  2005.30 |
+|  4096 |     32 |    8 |  33024 |   13.574 |  2413.97 |    1.841 |   139.07 |   15.415 |  2142.31 |
+|  4096 |     32 |   16 |  66048 |   27.176 |  2411.52 |    2.609 |   196.26 |   29.785 |  2217.49 |
+|  4096 |     32 |   32 | 132096 |   54.359 |  2411.23 |    3.905 |   262.20 |   58.264 |  2267.19 |
+|  8192 |     32 |    1 |   8224 |    3.491 |  2346.81 |    0.613 |    52.23 |    4.103 |  2004.21 |
+|  8192 |     32 |    2 |  16448 |    6.939 |  2361.03 |    0.981 |    65.21 |    7.921 |  2076.56 |
+|  8192 |     32 |    4 |  32896 |   13.888 |  2359.40 |    1.511 |    84.71 |   15.399 |  2136.21 |
+|  8192 |     32 |    8 |  65792 |   27.756 |  2361.18 |    2.034 |   125.86 |   29.790 |  2208.56 |
+|  8192 |     32 |   16 | 131584 |   55.554 |  2359.34 |    3.021 |   169.49 |   58.575 |  2246.41 |
+|  8192 |     32 |   32 | 263168 |  111.036 |  2360.89 |    4.537 |   225.72 |  115.573 |  2277.08 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       1919.36 ± 5.01 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         60.40 ± 0.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       1825.30 ± 6.37 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         56.94 ± 0.29 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1739.19 ± 6.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         52.51 ± 0.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1536.75 ± 4.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         49.33 ± 0.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1255.85 ± 3.26 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.99 ± 0.18 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2443.91 ± 7.47 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         58.72 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2309.84 ± 3.63 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         55.67 ± 0.35 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      2216.68 ± 10.16 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         52.87 ± 0.43 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1956.31 ± 6.39 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         49.45 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      1567.08 ± 11.79 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         42.76 ± 0.14 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
 
@@ -125,46 +125,46 @@ Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.398 |  1285.90 |    0.530 |    60.41 |    0.928 |   586.27 |
-|   512 |     32 |    2 |   1088 |    0.386 |  2651.65 |    0.948 |    67.50 |    1.334 |   815.38 |
-|   512 |     32 |    4 |   2176 |    0.666 |  3076.37 |    1.209 |   105.87 |    1.875 |  1160.71 |
-|   512 |     32 |    8 |   4352 |    1.325 |  3091.39 |    1.610 |   158.98 |    2.935 |  1482.65 |
-|   512 |     32 |   16 |   8704 |    2.664 |  3075.58 |    2.150 |   238.19 |    4.813 |  1808.39 |
-|   512 |     32 |   32 |  17408 |    5.336 |  3070.31 |    2.904 |   352.59 |    8.240 |  2112.50 |
-|  4096 |     32 |    1 |   4128 |    1.444 |  2836.81 |    0.581 |    55.09 |    2.025 |  2038.81 |
-|  4096 |     32 |    2 |   8256 |    2.872 |  2852.14 |    1.084 |    59.06 |    3.956 |  2086.99 |
-|  4096 |     32 |    4 |  16512 |    5.744 |  2852.32 |    1.440 |    88.90 |    7.184 |  2298.47 |
-|  4096 |     32 |    8 |  33024 |   11.463 |  2858.68 |    2.068 |   123.78 |   13.531 |  2440.65 |
-|  4096 |     32 |   16 |  66048 |   22.915 |  2859.95 |    3.018 |   169.67 |   25.933 |  2546.90 |
-|  4096 |     32 |   32 | 132096 |   45.956 |  2852.10 |    4.609 |   222.18 |   50.565 |  2612.39 |
-|  8192 |     32 |    1 |   8224 |    3.063 |  2674.72 |    0.693 |    46.20 |    3.755 |  2189.92 |
-|  8192 |     32 |    2 |  16448 |    6.109 |  2681.87 |    1.214 |    52.71 |    7.323 |  2245.98 |
-|  8192 |     32 |    4 |  32896 |   12.197 |  2686.63 |    1.682 |    76.11 |   13.878 |  2370.30 |
-|  8192 |     32 |    8 |  65792 |   24.409 |  2684.94 |    2.556 |   100.17 |   26.965 |  2439.95 |
-|  8192 |     32 |   16 | 131584 |   48.753 |  2688.50 |    3.994 |   128.20 |   52.747 |  2494.64 |
-|  8192 |     32 |   32 | 263168 |   97.508 |  2688.42 |    6.528 |   156.86 |  104.037 |  2529.57 |
+|   512 |     32 |    1 |    544 |    0.393 |  1303.73 |    0.548 |    58.36 |    0.941 |   578.10 |
+|   512 |     32 |    2 |   1088 |    0.387 |  2648.68 |    0.910 |    70.35 |    1.296 |   839.27 |
+|   512 |     32 |    4 |   2176 |    0.659 |  3107.63 |    1.302 |    98.33 |    1.961 |  1109.77 |
+|   512 |     32 |    8 |   4352 |    1.322 |  3099.35 |    1.669 |   153.42 |    2.990 |  1455.43 |
+|   512 |     32 |   16 |   8704 |    2.639 |  3104.63 |    2.212 |   231.44 |    4.851 |  1794.32 |
+|   512 |     32 |   32 |  17408 |    5.284 |  3100.80 |    2.955 |   346.53 |    8.239 |  2112.93 |
+|  4096 |     32 |    1 |   4128 |    1.417 |  2890.36 |    0.598 |    53.51 |    2.015 |  2048.45 |
+|  4096 |     32 |    2 |   8256 |    2.829 |  2895.62 |    1.019 |    62.82 |    3.848 |  2145.60 |
+|  4096 |     32 |    4 |  16512 |    5.656 |  2896.96 |    1.528 |    83.79 |    7.183 |  2298.71 |
+|  4096 |     32 |    8 |  33024 |   11.338 |  2890.02 |    2.127 |   120.36 |   13.465 |  2452.53 |
+|  4096 |     32 |   16 |  66048 |   22.709 |  2885.96 |    3.104 |   164.97 |   25.812 |  2558.79 |
+|  4096 |     32 |   32 | 132096 |   45.301 |  2893.35 |    4.723 |   216.80 |   50.024 |  2640.63 |
+|  8192 |     32 |    1 |   8224 |    3.022 |  2711.09 |    0.678 |    47.20 |    3.700 |  2222.89 |
+|  8192 |     32 |    2 |  16448 |    6.039 |  2713.01 |    1.149 |    55.70 |    7.188 |  2288.21 |
+|  8192 |     32 |    4 |  32896 |   12.050 |  2719.35 |    1.785 |    71.69 |   13.835 |  2377.67 |
+|  8192 |     32 |    8 |  65792 |   24.113 |  2717.90 |    2.629 |    97.39 |   26.741 |  2460.31 |
+|  8192 |     32 |   16 | 131584 |   48.178 |  2720.58 |    4.099 |   124.91 |   52.277 |  2517.06 |
+|  8192 |     32 |   32 | 263168 |   96.401 |  2719.31 |    6.696 |   152.93 |  103.097 |  2552.63 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2925.55 ± 4.25 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         62.80 ± 0.27 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2531.01 ± 6.79 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         55.86 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       2244.39 ± 5.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         45.95 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1783.17 ± 3.68 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.07 ± 0.10 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1241.90 ± 3.13 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         29.92 ± 0.06 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      2986.97 ± 18.87 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         61.06 ± 0.23 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2633.45 ± 6.26 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         54.77 ± 0.28 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2354.14 ± 3.84 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         48.02 ± 0.40 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1908.86 ± 4.25 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         40.23 ± 0.10 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1348.17 ± 2.00 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         30.21 ± 0.04 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
 
@@ -173,46 +173,46 @@ Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.211 |  2421.57 |    1.055 |    30.33 |    1.266 |   429.57 |
-|   512 |     32 |    2 |   1088 |    0.419 |  2441.34 |    1.130 |    56.65 |    1.549 |   702.32 |
-|   512 |     32 |    4 |   2176 |    0.873 |  2345.54 |    1.174 |   108.99 |    2.048 |  1062.74 |
-|   512 |     32 |    8 |   4352 |    1.727 |  2371.85 |    1.254 |   204.22 |    2.980 |  1460.19 |
-|   512 |     32 |   16 |   8704 |    3.452 |  2373.22 |    1.492 |   343.16 |    4.944 |  1760.56 |
-|   512 |     32 |   32 |  17408 |    6.916 |  2368.93 |    1.675 |   611.51 |    8.591 |  2026.36 |
-|  4096 |     32 |    1 |   4128 |    1.799 |  2277.26 |    1.084 |    29.51 |    2.883 |  1431.91 |
-|  4096 |     32 |    2 |   8256 |    3.577 |  2290.01 |    1.196 |    53.50 |    4.774 |  1729.51 |
-|  4096 |     32 |    4 |  16512 |    7.172 |  2284.36 |    1.313 |    97.50 |    8.485 |  1946.00 |
-|  4096 |     32 |    8 |  33024 |   14.341 |  2284.96 |    1.520 |   168.46 |   15.860 |  2082.18 |
-|  4096 |     32 |   16 |  66048 |   28.675 |  2285.44 |    1.983 |   258.21 |   30.658 |  2154.33 |
-|  4096 |     32 |   32 | 132096 |   57.354 |  2285.32 |    2.640 |   387.87 |   59.994 |  2201.82 |
-|  8192 |     32 |    1 |   8224 |    3.701 |  2213.75 |    1.119 |    28.59 |    4.820 |  1706.34 |
-|  8192 |     32 |    2 |  16448 |    7.410 |  2211.19 |    1.272 |    50.31 |    8.682 |  1894.56 |
-|  8192 |     32 |    4 |  32896 |   14.802 |  2213.83 |    1.460 |    87.68 |   16.261 |  2022.96 |
-|  8192 |     32 |    8 |  65792 |   29.609 |  2213.35 |    1.781 |   143.74 |   31.390 |  2095.93 |
-|  8192 |     32 |   16 | 131584 |   59.229 |  2212.96 |    2.495 |   205.17 |   61.725 |  2131.79 |
-|  8192 |     32 |   32 | 263168 |  118.449 |  2213.15 |    3.714 |   275.75 |  122.162 |  2154.25 |
+|   512 |     32 |    1 |    544 |    0.212 |  2420.12 |    1.100 |    29.10 |    1.311 |   414.85 |
+|   512 |     32 |    2 |   1088 |    0.428 |  2393.89 |    1.185 |    54.00 |    1.613 |   674.56 |
+|   512 |     32 |    4 |   2176 |    0.894 |  2290.41 |    1.229 |   104.17 |    2.123 |  1025.02 |
+|   512 |     32 |    8 |   4352 |    1.758 |  2330.36 |    1.319 |   194.15 |    3.076 |  1414.70 |
+|   512 |     32 |   16 |   8704 |    3.508 |  2335.21 |    1.543 |   331.90 |    5.051 |  1723.33 |
+|   512 |     32 |   32 |  17408 |    7.035 |  2328.93 |    1.738 |   589.21 |    8.773 |  1984.29 |
+|  4096 |     32 |    1 |   4128 |    1.831 |  2237.25 |    1.125 |    28.44 |    2.956 |  1396.42 |
+|  4096 |     32 |    2 |   8256 |    3.642 |  2249.48 |    1.253 |    51.07 |    4.895 |  1686.64 |
+|  4096 |     32 |    4 |  16512 |    7.274 |  2252.26 |    1.380 |    92.72 |    8.655 |  1907.81 |
+|  4096 |     32 |    8 |  33024 |   14.576 |  2248.09 |    1.617 |   158.29 |   16.193 |  2039.37 |
+|  4096 |     32 |   16 |  66048 |   29.138 |  2249.17 |    2.081 |   246.01 |   31.219 |  2115.63 |
+|  4096 |     32 |   32 | 132096 |   58.275 |  2249.19 |    2.814 |   363.87 |   61.089 |  2162.34 |
+|  8192 |     32 |    1 |   8224 |    3.757 |  2180.26 |    1.184 |    27.03 |    4.941 |  1664.37 |
+|  8192 |     32 |    2 |  16448 |    7.522 |  2178.05 |    1.341 |    47.73 |    8.863 |  1855.77 |
+|  8192 |     32 |    4 |  32896 |   15.043 |  2178.25 |    1.548 |    82.69 |   16.591 |  1982.74 |
+|  8192 |     32 |    8 |  65792 |   30.111 |  2176.49 |    1.937 |   132.13 |   32.048 |  2052.90 |
+|  8192 |     32 |   16 | 131584 |   60.405 |  2169.90 |    2.706 |   189.21 |   63.111 |  2084.97 |
+|  8192 |     32 |   32 | 263168 |  120.439 |  2176.58 |    3.993 |   256.46 |  124.432 |  2114.96 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2272.74 ± 4.68 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         30.66 ± 0.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2107.80 ± 9.55 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         29.71 ± 0.05 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1937.80 ± 6.75 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         28.86 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1641.12 ± 1.78 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.24 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1296.02 ± 2.67 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.78 ± 0.03 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2250.28 ± 6.41 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         29.43 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2100.19 ± 8.96 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         28.61 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2007.56 ± 4.16 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         27.38 ± 0.09 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1779.11 ± 6.42 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         25.72 ± 0.03 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1471.23 ± 1.71 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         22.51 ± 0.02 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/gemma-3-4b-it-qat-GGUF
 
@@ -221,44 +221,91 @@ Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.094 |  5434.73 |    0.394 |    81.21 |    0.488 |  1114.15 |
-|   512 |     32 |    2 |   1088 |    0.168 |  6091.68 |    0.498 |   128.52 |    0.666 |  1633.41 |
-|   512 |     32 |    4 |   2176 |    0.341 |  6010.68 |    0.542 |   236.37 |    0.882 |  2466.43 |
-|   512 |     32 |    8 |   4352 |    0.665 |  6161.46 |    0.678 |   377.74 |    1.342 |  3241.72 |
-|   512 |     32 |   16 |   8704 |    1.323 |  6193.19 |    0.902 |   567.41 |    2.225 |  3911.74 |
-|   512 |     32 |   32 |  17408 |    2.642 |  6202.03 |    1.231 |   832.03 |    3.872 |  4495.36 |
-|  4096 |     32 |    1 |   4128 |    0.701 |  5840.49 |    0.439 |    72.95 |    1.140 |  3621.23 |
-|  4096 |     32 |    2 |   8256 |    1.387 |  5906.82 |    0.574 |   111.48 |    1.961 |  4210.12 |
-|  4096 |     32 |    4 |  16512 |    2.758 |  5940.33 |    0.651 |   196.58 |    3.409 |  4843.33 |
-|  4096 |     32 |    8 |  33024 |    5.491 |  5967.56 |    0.876 |   292.40 |    6.367 |  5187.12 |
-|  4096 |     32 |   16 |  66048 |   10.978 |  5969.58 |    1.275 |   401.69 |   12.253 |  5390.38 |
-|  4096 |     32 |   32 | 132096 |   21.944 |  5972.93 |    1.992 |   514.16 |   23.936 |  5518.73 |
-|  8192 |     32 |    1 |   8224 |    1.402 |  5841.91 |    0.452 |    70.73 |    1.855 |  4434.12 |
-|  8192 |     32 |    2 |  16448 |    2.793 |  5865.34 |    0.637 |   100.55 |    3.430 |  4795.51 |
-|  8192 |     32 |    4 |  32896 |    5.564 |  5889.64 |    0.770 |   166.26 |    6.334 |  5193.95 |
-|  8192 |     32 |    8 |  65792 |   11.114 |  5896.44 |    1.122 |   228.07 |   12.237 |  5376.51 |
-|  8192 |     32 |   16 | 131584 |   22.210 |  5901.38 |    1.789 |   286.15 |   24.000 |  5482.74 |
-|  8192 |     32 |   32 | 263168 |   44.382 |  5906.56 |    3.044 |   336.38 |   47.426 |  5549.02 |
+|   512 |     32 |    1 |    544 |    0.092 |  5566.97 |    0.412 |    77.63 |    0.504 |  1078.95 |
+|   512 |     32 |    2 |   1088 |    0.161 |  6345.67 |    0.522 |   122.70 |    0.683 |  1593.06 |
+|   512 |     32 |    4 |   2176 |    0.325 |  6309.87 |    0.562 |   227.68 |    0.887 |  2453.87 |
+|   512 |     32 |    8 |   4352 |    0.643 |  6374.42 |    0.685 |   373.67 |    1.328 |  3277.94 |
+|   512 |     32 |   16 |   8704 |    1.277 |  6413.64 |    0.915 |   559.47 |    2.192 |  3970.01 |
+|   512 |     32 |   32 |  17408 |    2.518 |  6506.57 |    1.249 |   819.61 |    3.767 |  4620.64 |
+|  4096 |     32 |    1 |   4128 |    0.674 |  6079.68 |    0.453 |    70.60 |    1.127 |  3662.88 |
+|  4096 |     32 |    2 |   8256 |    1.335 |  6137.82 |    0.627 |   102.03 |    1.962 |  4208.11 |
+|  4096 |     32 |    4 |  16512 |    2.657 |  6167.35 |    0.749 |   170.92 |    3.405 |  4848.71 |
+|  4096 |     32 |    8 |  33024 |    5.307 |  6173.91 |    0.974 |   262.89 |    6.281 |  5257.53 |
+|  4096 |     32 |   16 |  66048 |   10.610 |  6176.96 |    1.379 |   371.42 |   11.988 |  5509.40 |
+|  4096 |     32 |   32 | 132096 |   21.213 |  6178.89 |    2.122 |   482.50 |   23.335 |  5660.82 |
+|  8192 |     32 |    1 |   8224 |    1.359 |  6027.34 |    0.467 |    68.52 |    1.826 |  4503.48 |
+|  8192 |     32 |    2 |  16448 |    2.699 |  6069.68 |    0.653 |    98.03 |    3.352 |  4906.68 |
+|  8192 |     32 |    4 |  32896 |    5.366 |  6106.74 |    0.818 |   156.55 |    6.184 |  5319.96 |
+|  8192 |     32 |    8 |  65792 |   10.755 |  6093.50 |    1.174 |   218.04 |   11.929 |  5515.22 |
+|  8192 |     32 |   16 | 131584 |   21.484 |  6100.82 |    1.829 |   279.90 |   23.314 |  5644.11 |
+|  8192 |     32 |   32 | 263168 |   42.950 |  6103.40 |    3.058 |   334.91 |   46.008 |  5720.05 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      5810.04 ± 21.71 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         84.54 ± 0.18 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       5288.04 ± 3.54 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         78.82 ± 1.37 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      4960.43 ± 16.64 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.13 ± 0.30 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |      4495.92 ± 31.11 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         72.37 ± 0.29 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      3746.90 ± 40.01 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         63.02 ± 0.20 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      5948.74 ± 10.61 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         81.05 ± 0.20 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      5652.69 ± 34.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         76.37 ± 0.58 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      5509.57 ± 40.69 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         71.61 ± 0.80 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      5340.86 ± 36.92 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.89 ± 0.34 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      5023.30 ± 13.52 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         62.28 ± 0.30 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.433 |  1181.83 |    0.693 |    46.16 |    1.126 |   482.94 |
+|   512 |     32 |    2 |   1088 |    0.439 |  2334.46 |    1.034 |    61.89 |    1.473 |   738.75 |
+|   512 |     32 |    4 |   2176 |    0.772 |  2654.46 |    1.459 |    87.76 |    2.230 |   975.77 |
+|   512 |     32 |    8 |   4352 |    1.541 |  2658.78 |    2.043 |   125.31 |    3.583 |  1214.47 |
+|   512 |     32 |   16 |   8704 |    3.083 |  2656.91 |    2.675 |   191.42 |    5.758 |  1511.62 |
+|   512 |     32 |   32 |  17408 |    6.159 |  2660.12 |    3.615 |   283.24 |    9.774 |  1780.98 |
+|  4096 |     32 |    1 |   4128 |    1.915 |  2139.30 |    0.725 |    44.14 |    2.640 |  1563.83 |
+|  4096 |     32 |    2 |   8256 |    3.834 |  2136.40 |    1.119 |    57.21 |    4.953 |  1666.81 |
+|  4096 |     32 |    4 |  16512 |    7.636 |  2145.72 |    1.631 |    78.49 |    9.266 |  1781.93 |
+|  4096 |     32 |    8 |  33024 |   15.295 |  2142.40 |    2.344 |   109.21 |   17.639 |  1872.20 |
+|  4096 |     32 |   16 |  66048 |   30.573 |  2143.62 |    3.773 |   135.70 |   34.346 |  1923.04 |
+|  4096 |     32 |   32 | 132096 |   61.282 |  2138.82 |    5.795 |   176.71 |   67.077 |  1969.31 |
+|  8192 |     32 |    1 |   8224 |    4.510 |  1816.24 |    0.760 |    42.11 |    5.270 |  1560.44 |
+|  8192 |     32 |    2 |  16448 |    9.036 |  1813.19 |    1.206 |    53.06 |   10.242 |  1605.91 |
+|  8192 |     32 |    4 |  32896 |   18.070 |  1813.43 |    1.783 |    71.80 |   19.852 |  1657.03 |
+|  8192 |     32 |    8 |  65792 |   36.125 |  1814.15 |    2.635 |    97.14 |   38.760 |  1697.41 |
+|  8192 |     32 |   16 | 131584 |   72.367 |  1811.20 |    4.954 |   103.34 |   77.322 |  1701.77 |
+|  8192 |     32 |   32 | 263168 |  144.501 |  1814.13 |    8.103 |   126.37 |  152.604 |  1724.51 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |          pp2048 |      2364.18 ± 11.43 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |            tg32 |         48.68 ± 0.12 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d4096 |       1684.13 ± 1.24 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d4096 |         44.62 ± 0.22 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d8192 |       1314.68 ± 1.41 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d8192 |         42.59 ± 0.11 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d16384 |        914.05 ± 3.32 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d16384 |         38.72 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d32768 |        567.20 ± 0.90 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d32768 |         32.65 ± 0.09 |
+
+build: 11fb327bf (7941)

benches/mac-m2-ultra/mac-m2-ultra.md

@@ -0,0 +1,298 @@
+## System info
+
+```bash
+uname -a
+Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
+
+g++ --version
+Apple clang version 17.0.0 (clang-1700.3.19.1)
+Target: arm64-apple-darwin25.2.0
+```
+
+## ggml-org/gpt-oss-20b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.215 |  2381.35 |    0.245 |   130.45 |    0.460 |  1181.81 |
+|   512 |     32 |    2 |   1088 |    0.379 |  2701.43 |    0.382 |   167.56 |    0.761 |  1429.67 |
+|   512 |     32 |    4 |   2176 |    0.721 |  2839.27 |    0.604 |   211.76 |    1.326 |  1641.32 |
+|   512 |     32 |    8 |   4352 |    1.433 |  2858.30 |    1.033 |   247.75 |    2.466 |  1764.57 |
+|   512 |     32 |   16 |   8704 |    2.853 |  2871.12 |    1.570 |   326.11 |    4.423 |  1967.77 |
+|   512 |     32 |   32 |  17408 |    5.699 |  2874.95 |    1.910 |   536.15 |    7.609 |  2287.88 |
+|  4096 |     32 |    1 |   4128 |    1.552 |  2638.56 |    0.334 |    95.72 |    1.887 |  2188.00 |
+|  4096 |     32 |    2 |   8256 |    3.084 |  2655.88 |    0.404 |   158.54 |    3.488 |  2366.86 |
+|  4096 |     32 |    4 |  16512 |    6.151 |  2663.78 |    0.652 |   196.39 |    6.802 |  2427.37 |
+|  4096 |     32 |    8 |  33024 |   12.288 |  2666.77 |    1.135 |   225.47 |   13.423 |  2460.27 |
+|  4096 |     32 |   16 |  66048 |   24.563 |  2668.12 |    1.762 |   290.55 |   26.325 |  2508.97 |
+|  4096 |     32 |   32 | 132096 |   49.114 |  2668.73 |    2.398 |   426.94 |   51.512 |  2564.35 |
+|  8192 |     32 |    1 |   8224 |    3.345 |  2448.78 |    0.275 |   116.46 |    3.620 |  2271.76 |
+|  8192 |     32 |    2 |  16448 |    6.665 |  2458.11 |    0.425 |   150.71 |    7.090 |  2319.91 |
+|  8192 |     32 |    4 |  32896 |   13.315 |  2460.92 |    0.691 |   185.21 |   14.006 |  2348.63 |
+|  8192 |     32 |    8 |  65792 |   26.611 |  2462.73 |    1.212 |   211.16 |   27.823 |  2364.62 |
+|  8192 |     32 |   16 | 131584 |   53.232 |  2462.27 |    1.919 |   266.83 |   55.151 |  2385.88 |
+|  8192 |     32 |   32 | 263168 |  110.455 |  2373.30 |    2.752 |   372.03 |  113.208 |  2324.64 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2713.40 ± 3.56 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        129.97 ± 3.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2324.59 ± 3.01 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        123.38 ± 0.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |      1989.82 ± 30.11 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        117.39 ± 0.33 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1556.54 ± 6.22 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        109.75 ± 0.42 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       1122.63 ± 1.45 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         98.25 ± 0.08 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gpt-oss-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.426 |  1200.92 |    0.361 |    88.56 |    0.788 |   690.64 |
+|   512 |     32 |    2 |   1088 |    0.683 |  1500.14 |    0.545 |   117.35 |    1.228 |   886.02 |
+|   512 |     32 |    4 |   2176 |    1.204 |  1701.56 |    0.847 |   151.19 |    2.050 |  1061.34 |
+|   512 |     32 |    8 |   4352 |    2.402 |  1705.20 |    1.455 |   176.00 |    3.857 |  1128.45 |
+|   512 |     32 |   16 |   8704 |    4.802 |  1705.90 |    2.349 |   217.93 |    7.152 |  1217.08 |
+|   512 |     32 |   32 |  17408 |    9.593 |  1707.85 |    3.665 |   279.42 |   13.258 |  1313.01 |
+|  4096 |     32 |    1 |   4128 |    2.581 |  1587.08 |    0.390 |    82.12 |    2.970 |  1389.67 |
+|  4096 |     32 |    2 |   8256 |    5.124 |  1598.79 |    0.589 |   108.62 |    5.713 |  1445.10 |
+|  4096 |     32 |    4 |  16512 |   10.231 |  1601.47 |    0.928 |   137.98 |   11.158 |  1479.80 |
+|  4096 |     32 |    8 |  33024 |   20.468 |  1600.94 |    1.606 |   159.38 |   22.074 |  1496.04 |
+|  4096 |     32 |   16 |  66048 |   40.924 |  1601.42 |    2.639 |   193.99 |   43.563 |  1516.15 |
+|  4096 |     32 |   32 | 132096 |   81.819 |  1601.98 |    4.466 |   229.29 |   86.284 |  1530.94 |
+|  8192 |     32 |    1 |   8224 |    5.517 |  1484.74 |    0.409 |    78.16 |    5.927 |  1387.58 |
+|  8192 |     32 |    2 |  16448 |   11.008 |  1488.43 |    0.622 |   102.92 |   11.629 |  1414.34 |
+|  8192 |     32 |    4 |  32896 |   22.002 |  1489.29 |    0.987 |   129.66 |   22.990 |  1430.90 |
+|  8192 |     32 |    8 |  65792 |   46.051 |  1423.11 |    1.858 |   137.79 |   47.909 |  1373.27 |
+|  8192 |     32 |   16 | 131584 |   97.680 |  1341.85 |    2.872 |   178.28 |  100.552 |  1308.62 |
+|  8192 |     32 |   32 | 263168 |  176.407 |  1486.02 |    5.048 |   202.85 |  181.455 |  1450.32 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1648.69 ± 1.80 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         85.60 ± 0.52 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1429.86 ± 1.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         82.03 ± 0.12 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1257.90 ± 1.81 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         78.23 ± 0.33 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1013.49 ± 0.70 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         73.20 ± 0.28 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        721.11 ± 0.58 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         65.52 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.243 |  2109.23 |    0.419 |    76.34 |    0.662 |   821.84 |
+|   512 |     32 |    2 |   1088 |    0.406 |  2521.40 |    0.575 |   111.36 |    0.981 |  1109.27 |
+|   512 |     32 |    4 |   2176 |    0.744 |  2751.65 |    0.841 |   152.22 |    1.585 |  1372.71 |
+|   512 |     32 |    8 |   4352 |    1.479 |  2770.20 |    1.330 |   192.48 |    2.809 |  1549.53 |
+|   512 |     32 |   16 |   8704 |    2.951 |  2776.20 |    2.572 |   199.05 |    5.523 |  1575.93 |
+|   512 |     32 |   32 |  17408 |    5.899 |  2777.64 |    2.603 |   393.34 |    8.502 |  2047.54 |
+|  4096 |     32 |    1 |   4128 |    1.901 |  2154.15 |    0.474 |    67.58 |    2.375 |  1738.14 |
+|  4096 |     32 |    2 |   8256 |    3.788 |  2162.89 |    0.652 |    98.17 |    4.439 |  1859.69 |
+|  4096 |     32 |    4 |  16512 |    7.564 |  2166.18 |    0.990 |   129.24 |    8.554 |  1930.34 |
+|  4096 |     32 |    8 |  33024 |   15.121 |  2166.98 |    1.632 |   156.82 |   16.754 |  1971.12 |
+|  4096 |     32 |   16 |  66048 |   30.241 |  2167.09 |    3.166 |   161.72 |   33.407 |  1977.04 |
+|  4096 |     32 |   32 | 132096 |   60.474 |  2167.42 |    3.780 |   270.93 |   64.254 |  2055.86 |
+|  8192 |     32 |    1 |   8224 |    4.733 |  1730.92 |    0.483 |    66.29 |    5.215 |  1576.85 |
+|  8192 |     32 |    2 |  16448 |    9.459 |  1732.09 |    0.722 |    88.58 |   10.182 |  1615.46 |
+|  8192 |     32 |    4 |  32896 |   18.912 |  1732.65 |    1.120 |   114.26 |   20.032 |  1642.14 |
+|  8192 |     32 |    8 |  65792 |   37.797 |  1733.91 |    1.873 |   136.67 |   39.670 |  1658.49 |
+|  8192 |     32 |   16 | 131584 |   84.133 |  1557.92 |    3.718 |   137.72 |   87.850 |  1497.82 |
+|  8192 |     32 |   32 | 263168 |  157.550 |  1663.88 |    4.854 |   210.98 |  162.403 |  1620.46 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2453.11 ± 1.70 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         78.97 ± 0.46 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1569.46 ± 1.97 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         71.18 ± 0.37 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1145.51 ± 1.16 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         65.11 ± 0.36 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        741.04 ± 0.74 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         56.87 ± 0.14 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        431.31 ± 0.31 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         45.26 ± 0.11 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.339 |  1509.22 |    0.409 |    78.17 |    0.749 |   726.67 |
+|   512 |     32 |    2 |   1088 |    0.646 |  1584.93 |    0.483 |   132.45 |    1.129 |   963.45 |
+|   512 |     32 |    4 |   2176 |    1.258 |  1627.50 |    0.585 |   218.67 |    1.844 |  1180.21 |
+|   512 |     32 |    8 |   4352 |    2.506 |  1634.41 |    1.005 |   254.83 |    3.511 |  1239.64 |
+|   512 |     32 |   16 |   8704 |    5.007 |  1635.99 |    1.595 |   321.07 |    6.602 |  1318.38 |
+|   512 |     32 |   32 |  17408 |   10.007 |  1637.19 |    1.676 |   611.12 |   11.683 |  1490.03 |
+|  4096 |     32 |    1 |   4128 |    2.730 |  1500.46 |    0.431 |    74.31 |    3.160 |  1306.12 |
+|  4096 |     32 |    2 |   8256 |    5.446 |  1504.33 |    0.524 |   122.04 |    5.970 |  1382.91 |
+|  4096 |     32 |    4 |  16512 |   10.875 |  1506.59 |    0.662 |   193.45 |   11.537 |  1431.28 |
+|  4096 |     32 |    8 |  33024 |   21.749 |  1506.61 |    1.158 |   221.11 |   22.907 |  1441.64 |
+|  4096 |     32 |   16 |  66048 |   43.477 |  1507.36 |    1.901 |   269.32 |   45.378 |  1455.49 |
+|  4096 |     32 |   32 | 132096 |   86.954 |  1507.37 |    2.325 |   440.42 |   89.279 |  1479.59 |
+|  8192 |     32 |    1 |   8224 |    5.940 |  1379.21 |    0.449 |    71.20 |    6.389 |  1287.20 |
+|  8192 |     32 |    2 |  16448 |   11.865 |  1380.84 |    0.559 |   114.59 |   12.424 |  1323.92 |
+|  8192 |     32 |    4 |  32896 |   23.723 |  1381.25 |    0.728 |   175.80 |   24.452 |  1345.35 |
+|  8192 |     32 |    8 |  65792 |   47.434 |  1381.63 |    1.279 |   200.09 |   48.713 |  1350.60 |
+|  8192 |     32 |   16 | 131584 |   94.864 |  1381.69 |    2.198 |   232.97 |   97.061 |  1355.68 |
+|  8192 |     32 |   32 | 263168 |  189.743 |  1381.57 |    3.052 |   335.50 |  192.795 |  1365.01 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1565.91 ± 0.86 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         79.68 ± 0.39 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1317.41 ± 1.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         74.70 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1134.65 ± 0.76 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         71.31 ± 0.12 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        886.46 ± 0.78 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         65.93 ± 0.06 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        612.21 ± 0.30 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         56.83 ± 0.02 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gemma-3-4b-it-qat-GGUF
+
+Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.186 |  2748.06 |    0.235 |   136.28 |    0.421 |  1291.78 |
+|   512 |     32 |    2 |   1088 |    0.342 |  2990.95 |    0.312 |   204.99 |    0.655 |  1662.15 |
+|   512 |     32 |    4 |   2176 |    0.662 |  3092.69 |    0.404 |   316.97 |    1.066 |  2041.21 |
+|   512 |     32 |    8 |   4352 |    1.317 |  3110.41 |    0.579 |   441.80 |    1.896 |  2294.97 |
+|   512 |     32 |   16 |   8704 |    2.625 |  3120.23 |    1.207 |   424.08 |    3.833 |  2270.93 |
+|   512 |     32 |   32 |  17408 |    5.242 |  3125.34 |    1.299 |   788.23 |    6.541 |  2661.19 |
+|  4096 |     32 |    1 |   4128 |    1.408 |  2909.90 |    0.296 |   108.07 |    1.704 |  2422.95 |
+|  4096 |     32 |    2 |   8256 |    2.793 |  2933.40 |    0.325 |   197.00 |    3.118 |  2648.25 |
+|  4096 |     32 |    4 |  16512 |    5.567 |  2943.22 |    0.440 |   291.07 |    6.006 |  2749.05 |
+|  4096 |     32 |    8 |  33024 |   11.114 |  2948.23 |    0.640 |   400.26 |   11.754 |  2809.59 |
+|  4096 |     32 |   16 |  66048 |   22.217 |  2949.76 |    1.327 |   385.83 |   23.544 |  2805.26 |
+|  4096 |     32 |   32 | 132096 |   44.420 |  2950.77 |    1.553 |   659.30 |   45.973 |  2873.36 |
+|  8192 |     32 |    1 |   8224 |    2.860 |  2864.58 |    0.250 |   127.90 |    3.110 |  2644.42 |
+|  8192 |     32 |    2 |  16448 |    5.702 |  2873.63 |    0.335 |   191.07 |    6.036 |  2724.77 |
+|  8192 |     32 |    4 |  32896 |   11.383 |  2878.69 |    0.456 |   280.72 |   11.839 |  2778.63 |
+|  8192 |     32 |    8 |  65792 |   22.750 |  2880.75 |    0.671 |   381.48 |   23.421 |  2809.14 |
+|  8192 |     32 |   16 | 131584 |   45.484 |  2881.74 |    1.406 |   364.04 |   46.890 |  2806.22 |
+|  8192 |     32 |   32 | 263168 |   90.956 |  2882.10 |    1.793 |   570.98 |   92.749 |  2837.41 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2923.59 ± 3.10 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        134.28 ± 1.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2748.21 ± 3.05 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        133.11 ± 0.08 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       2641.45 ± 2.31 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        125.85 ± 0.35 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       2446.20 ± 2.94 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        125.00 ± 0.12 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       2129.18 ± 7.43 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |        113.14 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.326 |  1568.69 |    0.522 |    61.28 |    0.849 |   641.09 |
+|   512 |     32 |    2 |   1088 |    0.528 |  1939.42 |    0.744 |    86.07 |    1.272 |   855.63 |
+|   512 |     32 |    4 |   2176 |    0.968 |  2114.85 |    1.105 |   115.85 |    2.073 |  1049.56 |
+|   512 |     32 |    8 |   4352 |    1.928 |  2124.62 |    1.684 |   151.99 |    3.612 |  1204.82 |
+|   512 |     32 |   16 |   8704 |    3.844 |  2131.34 |    3.141 |   162.99 |    6.985 |  1246.11 |
+|   512 |     32 |   32 |  17408 |    7.683 |  2132.38 |    3.924 |   260.95 |   11.608 |  1499.71 |
+|  4096 |     32 |    1 |   4128 |    3.280 |  1248.75 |    0.723 |    44.29 |    4.003 |  1031.33 |
+|  4096 |     32 |    2 |   8256 |    6.545 |  1251.63 |    0.930 |    68.85 |    7.475 |  1104.53 |
+|  4096 |     32 |    4 |  16512 |   13.080 |  1252.64 |    1.454 |    88.03 |   14.534 |  1136.12 |
+|  4096 |     32 |    8 |  33024 |   26.154 |  1252.90 |    2.388 |   107.20 |   28.542 |  1157.04 |
+|  4096 |     32 |   16 |  66048 |   52.297 |  1253.14 |    4.724 |   108.37 |   57.022 |  1158.30 |
+|  4096 |     32 |   32 | 132096 |  104.578 |  1253.34 |    7.266 |   140.93 |  111.844 |  1181.08 |
+|  8192 |     32 |    1 |   8224 |    9.623 |   851.31 |    0.767 |    41.72 |   10.390 |   791.54 |
+|  8192 |     32 |    2 |  16448 |   20.916 |   783.32 |    1.148 |    55.74 |   22.064 |   745.45 |
+|  8192 |     32 |    4 |  32896 |   43.509 |   753.14 |    1.833 |    69.82 |   45.342 |   725.51 |
+|  8192 |     32 |    8 |  65792 |   79.621 |   823.10 |    3.180 |    80.50 |   82.801 |   794.58 |
+|  8192 |     32 |   16 | 131584 |  153.770 |   852.39 |    6.502 |    78.74 |  160.272 |   821.00 |
+|  8192 |     32 |   32 | 263168 |  307.539 |   852.39 |   10.839 |    94.48 |  318.378 |   826.59 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1629.33 ± 0.27 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         59.58 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |        732.67 ± 0.42 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         47.44 ± 0.15 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |        474.33 ± 0.33 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         40.20 ± 0.20 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        277.46 ± 0.09 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         31.50 ± 0.93 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        151.44 ± 0.05 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         21.81 ± 0.01 |
+
+build: b828e18c7 (7948)

build-xcframework.sh

@@ -43,11 +43,6 @@ COMMON_CMAKE_ARGS=(
     -DGGML_OPENMP=${GGML_OPENMP}
 )
 
-XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
-MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
-MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
-echo "Detected Xcode version: $XCODE_VERSION"
-
 check_required_tool() {
     local tool=$1
     local install_message=$2
@@ -60,9 +55,12 @@ check_required_tool() {
 }
 echo "Checking for required tools..."
 check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
-check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
-check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
-check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "xcrun" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+
+XCODE_VERSION=$(xcrun xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
+MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
+MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
+echo "Detected Xcode version: $XCODE_VERSION"
 
 set -e
 
@@ -260,7 +258,7 @@ combine_static_libraries() {
 
     # Since we have multiple architectures libtool will find object files that do not
     # match the target architecture. We suppress these warnings.
-    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+    xcrun libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
 
     # Determine SDK, architectures, and install_name based on platform and simulator flag.
     local sdk=""
@@ -333,7 +331,7 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if command -v xcrun vtool &>/dev/null; then
+        if xcrun -f vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
@@ -451,10 +449,9 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
@@ -467,10 +464,9 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
@@ -528,13 +524,13 @@ combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
-xcodebuild -create-xcframework \
+xcrun xcodebuild -create-xcframework \
     -framework $(pwd)/build-ios-sim/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
     -framework $(pwd)/build-ios-device/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
     -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
+    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos/framework/llama.framework \
     -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos-sim/framework/llama.framework \

ci/run.sh

@@ -635,6 +635,29 @@ function gg_check_build_requirements {
     fi
 }
 
+function gg_run_test_backend_ops_cpu {
+    cd ${SRC}
+
+    cd build-ci-release
+
+    set -e
+
+    (time ./bin/test-backend-ops -b CPU ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops-cpu.log
+
+    set +e
+}
+
+function gg_sum_test_backend_ops_cpu {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test-backend-ops for CPU backend\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops-cpu.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
 ## main
 
 export LLAMA_LOG_PREFIX=1
@@ -663,6 +686,10 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
+if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
+    test $ret -eq 0 && gg_run test_backend_ops_cpu
+fi
+
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
     test $ret -eq 0 && gg_run rerank_tiny

cmake/common.cmake

@@ -32,4 +32,27 @@ function(llama_add_compile_flags)
             set(CXX_FLAGS "" PARENT_SCOPE)
         endif()
     endif()
+
+    if (NOT MSVC)
+        if (LLAMA_SANITIZE_THREAD)
+            message(STATUS "Using -fsanitize=thread")
+
+            add_compile_options(-fsanitize=thread)
+            link_libraries     (-fsanitize=thread)
+        endif()
+
+        if (LLAMA_SANITIZE_ADDRESS)
+            message(STATUS "Using -fsanitize=address")
+
+            add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+            link_libraries     (-fsanitize=address)
+        endif()
+
+        if (LLAMA_SANITIZE_UNDEFINED)
+            message(STATUS "Using -fsanitize=undefined")
+
+            add_compile_options(-fsanitize=undefined)
+            link_libraries     (-fsanitize=undefined)
+        endif()
+    endif()
 endfunction()

common/CMakeLists.txt

@@ -112,11 +112,7 @@ endif()
 
 # TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
 set(LLAMA_COMMON_EXTRA_LIBS build_info)
-
-if (LLAMA_HTTPLIB)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
-endif()
+set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)

common/arg.cpp

@@ -1301,7 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.kv_unified = value;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},
@@ -3437,16 +3437,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.ngram_size_m = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--spec-ngram-check-rate"}, "N",
-        string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
-        [](common_params & params, int value) {
-            if (value < 1) {
-                throw std::invalid_argument("ngram check rate must be at least 1");
-            }
-            params.speculative.ngram_check_rate = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--spec-ngram-min-hits"}, "N",
         string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),

common/chat.cpp

@@ -380,15 +380,46 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
     return msgs;
 }
 
-json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
+    if (!c.supports_string_content && !c.supports_typed_content) {
+        LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
+    }
+
+    bool only_string_accepted =  c.supports_string_content && !c.supports_typed_content;
+    bool only_typed_accepted  = !c.supports_string_content &&  c.supports_typed_content;
+
     json messages = json::array();
     for (const auto & msg : msgs) {
-        json jmsg = msg.to_json_oaicompat(concat_typed_text);
-        messages.push_back(jmsg);
+        if (only_string_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
+            messages.push_back(jmsg);
+        } else if (only_typed_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            if (jmsg.at("content").is_string()) {
+                jmsg["content"] = json::array({
+                    json{
+                        {"type", "text"},
+                        {"text", jmsg.at("content").get<std::string>()},
+                    }
+                });
+            }
+            messages.push_back(jmsg);
+        } else {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            messages.push_back(jmsg);
+        }
     }
     return messages;
 }
 
+// DEPRECATED: only used in tests
+json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+    jinja::caps c;
+    c.supports_string_content = true;
+    c.supports_typed_content = !concat_typed_text;
+    return render_message_to_json(msgs, c);
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -3020,7 +3051,7 @@ static common_chat_params common_chat_templates_apply_jinja(
         : *tmpls->template_default;
     const auto & src = tmpl.source();
     const auto & caps = tmpl.original_caps();
-    params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
+    params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
     params.reasoning_format = inputs.reasoning_format;

common/chat.h

@@ -240,6 +240,8 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+
+// DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);

common/common.cpp

@@ -1,7 +1,3 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "ggml.h"
 #include "gguf.h"
 
@@ -9,12 +5,12 @@
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
+#include "unicode.h"
 
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
-#include <codecvt>
 #include <chrono>
 #include <cstdarg>
 #include <cstring>
@@ -706,45 +702,28 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
         return false;
     }
 
-    std::u32string filename_utf32;
-    try {
-#if defined(__clang__)
-        // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
+    size_t offset = 0;
+    while (offset < filename.size()) {
+        utf8_parse_result result = parse_utf8_codepoint(filename, offset);
 
-        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-        filename_utf32 = converter.from_bytes(filename);
-
-        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
-        // or invalid encodings were encountered. Reject such attempts
-        std::string filename_reencoded = converter.to_bytes(filename_utf32);
-        if (filename_reencoded != filename) {
+        if (result.status != utf8_parse_result::SUCCESS) {
             return false;
         }
-    } catch (const std::exception &) {
-        return false;
-    }
+        uint32_t c = result.codepoint;
 
-    // Check for forbidden codepoints:
-    // - Control characters
-    // - Unicode equivalents of illegal characters
-    // - UTF-16 surrogate pairs
-    // - UTF-8 replacement character
-    // - Byte order mark (BOM)
-    // - Illegal characters: / \ : * ? " < > |
-    for (char32_t c : filename_utf32) {
+        if ((result.bytes_consumed == 2 && c < 0x80) ||
+            (result.bytes_consumed == 3 && c < 0x800) ||
+            (result.bytes_consumed == 4 && c < 0x10000)) {
+            return false;
+        }
+
+        // Check for forbidden codepoints:
+        // - Control characters
+        // - Unicode equivalents of illegal characters
+        // - UTF-16 surrogate pairs
+        // - UTF-8 replacement character
+        // - Byte order mark (BOM)
+        // - Illegal characters: / \ : * ? " < > |
         if (c <= 0x1F // Control characters (C0)
             || c == 0x7F // Control characters (DEL)
             || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
@@ -752,6 +731,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             || c == 0x2215 // Division Slash (forward slash equivalent)
             || c == 0x2216 // Set Minus (backslash equivalent)
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c > 0x10FFFF // Max Unicode limit
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
             || c == ':' || c == '*' // Illegal characters
@@ -762,6 +742,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             // Subdirectories not allowed, reject path separators
             return false;
         }
+        offset += result.bytes_consumed;
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -898,7 +879,8 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
+        defined(__OpenBSD__) || defined(__NetBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else if (std::getenv("HOME")) {
@@ -1242,7 +1224,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
             return res;
         }
 
-        int err = llama_apply_adapter_cvec(
+        int err = llama_set_adapter_cvec(
                 lctx,
                 cvec.data.data(),
                 cvec.data.size(),
@@ -1344,12 +1326,15 @@ std::string get_model_endpoint() {
 }
 
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    llama_clear_adapter_lora(ctx);
-    for (auto & la : lora) {
-        if (la.scale != 0.0f) {
-            llama_set_adapter_lora(ctx, la.ptr, la.scale);
-        }
+    std::vector<llama_adapter_lora *> loras;
+    std::vector<float> scales;
+
+    for (auto & la: lora) {
+        loras.push_back(la.ptr);
+        scales.push_back(la.scale);
     }
+
+    llama_set_adapters_lora(ctx, loras.data(), loras.size(), scales.data());
 }
 
 struct llama_model_params common_model_params_to_llama(common_params & params) {
@@ -1469,66 +1454,6 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
-//
-// Token utils
-//
-
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 //
 // Vocab utils
 //

common/common.h

@@ -269,7 +269,6 @@ struct common_params_speculative {
 
     uint16_t ngram_size_n     = 12; // ngram size for lookup
     uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_check_rate =  1; // check rate for ngram lookup
     uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
 
     std::shared_ptr<common_ngram_mod> ngram_mod;
@@ -780,16 +779,6 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
-//
-// Token utils
-//
-
-// longest common prefix
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
-
-// longet common subsequence
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
-
 //
 // Vocab utils
 //

common/debug.cpp

@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
     return v;
 }
 
+#define INDENT "    "
+
 template <bool abort>
 void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
     GGML_ASSERT(n > 0);
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         }
     }
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG_ERR("                                     [\n");
+        LOG(INDENT "[\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2 * n) {
-                LOG_ERR("                                      ..., \n");
+                LOG(INDENT INDENT "..., \n");
                 i2 = ne[2] - n;
             }
-            LOG_ERR("                                      [\n");
+            LOG(INDENT INDENT "[\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2 * n) {
-                    LOG_ERR("                                       ..., \n");
+                    LOG(INDENT INDENT INDENT "..., \n");
                     i1 = ne[1] - n;
                 }
-                LOG_ERR("                                       [");
+                LOG(INDENT INDENT INDENT "[");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2 * n) {
-                        LOG_ERR("..., ");
+                        LOG("   ..., ");
                         i0 = ne[0] - n;
                     }
                     const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
-                    LOG_ERR("%12.4f", v);
+                    LOG("%12.4f", v);
                     if (i0 < ne[0] - 1) {
-                        LOG_ERR(", ");
+                        LOG(", ");
                     }
                 }
-                LOG_ERR("],\n");
+                LOG("  ],\n");
             }
-            LOG_ERR("                                      ],\n");
+            LOG(INDENT INDENT "],\n");
         }
-        LOG_ERR("                                     ]\n");
-        LOG_ERR("                                     sum = %f\n", sum);
+        LOG(INDENT "]\n");
+        LOG(INDENT "sum = %f\n", sum);
     }
 
     if constexpr (abort) {
         if (std::isnan(sum)) {
-            LOG_ERR("encountered NaN - aborting\n");
+            LOG("encountered NaN - aborting\n");
             exit(0);
         }
     }
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
     }
 
     if (matches_filter) {
-        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
-                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
-                common_ggml_ne_string(t).c_str());
+        LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+            ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+            common_ggml_ne_string(t).c_str());
     }
 
     const bool is_host = ggml_backend_buffer_is_host(t->buffer);

common/download.cpp

@@ -19,9 +19,7 @@
 #include <thread>
 #include <vector>
 
-#if defined(LLAMA_USE_HTTPLIB)
 #include "http.h"
-#endif
 
 #ifndef __EMSCRIPTEN__
 #ifdef __linux__
@@ -114,44 +112,18 @@ static void write_etag(const std::string & path, const std::string & etag) {
 }
 
 static std::string read_etag(const std::string & path) {
-    std::string none;
     const std::string etag_path = path + ".etag";
-
-    if (std::filesystem::exists(etag_path)) {
-        std::ifstream etag_in(etag_path);
-        if (!etag_in) {
-            LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
-            return none;
-        }
-        std::string etag;
-        std::getline(etag_in, etag);
-        return etag;
+    if (!std::filesystem::exists(etag_path)) {
+        return {};
     }
-
-    // no etag file, but maybe there is an old .json
-    // remove this code later
-    const std::string metadata_path = path + ".json";
-
-    if (std::filesystem::exists(metadata_path)) {
-        std::ifstream metadata_in(metadata_path);
-        try {
-            nlohmann::json metadata_json;
-            metadata_in >> metadata_json;
-            LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
-                    metadata_json.dump().c_str());
-            if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
-                std::string etag = metadata_json.at("etag");
-                write_etag(path, etag);
-                if (!std::filesystem::remove(metadata_path)) {
-                    LOG_WRN("%s: failed to delete old .json metadata file: %s\n", __func__, metadata_path.c_str());
-                }
-                return etag;
-            }
-        } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-        }
+    std::ifstream etag_in(etag_path);
+    if (!etag_in) {
+        LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
+        return {};
     }
-    return none;
+    std::string etag;
+    std::getline(etag_in, etag);
+    return etag;
 }
 
 static bool is_http_status_ok(int status) {
@@ -168,8 +140,6 @@ std::pair<std::string, std::string> common_download_split_repo_tag(const std::st
     return {hf_repo, tag};
 }
 
-#if defined(LLAMA_USE_HTTPLIB)
-
 class ProgressBar {
     static inline std::mutex mutex;
     static inline std::map<const ProgressBar *, int> lines;
@@ -305,7 +275,10 @@ static bool common_pull_file(httplib::Client & cli,
     );
 
     if (!res) {
-        LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
+        LOG_ERR("%s: download failed: %s (status: %d)\n",
+                __func__,
+                httplib::to_string(res.error()).c_str(),
+                res ? res->status : -1);
         return false;
     }
 
@@ -344,62 +317,64 @@ static int common_download_file_single_online(const std::string        & url,
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
 
-    for (int i = 0; i < max_attempts; ++i) {
-        auto head = cli.Head(parts.path);
-        bool head_ok = head && head->status >= 200 && head->status < 300;
-        if (!head_ok) {
-            LOG_WRN("%s: HEAD invalid http status code received: %d\n", __func__, head ? head->status : -1);
-            if (file_exists) {
-                LOG_INF("%s: Using cached file (HEAD failed): %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            return head->status; // cannot use cached file, return raw status code
-            // TODO: maybe retry only on certain codes
-        }
-
-        std::string etag;
-        if (head_ok && head->has_header("ETag")) {
-            etag = head->get_header_value("ETag");
-        }
-
-        size_t total_size = 0;
-        if (head_ok && head->has_header("Content-Length")) {
-            try {
-                total_size = std::stoull(head->get_header_value("Content-Length"));
-            } catch (const std::exception& e) {
-                LOG_WRN("%s: Invalid Content-Length in HEAD response: %s\n", __func__, e.what());
-            }
-        }
-
-        bool supports_ranges = false;
-        if (head_ok && head->has_header("Accept-Ranges")) {
-            supports_ranges = head->get_header_value("Accept-Ranges") != "none";
-        }
-
-        bool should_download_from_scratch = false;
-        if (!last_etag.empty() && !etag.empty() && last_etag != etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
-                    last_etag.c_str(), etag.c_str());
-            should_download_from_scratch = true;
-        }
-
+    auto head = cli.Head(parts.path);
+    if (!head || head->status < 200 || head->status >= 300) {
+        LOG_WRN("%s: HEAD failed, status: %d\n", __func__, head ? head->status : -1);
         if (file_exists) {
-            if (!should_download_from_scratch) {
-                LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return -1;
-            }
+            LOG_INF("%s: using cached file (HEAD failed): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        return head ? head->status : -1;
+    }
+
+    std::string etag;
+    if (head->has_header("ETag")) {
+        etag = head->get_header_value("ETag");
+    }
+
+    size_t total_size = 0;
+    if (head->has_header("Content-Length")) {
+        try {
+            total_size = std::stoull(head->get_header_value("Content-Length"));
+        } catch (const std::exception& e) {
+            LOG_WRN("%s: invalid Content-Length in HEAD response: %s\n", __func__, e.what());
+        }
+    }
+
+    bool supports_ranges = false;
+    if (head->has_header("Accept-Ranges")) {
+        supports_ranges = head->get_header_value("Accept-Ranges") != "none";
+    }
+
+    if (file_exists) {
+        if (etag.empty()) {
+            LOG_INF("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (!last_etag.empty() && last_etag == etag) {
+            LOG_INF("%s: using cached file (same etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (remove(path.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+            return -1;
+        }
+    }
+
+    const std::string path_temporary = path + ".downloadInProgress";
+    int delay = retry_delay_seconds;
+
+    for (int i = 0; i < max_attempts; ++i) {
+        if (i) {
+            LOG_WRN("%s: retrying after %d seconds...\n", __func__, delay);
+            std::this_thread::sleep_for(std::chrono::seconds(delay));
+            delay *= retry_delay_seconds;
         }
 
-        const std::string path_temporary = path + ".downloadInProgress";
         size_t existing_size = 0;
 
         if (std::filesystem::exists(path_temporary)) {
-            if (supports_ranges && !should_download_from_scratch) {
+            if (supports_ranges) {
                 existing_size = std::filesystem::file_size(path_temporary);
             } else if (remove(path_temporary.c_str()) != 0) {
                 LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
@@ -407,32 +382,23 @@ static int common_download_file_single_online(const std::string        & url,
             }
         }
 
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (etag:%s)...\n",
-                __func__, common_http_show_masked_url(parts).c_str(), path_temporary.c_str(), etag.c_str());
-        const bool was_pull_successful = common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size);
-        if (!was_pull_successful) {
-            if (i + 1 < max_attempts) {
-                const int exponential_backoff_delay = std::pow(retry_delay_seconds, i) * 1000;
-                LOG_WRN("%s: retrying after %d milliseconds...\n", __func__, exponential_backoff_delay);
-                std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-            } else {
-                LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
+        LOG_INF("%s: downloading from %s to %s (etag:%s)...\n",
+                __func__, common_http_show_masked_url(parts).c_str(),
+                path_temporary.c_str(), etag.c_str());
+
+        if (common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size)) {
+            if (std::rename(path_temporary.c_str(), path.c_str()) != 0) {
+                LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+                return -1;
             }
-            continue;
+            if (!etag.empty()) {
+                write_etag(path, etag);
+            }
+            return head->status;
         }
-
-        if (std::rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return -1;
-        }
-        if (!etag.empty()) {
-            write_etag(path, etag);
-        }
-
-        return head->status; // TODO: use actual GET status?
     }
 
+    LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
     return -1; // max attempts reached
 }
 
@@ -798,30 +764,6 @@ std::string common_docker_resolve_model(const std::string & docker) {
     }
 }
 
-#else
-
-common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-bool common_download_model(const common_params_model &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-std::string common_docker_resolve_model(const std::string &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-int common_download_file_single(const std::string &,
-                                const std::string &,
-                                const std::string &,
-                                bool,
-                                const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-#endif // defined(LLAMA_USE_HTTPLIB)
-
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();

common/jinja/caps.cpp

@@ -63,7 +63,8 @@ static void caps_print_stats(value & v, const std::string & path) {
 
 std::map<std::string, bool> caps::to_map() const {
     return {
-        {"requires_typed_content", requires_typed_content},
+        {"supports_string_content", supports_string_content},
+        {"supports_typed_content", supports_typed_content},
         {"supports_tools", supports_tools},
         {"supports_tool_calls", supports_tool_calls},
         {"supports_parallel_tool_calls", supports_parallel_tool_calls},
@@ -89,7 +90,7 @@ caps caps_get(jinja::program & prog) {
         return v->stats.ops.find(op_name) != v->stats.ops.end();
     };
 
-    // case: typed content requirement
+    // case: typed content support
     caps_try_execute(
         prog,
         [&]() {
@@ -105,12 +106,16 @@ caps caps_get(jinja::program & prog) {
             // tools
             return json{nullptr};
         },
-        [&](bool, value & messages, value &) {
+        [&](bool success, value & messages, value &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
             if (has_op(content, "selectattr") || has_op(content, "array_access")) {
                 // accessed as an array
-                result.requires_typed_content = true;
+                result.supports_typed_content = true;
+            }
+            if (!success) {
+                // failed to execute with content as string
+                result.supports_string_content = false;
             }
         }
     );

common/jinja/caps.h

@@ -14,7 +14,9 @@ struct caps {
     bool supports_parallel_tool_calls = true;
     bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
 
-    bool requires_typed_content = false; // default: use string content
+    // one of the 2 content capabilities must be true
+    bool supports_string_content = true;
+    bool supports_typed_content = false;
 
     // for reporting on server
     std::map<std::string, bool> to_map() const;

common/jinja/runtime.cpp

@@ -144,6 +144,13 @@ value binary_expression::execute_impl(context & ctx) {
         return false;
     };
 
+    auto test_is_in = [&]() -> bool {
+        func_args args(ctx);
+        args.push_back(left_val);
+        args.push_back(right_val);
+        return global_builtins().at("test_is_in")(args)->as_bool();
+    };
+
     // Handle undefined and null values
     if (is_val<value_undefined>(left_val) || is_val<value_undefined>(right_val)) {
         if (is_val<value_undefined>(right_val) && (op.value == "in" || op.value == "not in")) {
@@ -223,19 +230,11 @@ value binary_expression::execute_impl(context & ctx) {
             return result;
         }
     } else if (is_val<value_array>(right_val)) {
-        auto & arr = right_val->as_array();
-        bool member = false;
-        for (const auto & item : arr) {
-            if (*left_val == *item) {
-                member = true;
-                break;
-            }
-        }
+        // case: 1 in [0, 1, 2]
+        bool member = test_is_in();
         if (op.value == "in") {
-            JJ_DEBUG("Checking membership: %s in Array is %d", left_val->type().c_str(), member);
             return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            JJ_DEBUG("Checking non-membership: %s not in Array is %d", left_val->type().c_str(), !member);
             return mk_val<value_bool>(!member);
         }
     }
@@ -252,22 +251,23 @@ value binary_expression::execute_impl(context & ctx) {
 
     // String membership
     if (is_val<value_string>(left_val) && is_val<value_string>(right_val)) {
-        auto left_str = left_val->as_string().str();
-        auto right_str = right_val->as_string().str();
+        // case: "a" in "abc"
+        bool member = test_is_in();
         if (op.value == "in") {
-            return mk_val<value_bool>(right_str.find(left_str) != std::string::npos);
+            return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            return mk_val<value_bool>(right_str.find(left_str) == std::string::npos);
+            return mk_val<value_bool>(!member);
         }
     }
 
     // Value key in object
     if (is_val<value_object>(right_val)) {
-        bool has_key = right_val->has_key(left_val);
+        // case: key in {key: value}
+        bool member = test_is_in();
         if (op.value == "in") {
-            return mk_val<value_bool>(has_key);
+            return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            return mk_val<value_bool>(!has_key);
+            return mk_val<value_bool>(!member);
         }
     }
 
@@ -446,6 +446,12 @@ value for_statement::execute_impl(context & ctx) {
 
     value iterable_val = iter_expr->execute(scope);
 
+    // mark the variable being iterated as used for stats
+    if (ctx.is_get_stats) {
+        iterable_val->stats.used = true;
+        iterable_val->stats.ops.insert("array_access");
+    }
+
     if (iterable_val->is_undefined()) {
         JJ_DEBUG("%s", "For loop iterable is undefined, skipping loop");
         iterable_val = mk_val<value_array>();

common/jinja/value.cpp

@@ -393,6 +393,33 @@ const func_builtins & global_builtins() {
         {"test_is_lt", test_compare_fn<value_compare_op::lt>},
         {"test_is_lessthan", test_compare_fn<value_compare_op::lt>},
         {"test_is_ne", test_compare_fn<value_compare_op::ne>},
+        {"test_is_in", [](const func_args & args) -> value {
+            args.ensure_count(2);
+            auto needle   = args.get_pos(0);
+            auto haystack = args.get_pos(1);
+            if (is_val<value_undefined>(haystack)) {
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_array>(haystack)) {
+                for (const auto & item : haystack->as_array()) {
+                    if (*needle == *item) {
+                        return mk_val<value_bool>(true);
+                    }
+                }
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_string>(haystack)) {
+                if (!is_val<value_string>(needle)) {
+                    throw raised_exception("'in' test expects args[1] as string when args[0] is string, got args[1] as " + needle->type());
+                }
+                return mk_val<value_bool>(
+                    haystack->as_string().str().find(needle->as_string().str()) != std::string::npos);
+            }
+            if (is_val<value_object>(haystack)) {
+                return mk_val<value_bool>(haystack->has_key(needle));
+            }
+            throw raised_exception("'in' test expects iterable as first argument, got " + haystack->type());
+        }},
         {"test_is_test", [](const func_args & args) -> value {
             args.ensure_vals<value_string>();
             auto & builtins = global_builtins();

common/ngram-map.cpp

@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
  * @return Vector of draft tokens, empty if no matching pattern is found
  */
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled) {
 
     // Simple implementation of self-speculative decoding without a draft model.
     //
     const size_t cur_len = tokens.size();
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (state.idx_last_check + state.config.check_rate > cur_len) {
-        llama_tokens draft_tokens;
-        return draft_tokens;
-    }
 
-    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
-    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+    const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
+    const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
 
     // vector for tokens we want to verify.
     // return empty vector if there is no match.
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
     }
     pattern.push_back(sampled); // add the last token to the pattern
 
-    // We do a search in the token history.
-    state.idx_last_check = cur_len;
-
     size_t match_pos = 0; // we ignore position 0, position 0 == no match
                           // search backwards, but skip the current match (we are currently there)
     for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
@@ -240,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
         GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
     }
 
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (map.idx_last_check + map.check_rate > cur_len) {
-        return;
+    if (map.idx_last_check  > cur_len) {
+        // Should not happen because of common_ngram_map_begin().
+        GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
     }
     map.idx_last_check = cur_len;
 
@@ -471,7 +461,7 @@ void common_ngram_map_draft(common_ngram_map & map,
             slot_max = v;
         }
     }
-    // What is sum of the other occurences?
+    // What is sum of the other occurrences?
     uint32_t sum_occur = 0;
     for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
         if (v == slot_max) {

common/ngram-map.h

@@ -24,26 +24,11 @@
 struct common_ngram_simple_config {
     uint16_t   size_ngram;      // size of n-grams to lookup in self-mode
     uint16_t   size_mgram;      // size of m-grams to draft in self-mode
-    uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
-};
-
-// current state (and config) of n-gram simple.
-struct common_ngram_simple_state {
-    common_ngram_simple_config config;
-
-    size_t idx_last_check = 0; // index of last check in context history (mutable)
-
-    common_ngram_simple_state(const common_ngram_simple_config & config)
-        : config(config) {}
 };
 
 // Searches for a n-gram in the history and checks whether a draft sequence should be generated.
-// state:              the ngram simple state to search in.
-// inp:                the tokens generated so far.
-// sampled:            the token that was just sampled.
-// draft:              vector to store the draft tokens, initially empty.
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled);
 
 
@@ -59,7 +44,7 @@ llama_tokens common_ngram_simple_draft(
 // statistics of a m-gram after a known n-gram
 struct common_ngram_map_value {
     size_t   value_idx =  0;  // index of value m-gram in token-history (0 if unused)
-    uint16_t value_num =  0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    uint16_t value_num =  0;  // number of occurrences of this value m-gram after the key n-gram (0 in an unused values-slot)
     int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
 };
 
@@ -68,7 +53,7 @@ struct common_ngram_map_key {
     size_t   key_idx;   // index of key n-gram in token-history
     size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
 
-    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    uint16_t key_num;   // number of occurrences of this key n-gram in token-history
     common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
 };
 
@@ -80,15 +65,14 @@ struct common_ngram_map {
     bool key_only;       // true if only key n-grams are used, no values.
 
     std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
-    uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
     uint16_t min_hits;   // minimum number of key hits to consider a draft
 
-    bool     show_key_map_stats = false; // true, if statitics of the key_map should be printed.
+    bool     show_key_map_stats = false; // true, if statistics of the key_map should be printed.
 
     common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
-                     uint16_t check_rate, uint16_t min_hits)
+                     uint16_t min_hits)
         : size_key(sz_key), size_value(sz_value), key_only(only_keys),
-          check_rate(check_rate), min_hits(min_hits) {
+          min_hits(min_hits) {
         key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
     }
 

common/speculative.cpp

@@ -113,13 +113,14 @@ static bool common_speculative_are_compatible(
 struct common_speculative_state {
     const enum common_speculative_type type;
 
-    // TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
-    // TODO: add n_call_begin, n_call_accept
-    size_t drafts_call_count       = 0; // number of times this implementation was called.
-    size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
-    size_t drafts_accepted_count   = 0; // number of times a draft or part was accepted by the target model.
-    size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
-    size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
+    size_t n_call_begin  = 0; // number of times this implementation was called for refresh.
+    size_t n_call_draft  = 0; // number of times this implementation was called for generation.
+    size_t n_call_accept = 0; // number of times this implementation was called for accumulation.
+
+    size_t n_gen_drafts = 0; // number of times a draft or part was generated by this implementation.
+    size_t n_acc_drafts = 0; // number of times a draft or part was accepted by the target model.
+    size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
+    size_t n_acc_tokens = 0; // number of tokens accepted by the target model.
 
     // TODO: track performance of most recent calls
     const bool gen_perf = true; // whether to generate performance stats.
@@ -463,12 +464,12 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
 
 // state of self-speculation (simple implementation, not ngram-map)
 struct common_speculative_state_ngram_simple : public common_speculative_state {
-    common_ngram_simple_state state;
+    common_ngram_simple_config config;
 
     common_speculative_state_ngram_simple(
             enum common_speculative_type type,
-            common_ngram_simple_state state)
-        : common_speculative_state(type), state(state) {}
+            common_ngram_simple_config config)
+        : common_speculative_state(type), config(config) {}
 
     void begin(const llama_tokens & prompt) override {
         GGML_UNUSED(prompt);
@@ -479,7 +480,8 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
         GGML_UNUSED(params);
     }
 
@@ -744,10 +746,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
     uint16_t size_key   = config.params.ngram_size_n;
     uint16_t size_value = config.params.ngram_size_m;
     bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
-    uint16_t check_rate = config.params.ngram_check_rate;
     uint16_t min_hits   = config.params.ngram_min_hits;
 
-    return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
+    return common_ngram_map(size_key, size_value, key_only, min_hits);
 }
 
 static common_speculative_state_ngram_cache create_state_ngram_cache(
@@ -797,6 +798,42 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
+bool common_speculative_is_compat(llama_context * ctx_tgt) {
+    auto * mem = llama_get_memory(ctx_tgt);
+    if (mem == nullptr) {
+        return false;
+    }
+
+    bool res = true;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = false;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = false;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx_tgt);
+
+    return res;
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
@@ -887,16 +924,14 @@ common_speculative * common_speculative_init(
 
                 uint16_t ngram_size_key   = ngram_map.size_key;
                 uint16_t mgram_size_value = ngram_map.size_value;
-                uint16_t check_rate       = ngram_map.check_rate;
 
-                auto config_simple = common_ngram_simple_config{
+                auto config_simple = common_ngram_simple_config {
                     /* .size_ngram      = */ ngram_size_key,
-                    /* .size_mgram      = */ mgram_size_value,
-                    /* .check_rate      = */ check_rate
+                    /* .size_mgram      = */ mgram_size_value
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
                     /* .type            = */ config.type,
-                    /* .state           = */ common_ngram_simple_state(config_simple)
+                    /* .state           = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;
@@ -951,12 +986,9 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
     }
 
     for (auto & impl : spec->impls) {
-        const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+        common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
         impl->begin(prompt);
-
-        const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-        impl->t_begin_us += t_now_us - t_start_us; // accumulate duration for this refresh
+        impl->n_call_begin++;
     }
 }
 
@@ -971,24 +1003,19 @@ llama_tokens common_speculative_draft(
 
     for (auto & impl : spec->impls) {
         {
-            const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+            common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
             impl->draft(params, prompt_tgt, id_last, result);
-
-            const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-
-            impl->drafts_call_count++;
-            impl->t_draft_us += t_now_us - t_start_us; // accumulate duration for this implementation
+            impl->n_call_draft++;
         }
 
         if (!result.empty()) {
             LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                     common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
-                    impl.get()->drafts_call_count, result.size());
+                    impl.get()->n_call_draft, result.size());
 
             spec->curr_impl = impl.get(); // set current implementation for stats
-            impl->drafts_generated_count++;
-            impl->drafts_generated_tokens += result.size();
+            impl->n_gen_drafts++;
+            impl->n_gen_tokens += result.size();
 
             break; // We have a draft, so break out of the loop and return it.
         }
@@ -1006,15 +1033,16 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
 
     GGML_ASSERT(impl);
 
-    const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-    if (n_accepted > 0) {
-        impl->drafts_accepted_count++;
-        impl->drafts_accepted_tokens += n_accepted;
-    }
+    {
+        common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
+        if (n_accepted > 0) {
+            impl->n_acc_drafts++;
+            impl->n_acc_tokens += n_accepted;
+        }
 
-    impl->accept(n_accepted);
-    const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-    impl->t_accept_us += t_now_us - t_start_us; // accumulate duration for this acculumulation
+        impl->accept(n_accepted);
+        impl->n_call_accept++;
+    }
 }
 
 void common_speculative_print_stats(const common_speculative * spec) {
@@ -1034,13 +1062,13 @@ void common_speculative_print_stats(const common_speculative * spec) {
             str_perf = "";
         }
 
-        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+        LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
                 common_speculative_type_to_str(impl->type).c_str(),
-                impl->drafts_call_count,
-                impl->drafts_generated_count,
-                impl->drafts_accepted_count,
-                impl->drafts_generated_tokens,
-                impl->drafts_accepted_tokens,
+                impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
+                impl->n_gen_drafts,
+                impl->n_acc_drafts,
+                impl->n_gen_tokens,
+                impl->n_acc_tokens,
                 str_perf.c_str());
     }
 }

common/speculative.h

@@ -14,6 +14,10 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
+// check if the llama_context is compatible for speculative decoding
+// note: clears the memory of the context
+bool common_speculative_is_compat(llama_context * ctx_tgt);
+
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);

convert_hf_to_gguf_update.py

@@ -148,6 +148,7 @@ models = [
     {"name": "youtu",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
     {"name": "solar-open",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
     {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
+    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", }
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/backend/SYCL.md

@@ -22,12 +22,11 @@
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
-- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
 ### Llama.cpp + SYCL
 
 The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
-SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
+SYCL cross-platform capabilities enable support for other vendor GPUs as well.
 
 ## Recommended Release
 
@@ -42,6 +41,9 @@ The following releases are verified and recommended:
 
 ## News
 
+- 2026.02
+  - Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. User can't build up the software for Nvidia & AMD GPU.
+
 - 2025.11
   - Support malloc memory on device more than 4GB.
 
@@ -111,8 +113,8 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc A-Series              | Support | Arc A770, Arc A730M, Arc A750         |
-| Intel Arc B-Series              | Support | Arc B580         |
+| Intel Arc A-Series            | Support | Arc A770, Arc A730M, Arc A750         |
+| Intel Arc B-Series            | Support | Arc B580                              |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
 | Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
 
@@ -127,20 +129,7 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 
 ### Other Vendor GPU
 
-**Verified devices**
-
-| Nvidia GPU               | Status    | Verified Model |
-|--------------------------|-----------|----------------|
-| Ampere Series            | Supported | A100, A4000    |
-| Ampere Series *(Mobile)* | Supported | RTX 40 Series  |
-
-| AMD GPU                  | Status       | Verified Model |
-|--------------------------|--------------|----------------|
-| Radeon Pro               | Experimental | W6800          |
-| Radeon RX                | Experimental | 6700 XT        |
-
-Note: AMD GPU support is highly experimental and is incompatible with F16.
-Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
+NA
 
 ## Docker
 
@@ -149,11 +138,11 @@ The docker build option is currently limited to *Intel GPU* targets.
 ### Build image
 
 ```sh
-# Using FP16
-docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
-
 # Using FP32
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=OFF" --target light -f .devops/intel.Dockerfile .
+
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
 ```
 
 *Notes*:
@@ -212,14 +201,6 @@ Platform #0: Intel(R) OpenCL HD Graphics
  `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```
 
-- **Nvidia GPU**
-
-In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
-
-- **AMD GPU**
-
-To target AMD GPUs with SYCL, the ROCm stack must be installed first.
-
 2. **Install Intel® oneAPI Base toolkit**
 
 SYCL backend depends on:
@@ -248,23 +229,6 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 |2025.1|
 |2024.1|
 
-- **Adding support to Nvidia GPUs**
-
-**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
-
-**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
-
-```sh
-git clone https://github.com/oneapi-src/oneDNN.git
-cd oneDNN
-cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-cmake --build build-nvidia --config Release
-```
-
-- **Adding support to AMD GPUs**
-
-**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
-
 3. **Verify installation and environment**
 
 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -285,25 +249,6 @@ When targeting an intel GPU, the user should expect one or more devices among th
 [opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) UHD Graphics 730 OpenCL 3.0 NEO  [24.39.31294]
 ```
 
-- **Nvidia GPU**
-
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
-
-```
-[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
-[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
-[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
-```
-
-- **AMD GPU**
-
-For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
-
-```
-[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
-[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
-```
-
 ### II. Build llama.cpp
 
 #### Intel GPU
@@ -332,47 +277,6 @@ It is possible to come across some precision issues when running tests that stem
 instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS`
 as `-cl-fp32-correctly-rounded-divide-sqrt`
 
-#### Nvidia GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with Nvidia BLAS acceleration through SYCL
-# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
-GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
-
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
-It is possible to come across some precision issues when running tests that stem from using faster
-instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler.
-
-#### AMD GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with rocBLAS acceleration through SYCL
-
-## AMD
-# Use FP32, FP16 is not supported
-# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
-GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
 ### III. Run the inference
 
 #### Retrieve and prepare model
@@ -766,15 +670,15 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
-| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
-| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
+| GGML_SYCL_TARGET   | INTEL *(default)*                     | Set the SYCL target device type.            |
+| GGML_SYCL_DEVICE_ARCH | Optional                           | Set the SYCL device architecture. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
-| GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
+| GGML_SYCL_GRAPH    | OFF *(default)* \|ON *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
-1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.
 
 #### Runtime
 
@@ -782,7 +686,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
-| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|

docs/backend/VirtGPU.md

@@ -0,0 +1,180 @@
+# GGML-VirtGPU Backend
+
+The GGML-VirtGPU backend enables GGML applications to run machine
+learning computations on host hardware while the application itself
+runs inside a virtual machine.  It uses host-guest shared memory to
+efficiently share data buffers between the two sides.
+
+This backend relies on the virtio-gpu, and VirglRenderer API Remoting
+(APIR) component. The backend is split into two libraries:
+- a GGML implementation (the "remoting frontend"), running in the
+  guest and interacting with the virtgpu device
+- a VirglRenderer APIR compatible library (the "remoting backend"),
+  running in the host and interacting with Virglrenderer and an actual
+  GGML device backend.
+
+## OS support
+
+| OS       | Status            | Backend     | CI testing  | Notes
+| -------- | ----------------- | ----------- | ----------- | -----
+| MacOS 14 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14
+| MacOS 15 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14 or MacOS 15
+| MacOS 26 | Not tested        |             |             |
+| Linux    | Under development | ggml-vulkan | not working | Working locally, CI running into deadlocks
+
+
+## Architecture Overview
+
+The GGML-VirtGPU backend consists of three main components:
+
+```mermaid
+graph TD
+    %% Nodes
+
+ subgraph GuestVM ["Guest VM - Frontend"]
+        App([GGML Application<br/>llama.cpp, etc.])
+
+        direction TB
+        Interface[GGML Backend Interface]
+        Comm["GGML-VirtGPU<br/>(hypercalls + shared mem)"]
+
+        App --> Interface
+        Interface --> Comm
+    end
+
+    API[virtio-gpu / virglrenderer API]
+
+    subgraph HostSystem [Host System - Backend]
+        direction TB
+        Dispatcher[GGML-VirtGPU-Backend]
+        BackendLib[GGML Backend library<br/>Metal / Vulkan / CPU / ...]
+
+        Dispatcher --> BackendLib
+    end
+
+    %% Connections
+    Comm --> API
+    API --> HostSystem
+```
+
+### Key Components
+
+1. **Guest-side Frontend** (`ggml-virtgpu/`): Implements the GGML backend interface and forwards operations to the host
+2. **Host-side Backend** (`ggml-virtgpu/backend/`): Receives forwarded operations and executes them on actual hardware backends
+3. **Communication Layer**: Uses virtio-gpu hypercalls and shared memory for efficient data transfer
+
+## Features
+
+- **Dynamic backend loading** on the host side (CPU, CUDA, Metal, etc.)
+- **Zero-copy data transfer** via host-guest shared memory pages
+
+## Communication Protocol
+
+### Hypercalls and Shared Memory
+
+The backend uses two primary communication mechanisms:
+
+1. **Hypercalls (`DRM_IOCTL_VIRTGPU_EXECBUFFER`)**: Trigger remote execution from guest to host
+2. **Shared Memory Pages**: Zero-copy data transfer for tensors and parameters
+
+#### Shared Memory Layout
+
+Each connection uses two shared memory buffers:
+
+- **Data Buffer** (24 MiB): For command/response data and tensor transfers
+- **Reply Buffer** (16 KiB): For command replies and status information
+- **Data Buffers**: Dynamically allocated host-guest shared buffers
+  served as GGML buffers.
+
+### APIR Protocol
+
+The Virglrender API Remoting protocol defines three command types:
+
+- `HANDSHAKE`: Protocol version negotiation and capability discovery
+- `LOADLIBRARY`: Dynamic loading of backend libraries on the host
+- `FORWARD`: API function call forwarding
+
+### Binary Serialization
+
+Commands and data are serialized using a custom binary protocol with:
+
+- Fixed-size encoding for basic types
+- Variable-length arrays with size prefixes
+- Buffer bounds checking
+- Error recovery mechanisms
+
+## Supported Operations
+
+### Device Operations
+- Device enumeration and capability queries
+- Memory information (total/free)
+- Backend type detection
+
+### Buffer Operations
+- Buffer allocation and deallocation
+- Tensor data transfer (host ↔ guest)
+- Memory copying and clearing
+
+### Computation Operations
+- Graph execution forwarding
+
+## Build Requirements
+
+### Guest-side Dependencies
+- `libdrm` for DRM/virtio-gpu communication
+- C++20 compatible compiler
+- CMake 3.14+
+
+### Host-side Dependencies
+- virglrenderer with APIR support (pending upstream review)
+- Target backend libraries (libggml-metal, libggml-vulkan, etc.)
+
+## Configuration
+
+### Environment Variables
+
+- `GGML_VIRTGPU_BACKEND_LIBRARY`: Path to the host-side backend library
+- `GGML_VIRTGPU_DEBUG`: Enable debug logging
+
+### Build Options
+
+- `GGML_VIRTGPU`: Enable the VirtGPU backend (`ON` or `OFF`, default: `OFF`)
+- `GGML_VIRTGPU_BACKEND`: Build the host-side backend component (`ON`, `OFF` or `ONLY`, default: `OFF`)
+
+### System Requirements
+
+- VM with virtio-gpu support
+- VirglRenderer with APIR patches
+- Compatible backend libraries on host
+
+## Limitations
+
+- **VM-specific**: Only works in virtual machines with virtio-gpu support
+- **Host dependency**: Requires properly configured host-side backend
+- **Latency**: Small overhead from VM escaping for each operation
+
+
+* This work is pending upstream changes in the VirglRenderer
+  project.
+  * The backend can be tested with Virglrenderer compiled from source
+  using this PR:
+  https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590
+* This work is pending changes in the VMM/hypervisor running the
+  virtual machine, which need to know how to route the newly
+  introduced APIR capset.
+  * The environment variable `VIRGL_ROUTE_VENUS_TO_APIR=1` allows
+    using the Venus capset, until the relevant hypervisors have been
+    patched. However, setting this flag breaks the Vulkan/Venus normal
+    behavior.
+  * The environment variable `GGML_REMOTING_USE_APIR_CAPSET` tells the
+    `ggml-virtgpu` backend to use the APIR capset. This will become
+    the default when the relevant hypervisors have been patched.
+
+* This work focused on improving the performance of llama.cpp running
+  on MacOS containers, and is mainly tested on this platform. The
+  linux support (via `krun`) is in progress.
+
+## See Also
+
+- [Development and Testing](VirtGPU/development.md)
+- [Backend configuration](VirtGPU/configuration.md)

docs/backend/VirtGPU/configuration.md

@@ -0,0 +1,174 @@
+# GGML-VirtGPU Backend Configuration
+
+This document describes the environment variables used by the ggml-virtgpu backend system, covering both the frontend (guest-side) and backend (host-side) components.
+
+## Environment Variables Overview
+
+The ggml-virtgpu backend uses environment variables for configuration across three main components:
+- **Frontend (Guest)**: GGML applications running in VMs
+- **Hypervisor**: Virglrenderer/APIR system
+- **Backend (Host)**: Host-side GGML backend integration
+
+## Frontend (Guest-side) Configuration
+
+### GGML_REMOTING_USE_APIR_CAPSET
+- **Location**: `ggml/src/ggml-virtgpu/virtgpu.cpp`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Controls which virtio-gpu capability set to use for communication
+- **Values**:
+  - Set (any value): Use the APIR capset (long-term setup)
+  - Unset: Use the Venus capset (easier for testing with an unmodified hypervisor)
+- **Default**: Unset (Venus capset)
+- **Usage**:
+  ```bash
+  export GGML_REMOTING_USE_APIR_CAPSET=1  # Use APIR capset
+  # or leave unset for Venus capset
+  ```
+
+## Hypervisor (Virglrenderer/APIR) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### VIRGL_APIR_BACKEND_LIBRARY
+- **Location**: `virglrenderer/src/apir/apir-context.c`
+- **Configuration Key**: `apir.load_library.path`
+- **Type**: File path string
+- **Purpose**: Path to the APIR backend library that virglrenderer should dynamically load
+- **Required**: Yes
+- **Example**:
+  ```bash
+  export VIRGL_APIR_BACKEND_LIBRARY="/path/to/libggml-remotingbackend.so"
+  ```
+
+### VIRGL_ROUTE_VENUS_TO_APIR
+- **Location**: `virglrenderer/src/apir/apir-renderer.h`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Temporary workaround to route Venus capset calls to APIR during hypervisor transition period
+- **Status**: will be removed once hypervisors support APIR natively
+- **Warning**: Breaks normal Vulkan/Venus functionality
+- **Usage**:
+  ```bash
+  export VIRGL_ROUTE_VENUS_TO_APIR=1  # For testing with an unmodified hypervisor
+  ```
+
+### VIRGL_APIR_LOG_TO_FILE
+- **Location**: `virglrenderer/src/apir/apir-renderer.c`
+- **Environment Variable**: `VIRGL_APIR_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the VirglRenderer APIR component to specified file
+- **Required**: No (optional debugging)
+- **Default**: Logging to `stderr`
+- **Usage**:
+  ```bash
+  export VIRGL_APIR_LOG_TO_FILE="/tmp/apir-debug.log"
+  ```
+
+## Backend (Host-side) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_PATH
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+- **Configuration Key**: `ggml.library.path`
+- **Type**: File path string
+- **Purpose**: Path to the actual GGML backend library (Metal, CUDA, Vulkan, etc.)
+- **Required**: **Yes** - backend initialization fails without this
+- **Examples**:
+  ```bash
+  # macOS with Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+
+  # Linux with CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-cuda.so"
+
+  # macOS or Linux with Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-vulkan.so"
+  ```
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_REG
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+- **Configuration Key**: `ggml.library.reg`
+- **Type**: Function symbol name string
+- **Purpose**: Name of the backend registration function to call after loading the library
+- **Required**: No (defaults to `ggml_backend_init`)
+- **Default**: `ggml_backend_init`
+- **Examples**:
+  ```bash
+  # Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+  # CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_cuda_reg"
+
+  # Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_vulkan_reg"
+
+  # Generic fallback (default)
+  # export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_init"
+  ```
+
+### APIR_LLAMA_CPP_LOG_TO_FILE
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp:62`
+- **Environment Variable**: `APIR_LLAMA_CPP_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the GGML backend to specified file
+- **Required**: No (optional debugging)
+- **Usage**:
+  ```bash
+  export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml-backend-debug.log"
+  ```
+
+## Configuration Flow
+
+The configuration system works as follows:
+
+1. **Hypervisor Setup**: Virglrenderer loads the APIR backend library specified by `VIRGL_APIR_BACKEND_LIBRARY`
+
+2. **Context Creation**: When an APIR context is created, it populates a configuration table with environment variables:
+   - `apir.load_library.path` ← `VIRGL_APIR_BACKEND_LIBRARY`
+   - `ggml.library.path` ← `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+   - `ggml.library.reg` ← `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+   - this step will eventually be performed by the hypervisor itself, with command-line arguments instead of environment variables.
+
+3. **Backend Initialization**: The backend queries the configuration via callbacks:
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.path")` returns the library path
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.reg")` returns the registration function
+
+4. **Library Loading**: The backend dynamically loads and initializes the specified GGML library
+
+## Error Messages
+
+Common error scenarios and their messages:
+
+- **Missing library path**: `"cannot open the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_PATH' not defined"`
+- **Missing registration function**: `"cannot register the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_REG' not defined"`
+
+## Example Complete Configuration
+
+Here's an example configuration for a macOS host with Metal backend:
+
+```bash
+# Hypervisor environment
+export VIRGL_APIR_BACKEND_LIBRARY="/opt/llama.cpp/lib/libggml-virtgpu-backend.dylib"
+
+# Backend configuration
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+# Optional logging
+export VIRGL_APIR_LOG_TO_FILE="/tmp/apir.log"
+export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml.log"
+
+# Guest configuration
+export GGML_REMOTING_USE_APIR_CAPSET=1
+```

docs/backend/VirtGPU/development.md

@@ -0,0 +1,220 @@
+# Development and Testing
+
+## Development
+
+### Code Generation
+
+The backend uses code generation from YAML configuration:
+
+```bash
+# Regenerate protocol code
+cd ggml-virtgpu/
+python regenerate_remoting.py
+```
+
+### Adding New Operations
+
+1. Add function definition to `ggmlremoting_functions.yaml`
+2. Regenerate code with `regenerate_remoting.py`
+3. Implement guest-side forwarding in `virtgpu-forward-*.cpp`
+4. Implement host-side handling in `backend-dispatched-*.cpp`
+
+## Testing
+
+This document provides instructions for building and testing the GGML-VirtGPU backend on macOS with containers.
+
+### Prerequisites
+
+The testing setup requires:
+
+- macOS host system
+- Container runtime with `libkrun` provider (podman machine)
+- Access to development patchset for VirglRenderer
+
+### Required Patchsets
+
+The backend requires patches that are currently under review:
+
+- **Virglrenderer APIR upstream PR**: https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590 (for reference)
+- **MacOS Virglrenderer (for krunkit)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-macos
+- **Linux Virglrenderer (for krun)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-linux
+
+### Build Instructions
+
+#### 1. Build ggml-virtgpu-backend (Host-side, macOS)
+
+```bash
+# Build the backend that runs natively on macOS
+mkdir llama.cpp
+cd llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
+
+cmake -S . -B $LLAMA_MAC_BUILD \
+      -DGGML_NATIVE=OFF \
+      -DLLAMA_CURL=ON \
+      -DGGML_REMOTINGBACKEND=ONLY \
+      -DGGML_METAL=ON
+
+TARGETS="ggml-metal"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $TARGETS
+
+# Build additional tools for native benchmarking
+EXTRA_TARGETS="llama-run llama-bench"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
+```
+
+#### 2. Build virglrenderer (Host-side, macOS)
+
+```bash
+# Build virglrenderer with APIR support
+mkdir virglrenderer
+git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
+cd src
+
+VIRGL_BUILD_DIR=$PWD/build
+
+# -Dvenus=true and VIRGL_ROUTE_VENUS_TO_APIR=1 route the APIR requests via the Venus backend, for easier testing without a patched hypervisor
+
+meson setup $VIRGL_BUILD_DIR \
+      -Dvenus=true \
+      -Dapir=true
+
+ninja -C $VIRGL_BUILD_DIR
+```
+
+#### 3. Build ggml-virtgpu (Guest-side, Linux)
+
+Option A: Build from a script:
+
+```bash
+# Inside a Linux container
+mkdir llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_LINUX_BUILD=$PWD//build-virtgpu
+
+cmake -S . -B $LLAMA_LINUX_BUILD \
+      -DGGML_VIRTGPU=ON
+
+ninja -C $LLAMA_LINUX_BUILD
+```
+
+Option B: Build container image with frontend:
+
+```bash
+cat << EOF > remoting.containerfile
+FROM quay.io/fedora/fedora:43
+USER 0
+
+WORKDIR /app/remoting
+
+ARG LLAMA_CPP_REPO="https://github.com/ggml-org/llama.cpp.git"
+ARG LLAMA_CPP_VERSION="master"
+ARG LLAMA_CPP_CMAKE_FLAGS="-DGGML_VIRTGPU=ON"
+ARG LLAMA_CPP_CMAKE_BUILD_FLAGS="--parallel 4"
+
+RUN dnf install -y git cmake gcc gcc-c++ libcurl-devel libdrm-devel
+
+RUN git clone "\${LLAMA_CPP_REPO}" src \\
+ && git -C src fetch origin \${LLAMA_CPP_VERSION} \\
+ && git -C src reset --hard FETCH_HEAD
+
+RUN mkdir -p build \\
+ && cd src \\
+ && set -o pipefail \\
+ && cmake -S . -B ../build \${LLAMA_CPP_CMAKE_FLAGS} \\
+ && cmake --build ../build/ \${LLAMA_CPP_CMAKE_BUILD_FLAGS}
+
+ENTRYPOINT ["/app/remoting/src/build/bin/llama-server"]
+EOF
+
+mkdir -p empty_dir
+podman build -f remoting.containerfile ./empty_dir -t localhost/llama-cpp.virtgpu
+```
+
+### Environment Setup
+
+#### Set krunkit Environment Variables
+
+```bash
+# Define the base directories (adapt these paths to your system)
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build
+LLAMA_MAC_BUILD=$HOME/remoting/llama.cpp/build-backend
+
+# For krunkit to load the custom virglrenderer library
+export DYLD_LIBRARY_PATH=$VIRGL_BUILD_DIR/src
+
+# For Virglrenderer to load the ggml-remotingbackend library
+export VIRGL_APIR_BACKEND_LIBRARY="$LLAMA_MAC_BUILD/bin/libggml-virtgpu-backend.dylib"
+
+# For llama.cpp remotingbackend to load the ggml-metal backend
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="$LLAMA_MAC_BUILD/bin/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG=ggml_backend_metal_reg
+```
+
+#### Launch Container Environment
+
+```bash
+# Set container provider to libkrun
+export CONTAINERS_MACHINE_PROVIDER=libkrun
+podman machine start
+```
+
+#### Verify Environment
+
+Confirm that krunkit is using the correct virglrenderer library:
+
+```bash
+lsof -c krunkit | grep virglrenderer
+# Expected output:
+# krunkit 50574 user  txt  REG  1,14  2273912  10849442 ($VIRGL_BUILD_DIR/src)/libvirglrenderer.1.dylib
+```
+
+### Running Tests
+
+#### Launch Test Container
+
+```bash
+# Optional model caching
+mkdir -p models
+PODMAN_CACHE_ARGS="-v models:/models --user root:root --cgroupns host --security-opt label=disable -w /models"
+
+podman run $PODMAN_CACHE_ARGS -it --rm --device /dev/dri localhost/llama-cpp.virtgpu
+```
+
+#### Test llama.cpp in Container
+
+```bash
+
+# Run performance benchmark
+/app/remoting/build/bin/llama-bench -m ./llama3.2
+```
+
+Expected output (performance may vary):
+```
+| model                          |       size |     params | backend    | ngl |          test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         pp512 |        991.30 ± 0.66 |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         tg128 |         85.71 ± 0.11 |
+```
+
+### Troubleshooting
+
+#### SSH Environment Variable Issues
+
+⚠️ **Warning**: Setting `DYLD_LIBRARY_PATH` from SSH doesn't work on macOS. Here is a workaround:
+
+**Workaround 1: Replace system library**
+```bash
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build  # ⚠️ adapt to your system
+BREW_VIRGL_DIR=/opt/homebrew/Cellar/virglrenderer/0.10.4d/lib
+VIRGL_LIB=libvirglrenderer.1.dylib
+
+cd $BREW_VIRGL_DIR
+mv $VIRGL_LIB ${VIRGL_LIB}.orig
+ln -s $VIRGL_BUILD_DIR/src/$VIRGL_LIB
+```

docs/backend/snapdragon/README.md

@@ -35,7 +35,7 @@ Adapt below build commands accordingly.
 Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
 ```
-[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+[d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
 Preset CMake variables:

docs/build-s390x.md

@@ -242,10 +242,10 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 |------------|-------------|------|-------|
 | FP32       | ✅           | ✅    | ❓     |
 | FP16       | ✅           | ✅    | ❓     |
-| BF16       | 🚫           | ✅    | ❓     |
+| BF16       | ✅           | ✅    | ❓     |
 | Q4_0       | ✅           | ❓    | ❓     |
 | Q4_1       | ✅           | ❓    | ❓     |
-| MXFP4      | 🚫           | ❓    | ❓     |
+| MXFP4      | ✅           | ❓    | ❓     |
 | Q5_0       | ✅           | ❓    | ❓     |
 | Q5_1       | ✅           | ❓    | ❓     |
 | Q8_0       | ✅           | ❓    | ❓     |
@@ -272,4 +272,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Feb 15, 2026.

docs/build.md

@@ -252,9 +252,7 @@ CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.ggu
 
 The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.
 
-**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
-
-See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
+Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
 
 ### Unified Memory
 

docs/ops.md

@@ -22,7 +22,7 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
@@ -113,7 +113,7 @@ Legend:
 |                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |

docs/ops/SYCL.csv

@@ -77,8 +77,8 @@
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -161,8 +161,8 @@
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -9677,168 +9677,168 @@
 "SYCL0","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","1","yes","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=100,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=100,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=500,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9847,16 +9847,16 @@
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9865,16 +9865,16 @@
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9883,16 +9883,16 @@
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9901,16 +9901,16 @@
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9919,16 +9919,16 @@
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9937,51 +9937,51 @@
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL"

docs/speculative.md

@@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
                                         of lookup n-gram (default: 12)
 --spec-ngram-size-m N                   ngram size M for ngram-simple/ngram-map speculative decoding, length
                                         of draft m-gram (default: 48)
---spec-ngram-check-rate N               ngram check rate for ngram-simple/ngram-map speculative decoding
-                                        (default: 1)
 --spec-ngram-min-hits N                 minimum hits for ngram-map speculative decoding (default: 1)
 ```
 
@@ -153,10 +151,6 @@ Sets the size M of the draft m-gram for n-gram map based speculative decoding.
 The m-gram size determines how many tokens to draft when a match is found.
 Larger values can provide more speedup but may reduce acceptance rate.
 
-### `--spec-ngram-check-rate R`
-
-This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
-
 ### `--spec-ngram-min-hits H`
 
 This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
@@ -175,7 +169,12 @@ draft acceptance rate = 0.70312 (   90 accepted /   128 generated)
 statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
 ```
 
-- `#calls`: number of calls of this implementations
+```
+statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
+```
+
+
+- `#calls(b,g,a)`: number of calls of begin (new prompt), generation and accumulation of this implementations
 - `#gen drafts`: number of drafts generated by this implementation
 - `#acc drafts`: number of drafts accepted (partially) by the main model
 - `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)

examples/model-conversion/scripts/utils/tensor-info.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+from safetensors import safe_open
+
+
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
+
+
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX
+
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})
+
+    return None
+
+
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return list(weight_map.keys())
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(model_path: Path, unique: bool = False):
+    tensor_names = get_all_tensor_names(model_path)
+
+    if unique:
+        seen = set()
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen.add(normalized)
+                print(normalized)
+    else:
+        for tensor_name in sorted(tensor_names):
+            print(tensor_name)
+
+
+def print_tensor_info(model_path: Path, tensor_name: str):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",  # optional (if --list is used for example)
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list",
+        action="store_true",
+        help="List unique tensor patterns in the model (layer numbers replaced with #)"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list:
+        list_all_tensors(model_path, unique=True)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name)
+
+
+if __name__ == "__main__":
+    main()

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 5)
+set(GGML_VERSION_PATCH 7)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)

ggml/include/ggml-cpu.h

@@ -19,6 +19,9 @@ extern "C" {
         // abort ggml_graph_compute when true
         ggml_abort_callback abort_callback;
         void *              abort_callback_data;
+
+        // use only reference implementations
+        bool use_ref;
     };
 
     // numa strategies
@@ -132,6 +135,8 @@ extern "C" {
     GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
     GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
+    GGML_BACKEND_API void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref);
+
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);

ggml/include/ggml-virtgpu.h

@@ -7,8 +7,6 @@
 extern "C" {
 #endif
 
-#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
 
 #ifdef  __cplusplus

ggml/src/ggml-backend-reg.cpp

@@ -471,9 +471,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     int best_score = 0;
     fs::path best_path;
+    std::error_code ec;
 
     for (const auto & search_path : search_paths) {
-        if (std::error_code ec; !fs::exists(search_path, ec)) {
+        if (!fs::exists(search_path, ec)) {
             if (ec) {
                 GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
             } else {
@@ -483,7 +484,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         }
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
         for (const auto & entry : dir_it) {
-            if (entry.is_regular_file()) {
+            if (entry.is_regular_file(ec)) {
                 auto filename = entry.path().filename();
                 auto ext = entry.path().extension();
                 if (filename.native().find(file_prefix) == 0 && ext == file_extension) {

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -3286,130 +3286,223 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context & ctx, ggml_tensor
 }
 
 /**
- * @brief Performs expert-specific matrix multiplication (MoE) with
- * quantized precision using the CANN backend.
+ * @brief Performs quantized matrix multiplication for Mixture of Experts (MoE)
+ * models using the CANN backend.
  *
- * This function executes a matrix multiplication operation tailored for
- * Mixture of Experts (MoE) models, where the input tensor is multiplied
- * with expert-specific quantized weight matrices. It leverages the CANN
- * backend to perform efficient low-precision computations and stores the
- * quantized result in the destination tensor `dst`.
+ * This function implements MUL_MAT_ID operation for quantized weight matrices
+ * (Q4_0 and Q8_0 formats). It selects expert-specific weight matrices based on
+ * the provided expert indices, and computes matrix multiplication using CANN's
+ * WeightQuantBatchMatmulV2 operator.
  *
- * Quantization techniques reduce memory footprint and improve performance
- * by using lower-bit representations (e.g., int8) instead of floating-point.
- * This function is designed to work with such formats and may incorporate
- * optimizations like identity-based fast paths or routing masks for sparse
- * expert selection.
+ * The function performs the following steps:
+ * 1. Converts input/output tensors to F16 format if necessary
+ * 2. Uses IndexSelect to extract expert-specific weights and scales based on indices
+ * 3. Performs quantized matrix multiplication for each expert using WeightQuantBatchMatmulV2
+ * 4. Converts output back to the target type if needed
  *
- * @param ctx The context for executing CANN backend operations.
- * @param dst The destination tensor where the quantized MoE multiplication result
- * will be stored.
+ * Tensor shapes:
+ * - dst:  [M, K, N, 1] - output tensor
+ * - src0: [D, M, A, 1] - quantized weight matrices (Q4_0 or Q8_0)
+ * - src1: [D, B, N, 1] - input activations (B = K for per-expert input, or B = 1 for broadcast)
+ * - ids:  [K, N] - expert indices for routing
  *
- * @note This function assumes quantized data types and is designed for
- * MoE architectures with potential sparse expert routing.
+ * @param ctx The CANN backend context for operation execution.
+ * @param dst The destination tensor where the multiplication result will be stored.
+ *
+ * @note Only Q4_0 and Q8_0 quantization formats are supported.
+ * @note The function handles automatic type conversion to/from F16 as needed by the hardware.
  */
 static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    // TODO: Use aclnnGroupedMatMul
-    //dst   [M, K, N, 1]
-    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
-    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
-    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+    // dst:  [M, K, N, 1]
+    // src0: [D, M, A, 1] - quantized weights
+    // src1: [D, B, N, 1] - input activations, B = K or B = 1
+    // ids:  [K, N] - expert indices
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+    ggml_tensor * ids  = dst->src[2];
 
-    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT(dst->ne[3] == 1);
+    GGML_ASSERT(src1->ne[2] == ids->ne[1]);
 
-    // copy index from npu to cpu
-    int64_t n_as  = ne02;        // A
-    int64_t n_ids = ids->ne[0];  // K
+    const int64_t        n_batches        = ids->ne[1];
+    const int64_t        n_select_experts = ids->ne[0];
+    const enum ggml_type type             = src0->type;
 
-    std::vector<char> ids_host(ggml_nbytes(ids));
-    ACL_CHECK(aclrtMemcpyAsync(ids_host.data(), ggml_nbytes(ids), ids->data, ggml_nbytes(ids),
-                               ACL_MEMCPY_DEVICE_TO_HOST, ctx.stream()));
-    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+    const int32_t group_size = QK8_0;  // Both Q4_0 and Q8_0 use group size of 32
+    GGML_ASSERT(group_size == QK4_0);
 
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *) dst->data;
+    // Calculate element size for quantized weights
+    const float weight_elem_size =
+        (type == GGML_TYPE_Q4_0) ? 0.5f :
+        (type == GGML_TYPE_Q8_0) ? 1.0f :
+                                   (GGML_ABORT("MUL_MAT_ID only supports Q4_0 and Q8_0"), 0.0f);
 
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
+    // Calculate scale offset in memory
+    const size_t weight_size     = src0->ne[0] * src0->ne[1] * src0->ne[2] * weight_elem_size;
+    const size_t scale_elem_size = sizeof(uint16_t);
+    char *       scale_data      = (char *) src0->data + weight_size;
 
-    const enum ggml_type type = dst->src[0]->type;
-    float                weight_elem_size;
-    if (type == GGML_TYPE_Q4_0) {
-        weight_elem_size = float(sizeof(uint8_t)) / 2;
-    } else if (type == GGML_TYPE_Q8_0) {
-        weight_elem_size = float(sizeof(uint8_t));
-    } else {
-        GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
-    }
+    // Allocate buffers for selected expert weights and scales
+    const size_t         selected_weight_size = src0->ne[0] * src0->ne[1] * n_select_experts * weight_elem_size;
+    ggml_cann_pool_alloc selected_weight_alloc(ctx.pool(), selected_weight_size);
+    void *               selected_weight_buffer = selected_weight_alloc.get();
 
-    // src0_row [D, M, 1, 1] weight without permute
-    src0_row.ne[2]       = 1;
-    src0_row.ne[3]       = 1;
-    src0_row.nb[0]       = weight_elem_size;
-    src0_row.nb[1]       = weight_elem_size * ne00;
-    src0_row.nb[2]       = weight_elem_size * ne00;
-    src0_row.nb[3]       = weight_elem_size * ne00;
-    size_t weight_stride = ne00 * ne01 * weight_elem_size;
-    size_t weight_size   = weight_stride * ne02 * ne03;
+    const size_t selected_scale_size = (src0->ne[0] / group_size) * src0->ne[1] * n_select_experts * scale_elem_size;
+    ggml_cann_pool_alloc selected_scale_alloc(ctx.pool(), selected_scale_size);
+    void *               selected_scale_buffer = selected_scale_alloc.get();
 
-    // scale [D, M, 1, 1] -> scale && permute
-    size_t scale_elem_size = sizeof(uint16_t);
-    size_t scale_stride    = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
+    // Helper lambda to allocate and cast tensor to F16 if needed
+    constexpr size_t f16_elem_size      = sizeof(uint16_t);
+    auto             prepare_f16_buffer = [&](ggml_tensor * tensor, ggml_cann_pool_alloc & allocator,
+                                  bool need_cast = false) -> void * {
+        if (tensor->type == GGML_TYPE_F16) {
+            return tensor->data;
+        }
 
-    // src1_row [D, 1, 1, 1] -> input
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
+        size_t total_size = f16_elem_size;
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            total_size *= tensor->ne[i];
+        }
+        void * buffer = allocator.alloc(total_size);
 
-    // dst_row [M, 1, 1, 1] -> out
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
+        if (need_cast == false) {
+            return buffer;
+        }
 
-    //create weight for one row
-    ggml_cann_pool_alloc weight_allocator(ctx.pool());
-    void *               weight_buffer = weight_allocator.alloc(nb02);
-    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-        for (int64_t id = 0; id < n_ids; id++) {
-            // expert index
-            int32_t i02 = *(int32_t *) (ids_host.data() + iid1 * ids->nb[1] + id * ids->nb[0]);
-            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+        int64_t ne[GGML_MAX_DIMS];
+        size_t  nb[GGML_MAX_DIMS] = { f16_elem_size };
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            ne[i] = tensor->ne[i];
+            if (i > 0) {
+                nb[i] = nb[i - 1] * ne[i - 1];
+            }
+        }
 
-            // If B = 1 (broadcast), always use 0; otherwise, use id.
-            int64_t i11 = (ne11 == 1 ? 0 : id);
-            int64_t i12 = iid1;
+        acl_tensor_ptr src_tensor = ggml_cann_create_tensor(tensor);
+        acl_tensor_ptr f16_tensor = ggml_cann_create_tensor(buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
+        aclnn_cast(ctx, src_tensor.get(), f16_tensor.get(), ACL_FLOAT16);
 
-            int64_t i1 = id;
-            int64_t i2 = i12;
+        return buffer;
+    };
 
-            void * src0_tmp_ptr  = src0_original + i02 * weight_stride;
-            void * scale_tmp_ptr = src0_original + weight_size + i02 * scale_stride;
-            void * src1_tmp_ptr  = src1_original + i11 * nb11 + i12 * nb12;
-            void * dst_tmp_ptr   = dst_original + i1 * nb1 + i2 * nb2;
+    // Prepare input and output buffers
+    ggml_cann_pool_alloc input_alloc(ctx.pool());
+    void *               input_buffer = prepare_f16_buffer(src1, input_alloc, true);
 
-            // mem cpy
-            ACL_CHECK(aclrtMemcpyAsync(weight_buffer, weight_stride, src0_tmp_ptr, weight_stride,
-                                       ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
-            void * scale_buffer = (char *) weight_buffer + weight_stride;
-            ACL_CHECK(aclrtMemcpyAsync(scale_buffer, scale_stride, scale_tmp_ptr, scale_stride,
-                                       ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    ggml_cann_pool_alloc output_alloc(ctx.pool());
+    void *               output_buffer = prepare_f16_buffer(dst, output_alloc, false);
 
-            src0_row.data  = weight_buffer;
-            src1_row.data  = src1_tmp_ptr;
-            dst_row.data   = dst_tmp_ptr;
-            dst_row.src[0] = &src0_row;
-            dst_row.src[1] = &src1_row;
+    // Process each batch
+    for (int64_t batch_idx = 0; batch_idx < n_batches; batch_idx++) {
+        // Create index tensor for current batch
+        const size_t   index_offset  = batch_idx * ids->nb[1];
+        acl_tensor_ptr batch_indices = ggml_cann_create_tensor(ids, ids->ne, ids->nb, 1, ACL_FORMAT_ND, index_offset);
 
-            ggml_cann_mul_mat(ctx, &dst_row);
+        // Select quantized weights using expert indices
+        // Q4_0 stores 2 values per byte, Q8_0 stores 1 value per byte
+        const int64_t weight_d         = (type == GGML_TYPE_Q4_0) ? src0->ne[0] / 2 : src0->ne[0];
+        const int64_t weight_m         = src0->ne[1];
+        const int64_t weight_n_experts = src0->ne[2];
+
+        int64_t weight_ne[3] = { weight_d, weight_m, weight_n_experts };
+        size_t  weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t), weight_d * weight_m * sizeof(int8_t) };
+
+        acl_tensor_ptr all_weights =
+            ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb, 3);
+
+        int64_t selected_weight_ne[3] = { weight_d, weight_m, n_select_experts };
+        size_t  selected_weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t),
+                                          weight_d * weight_m * sizeof(int8_t) };
+
+        acl_tensor_ptr selected_weights = ggml_cann_create_tensor(selected_weight_buffer, ACL_INT8, sizeof(int8_t),
+                                                                  selected_weight_ne, selected_weight_nb, 3);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_weights.get(), 0, batch_indices.get(), selected_weights.get());
+
+        // Select scales using the same expert indices
+        const int64_t scale_d     = src0->ne[0] / group_size;
+        int64_t       scale_ne[3] = { scale_d, weight_m, weight_n_experts };
+        size_t scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size, scale_d * weight_m * scale_elem_size };
+
+        acl_tensor_ptr all_scales =
+            ggml_cann_create_tensor(scale_data, ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 3);
+
+        int64_t selected_scale_ne[3] = { scale_d, weight_m, n_select_experts };
+        size_t  selected_scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size,
+                                         scale_d * weight_m * scale_elem_size };
+
+        acl_tensor_ptr selected_scales = ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size,
+                                                                 selected_scale_ne, selected_scale_nb, 3);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_scales.get(), 0, batch_indices.get(), selected_scales.get());
+
+        // Process each expert for current batch
+        // IndexSelect output layout: [D, M, K] in contiguous format
+        // WeightQuantBatchMatmulV2 expects: [M, D] with row-major stride
+        for (int64_t expert_idx = 0; expert_idx < n_select_experts; expert_idx++) {
+            // Determine input offset: broadcast if src1->ne[1]==1, otherwise use per-expert input
+            const size_t input_offset =
+                (batch_idx * src1->ne[1] + (src1->ne[1] == 1 ? 0 : expert_idx)) * src1->ne[0] * f16_elem_size;
+            const size_t output_offset = (batch_idx * dst->ne[1] + expert_idx) * dst->ne[0] * f16_elem_size;
+
+            // Create weight view for current expert: [D, M, K] -> [M, D]
+            int64_t      weight_view_ne[2]  = { weight_m, src0->ne[0] };
+            float        weight_view_nb[2]  = { src0->ne[0] * weight_elem_size, weight_elem_size };
+            const size_t weight_view_offset = expert_idx * selected_weight_nb[2];
+
+            acl_tensor_ptr weight_view =
+                ggml_cann_create_tensor(selected_weight_buffer, ggml_cann_type_mapping(type), weight_elem_size,
+                                        weight_view_ne, weight_view_nb, 2, ACL_FORMAT_ND, weight_view_offset);
+
+            // Create scale view for current expert: [D, M, K] -> [M, D]
+            int64_t      scale_view_ne[2]  = { weight_m, scale_d };
+            size_t       scale_view_nb[2]  = { selected_scale_nb[1], selected_scale_nb[0] };
+            const size_t scale_view_offset = expert_idx * selected_scale_nb[2];
+
+            acl_tensor_ptr scale_view =
+                ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size, scale_view_ne,
+                                        scale_view_nb, 2, ACL_FORMAT_ND, scale_view_offset);
+
+            // Create input activation tensor [D, 1]
+            int64_t input_ne[2] = { src1->ne[0], 1 };
+            size_t  input_nb[2] = { f16_elem_size, src1->ne[0] * f16_elem_size };
+
+            acl_tensor_ptr input_tensor = ggml_cann_create_tensor(input_buffer, ACL_FLOAT16, f16_elem_size, input_ne,
+                                                                  input_nb, 2, ACL_FORMAT_ND, input_offset);
+
+            // Create output tensor [M, 1]
+            int64_t output_ne[2] = { dst->ne[0], 1 };
+            size_t  output_nb[2] = { f16_elem_size, dst->ne[0] * f16_elem_size };
+
+            acl_tensor_ptr output_tensor = ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, output_ne,
+                                                                   output_nb, 2, ACL_FORMAT_ND, output_offset);
+
+            // Perform quantized matrix multiplication
+            GGML_CANN_CALL_ACLNN_OP(ctx, WeightQuantBatchMatmulV2, input_tensor.get(), weight_view.get(),
+                                    scale_view.get(), nullptr, nullptr, nullptr, nullptr, group_size,
+                                    output_tensor.get());
         }
     }
-    return;
+
+    // Cast output back to original type if we used a temporary F16 buffer
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t ne[GGML_MAX_DIMS];
+        size_t  nb[GGML_MAX_DIMS] = { f16_elem_size };
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            ne[i] = dst->ne[i];
+            if (i > 0) {
+                nb[i] = nb[i - 1] * ne[i - 1];
+            }
+        }
+
+        acl_tensor_ptr f16_output =
+            ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
+        acl_tensor_ptr dst_tensor = ggml_cann_create_tensor(dst);
+
+        aclnn_cast(ctx, f16_output.get(), dst_tensor.get(), ggml_cann_type_mapping(dst->type));
+    }
 }
 
 void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cann/ggml-cann.cpp

@@ -794,19 +794,44 @@ struct ggml_backend_cann_buffer_context {
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };
 
+// cann buffer type
 /**
- * @brief Check if a buffer is a CANN buffer.
- *
- * This function checks if a given buffer is a CANN buffer by comparing its
- * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
- *
- * @param buffer The buffer to check.
- * @return true if the buffer is a CANN buffer, false otherwise.
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
  */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
+struct ggml_backend_cann_buffer_type_context {
+    int32_t     device; /**< Device identifier associated with the buffer context. */
+    std::string name;   /**< Name associated with the buffer context. */
+};
 
-static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_cann(buffer->buft);
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
+
+    return buft_ctx->name.c_str();
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
 /**
@@ -1271,7 +1296,7 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
 static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                                 const ggml_tensor *   src,
                                                 ggml_tensor *         dst) {
-    if (ggml_backend_buffer_is_cann(src->buffer)) {
+    if (ggml_backend_buft_is_cann(src->buffer->buft)) {
         ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
         ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
 
@@ -1335,31 +1360,6 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .reset           = */ NULL,
 };
 
-// cann buffer type
-/**
- * @brief Structure representing context information for a specific backend
- * buffer type.
- */
-struct ggml_backend_cann_buffer_type_context {
-    int32_t     device; /**< Device identifier associated with the buffer context. */
-    std::string name;   /**< Name associated with the buffer context. */
-};
-
-/**
- * @brief Retrieves the name associated with a CANN buffer type.
- *
- * This function returns the descriptive name associated with the specified
- * CANN buffer type context.
- *
- * @param buft Pointer to the buffer type context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
-    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
-
-    return buft_ctx->name.c_str();
-}
-
 /**
  * @brief Allocates a new CANN buffer of the specified type and size.
  *
@@ -1997,7 +1997,7 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t      backend_src,
 
     GGML_ASSERT(!is_matmul_weight((const ggml_tensor *) src));
 
-    if (!ggml_backend_buffer_is_cann(src->buffer) || !ggml_backend_buffer_is_cann(dst->buffer)) {
+    if (!ggml_backend_buft_is_cann(src->buffer->buft) || !ggml_backend_buft_is_cann(dst->buffer->buft)) {
         return false;
     }
 
@@ -2523,21 +2523,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
     GGML_UNUSED(dev);
 }
 
-/**
- * @brief Checks if the backend buffer type is associated with the CANN backend.
- *
- * This function checks whether the provided backend buffer type is associated
- * with the CANN backend based on the comparison of its name retrieval function
- * pointer.
- *
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the buffer type is associated with the CANN
- * backend, otherwise false.
- */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *

ggml/src/ggml-cpu/CMakeLists.txt

@@ -569,27 +569,24 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             cmake_policy(SET CMP0135 NEW)
         endif()
 
+        # TODO: Use FetchContent_MakeAvailable with EXCLUDE_FROM_ALL after bumping minimum CMake version to 3.28+
+        # Using FetchContent_Populate instead to avoid EXCLUDE_FROM_ALL which requires CMake 3.28
         FetchContent_Declare(KleidiAI_Download
             URL ${KLEIDIAI_DOWNLOAD_URL}
             DOWNLOAD_EXTRACT_TIMESTAMP NEW
             URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
 
-        FetchContent_MakeAvailable(KleidiAI_Download)
         FetchContent_GetProperties(KleidiAI_Download
             SOURCE_DIR  KLEIDIAI_SRC
             POPULATED   KLEIDIAI_POPULATED)
 
         if (NOT KLEIDIAI_POPULATED)
-            message(FATAL_ERROR "KleidiAI source downloaded failed.")
+            FetchContent_Populate(KleidiAI_Download)
+            FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
         endif()
 
         add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
 
-        # Remove kleidiai target after fetching it
-        if (TARGET kleidiai)
-            set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
-        endif()
-
         list(APPEND GGML_CPU_SOURCES
             ggml-cpu/kleidiai/kleidiai.cpp
             ggml-cpu/kleidiai/kernels.cpp

ggml/src/ggml-cpu/arch-fallback.h

@@ -43,6 +43,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -55,7 +56,8 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
-#    define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -76,6 +78,7 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -84,6 +87,7 @@
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -107,6 +111,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -119,6 +124,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -143,6 +149,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -155,6 +162,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -186,6 +194,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -197,6 +206,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -227,6 +237,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -239,6 +250,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -271,6 +283,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -283,6 +296,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -1072,6 +1072,195 @@ void ggml_gemv_q5_K_8x8_q8_K(int                        n,
     ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q6_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4;
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo    = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi    = vdupq_n_u8(0x30);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[2];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q6_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q6_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q6_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q6_d_1, q8_d);
+
+            int32x4_t acc[col_groups];
+            for (int i = 0; i < col_groups; i++) {
+                acc[i] = vdupq_n_s32(0);
+            }
+
+            // Load all 16 scales once and widen to int16 (Q6_K has 16 scales per block)
+            // Reused for bias and dequantization later
+            int16_t q6_scales[16 * 8];
+            for (int i = 0; i < 16; i++) {
+                int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                vst1q_s16(q6_scales + i * 8, scales);
+            }
+
+            // Compute bias per column using q8 bsums and preloaded scales to skip the -32 shift
+            int32x4_t bias_lo = vdupq_n_s32(0);
+            int32x4_t bias_hi = vdupq_n_s32(0);
+
+            // Load bsums in chunks of 4 to process with vectorized operations
+            for (int i = 0; i < 16; i += 4) {
+                int16x4_t bsums_vec   = vld1_s16(q8_ptr[b].bsums + i);
+                int16x4_t scales_lo_0 = vld1_s16(q6_scales + (i + 0) * 8);
+                int16x4_t scales_hi_0 = vld1_s16(q6_scales + (i + 0) * 8 + 4);
+                int16x4_t scales_lo_1 = vld1_s16(q6_scales + (i + 1) * 8);
+                int16x4_t scales_hi_1 = vld1_s16(q6_scales + (i + 1) * 8 + 4);
+                int16x4_t scales_lo_2 = vld1_s16(q6_scales + (i + 2) * 8);
+                int16x4_t scales_hi_2 = vld1_s16(q6_scales + (i + 2) * 8 + 4);
+                int16x4_t scales_lo_3 = vld1_s16(q6_scales + (i + 3) * 8);
+                int16x4_t scales_hi_3 = vld1_s16(q6_scales + (i + 3) * 8 + 4);
+
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_0, bsums_vec, 0);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_0, bsums_vec, 0);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_1, bsums_vec, 1);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_1, bsums_vec, 1);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_2, bsums_vec, 2);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_2, bsums_vec, 2);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_3, bsums_vec, 3);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_3, bsums_vec, 3);
+            }
+            bias_lo = vshlq_n_s32(bias_lo, 5);
+            bias_hi = vshlq_n_s32(bias_hi, 5);
+
+            // Process two 128-value halves per superblock
+            for (int half = 0; half < 2; half++) {
+                const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                // A subblock (sb) is a set of weights that share the scale
+                // Since q6_K scales are per 16 elements
+                // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves)
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    const int8_t * q8_base_l = q8_ptr[b].qs + half * 128 + sb * 16;
+                    const int8_t * q8_base_h = q8_base_l + 64;
+
+                    // Load and duplicate q8 values (each register covers four interleaved columns of q6)
+                    int8x16_t q8_l[4];
+                    int8x16_t q8_h[4];
+                    for (int i = 0; i < 4; i++) {
+                        q8_l[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_l + i * 4));
+                        q8_h[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_h + i * 4));
+                    }
+
+                    const int ql_off_base = sb * QK_K / 2;
+                    const int qh_off_base = ql_off_base & 255;  // wraps after 256 bytes
+
+                    // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
+                    uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
+                    uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
+                    uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
+                    uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
+
+                    // Adjust qh for subblocks 2 and 3 (shift right by 2)
+                    if (sb > 1) {
+                        q6_qh_0.val[0] = vshrq_n_u8(q6_qh_0.val[0], 2);
+                        q6_qh_0.val[1] = vshrq_n_u8(q6_qh_0.val[1], 2);
+                        q6_qh_0.val[2] = vshrq_n_u8(q6_qh_0.val[2], 2);
+                        q6_qh_0.val[3] = vshrq_n_u8(q6_qh_0.val[3], 2);
+                        q6_qh_1.val[0] = vshrq_n_u8(q6_qh_1.val[0], 2);
+                        q6_qh_1.val[1] = vshrq_n_u8(q6_qh_1.val[1], 2);
+                        q6_qh_1.val[2] = vshrq_n_u8(q6_qh_1.val[2], 2);
+                        q6_qh_1.val[3] = vshrq_n_u8(q6_qh_1.val[3], 2);
+                    }
+
+                    const uint8x16_t q6_ql[8] = { q6_ql_0.val[0], q6_ql_0.val[1], q6_ql_0.val[2], q6_ql_0.val[3],
+                                                  q6_ql_1.val[0], q6_ql_1.val[1], q6_ql_1.val[2], q6_ql_1.val[3] };
+                    const uint8x16_t q6_qh[8] = { q6_qh_0.val[0], q6_qh_0.val[1], q6_qh_0.val[2], q6_qh_0.val[3],
+                                                  q6_qh_1.val[0], q6_qh_1.val[1], q6_qh_1.val[2], q6_qh_1.val[3] };
+
+                    // Process column groups (0-3, 4-7)
+                    for (int g = 0; g < col_groups; g++) {
+                        int32x4_t sb_acc_l = vdupq_n_s32(0);
+                        int32x4_t sb_acc_h = vdupq_n_s32(0);
+
+                        for (int chunk = 0; chunk < 4; chunk++) {
+                            const int idx = chunk * 2 + g;
+
+                            const uint8x16_t q6_qs_l = q6_ql[idx];
+                            const uint8x16_t q6_qs_h = q6_qh[idx];
+
+                            // Extract high 2 bits for upper nibble reconstruction
+                            const uint8x16_t q6_qs_hh = vandq_u8(q6_qs_h, mask_hi);
+
+                            // q6 = (low4 | high2<<4), without -32 bias (handled via bsums)
+                            const int8x16_t q6_l =
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_l, m4b), vandq_u8(q6_qs_h, mask_lo), 4));
+                            const int8x16_t q6_h = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_l, 4), q6_qs_hh));
+
+                            sb_acc_l = vdotq_s32(sb_acc_l, q6_l, q8_l[chunk]);
+                            sb_acc_h = vdotq_s32(sb_acc_h, q6_h, q8_h[chunk]);
+                        }
+
+                        const int scale_idx_l = half * 8 + sb;
+                        const int scale_idx_h = half * 8 + sb + 4;
+
+                        const int32x4_t scale_vec_l = vmovl_s16(vld1_s16(q6_scales + scale_idx_l * 8 + g * 4));
+                        const int32x4_t scale_vec_h = vmovl_s16(vld1_s16(q6_scales + scale_idx_h * 8 + g * 4));
+
+                        acc[g] = vmlaq_s32(acc[g], sb_acc_l, scale_vec_l);
+                        acc[g] = vmlaq_s32(acc[g], sb_acc_h, scale_vec_h);
+                    }
+                }
+            }  // for half
+
+            // Bias correction
+            acc[0] = vsubq_s32(acc[0], bias_lo);
+            acc[1] = vsubq_s32(acc[1], bias_hi);
+
+            // Apply superblock scale (no mins for q6_K)
+            // acc[g] has [c0, c1, c2, c3]
+            float32x4_t w_0123 = vmulq_f32(vcvtq_f32_s32(acc[0]), sb_scale_0);
+            float32x4_t w_4567 = vmulq_f32(vcvtq_f32_s32(acc[1]), sb_scale_1);
+
+            acc_f32[0] = vaddq_f32(acc_f32[0], w_0123);
+            acc_f32[1] = vaddq_f32(acc_f32[1], w_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q6_K_8x8_q8_K(int                        n,
                              float * GGML_RESTRICT      s,
                              size_t                     bs,
@@ -1177,15 +1366,14 @@ void ggml_gemv_q6_K_8x8_q8_K(int                        n,
                         q8_h[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_h + i * 8));
                     }
 
-                    // TODO: Test other qh repack patterns to reduce loads
                     const int ql_off_base = sb * QK_K / 2;
                     const int qh_off_base = ql_off_base & 255;  // wraps after 256 bytes
 
                     // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
-                    ggml_uint8x16x4_t q6_ql_0 = ggml_vld1q_u8_x4(ql_base + ql_off_base);
-                    ggml_uint8x16x4_t q6_ql_1 = ggml_vld1q_u8_x4(ql_base + ql_off_base + 64);
-                    ggml_uint8x16x4_t q6_qh_0 = ggml_vld1q_u8_x4(qh_base + qh_off_base);
-                    ggml_uint8x16x4_t q6_qh_1 = ggml_vld1q_u8_x4(qh_base + qh_off_base + 64);
+                    uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
+                    uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
+                    uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
+                    uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
 
                     // Adjust qh for subblocks 2 and 3 (shift right by 2)
                     if (sb > 1) {
@@ -3038,6 +3226,316 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (svcntb() * 8 == 256) {
+        constexpr int    q8_k_blocklen = 4;
+        const svuint8_t m4b_1          = svdup_n_u8(0x0f);
+        // 8 accumulators: 2 row pairs × 4 col pairs
+        svfloat32_t acc_f32_01, acc_f32_23, acc_f32_45, acc_f32_67;
+        uint32_t idx_arr[8] = { 0, 2, 4, 6,  1, 3, 5, 7 };
+        svbool_t pg = svptrue_pat_b32(SV_VL8);
+        svuint32_t idx = svld1(pg, idx_arr);
+
+        static const uint32_t idx_data[8] = {0, 4, 2, 6, 1, 5, 3, 7};
+        svuint32_t idx1 = svld1_u32(svptrue_b32(), idx_data);
+
+        for (int y = 0; y < nr / q8_k_blocklen; y++) {
+            const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_Kx8 * GGML_RESTRICT q4_ptr = (const block_q4_Kx8 *) vx + (x * nb);
+
+                acc_f32_01 = svdup_n_f32(0);
+                acc_f32_23 = svdup_n_f32(0);
+                acc_f32_45 = svdup_n_f32(0);
+                acc_f32_67 = svdup_n_f32(0);
+
+                for (int b = 0; b < nb; b++) {
+                    // bsums pairs belongs to the same q8_k subblock
+                    // 64 elemnts loaded and made sum of 0-7 and 8-15 sum || 16-23 and 24 - 31 sum
+                    const int16x8_t bsums[4]{
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                        vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                    };
+
+                    int32_t bsums_arr32[4][8];
+
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        int16x8_t v16 = bsums[q8_row];
+
+                        // low 4
+                        int32x4_t v32_lo = vmovl_s16(vget_low_s16(v16));
+                        vst1q_s32(&bsums_arr32[q8_row][0], v32_lo);
+
+                        // high 4
+                        int32x4_t v32_hi = vmovl_s16(vget_high_s16(v16));
+                        vst1q_s32(&bsums_arr32[q8_row][4], v32_hi);
+                    }
+
+                    svint32_t sb_acc_0 = svdup_n_s32(0);
+                    svint32_t sb_acc_2 = svdup_n_s32(0);
+
+                    svint32_t acc_00 = svdup_n_s32(0);
+                    svint32_t acc_11 = svdup_n_s32(0);
+                    svint32_t acc_22 = svdup_n_s32(0);
+                    svint32_t acc_33 = svdup_n_s32(0);
+                    svint32_t acc_44 = svdup_n_s32(0);
+                    svint32_t acc_55 = svdup_n_s32(0);
+                    svint32_t acc_66 = svdup_n_s32(0);
+                    svint32_t acc_77 = svdup_n_s32(0);
+
+                    svint32_t bias_acc_00 = svdup_n_s32(0);
+                    svint32_t bias_acc_22 = svdup_n_s32(0);
+                    svint32_t bias_acc_44 = svdup_n_s32(0);
+                    svint32_t bias_acc_66 = svdup_n_s32(0);
+
+                    for (int sb = 0; sb < QK_K / 64; sb++) {
+                        // Need scales for the low and high nibbles
+                        // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                        svint32_t block_scale_0, block_scale_1, block_scale_2, block_scale_3;
+                        svint32_t q4sb_mins_0, q4sb_mins_1;
+                        {
+                            // 2-superblock I am working on
+                            const int offset = sb * 24 + 0 * 12;
+                            const uint8_t * scales_in = &q4_ptr[b].scales[offset];
+
+                            const int offset1 = sb * 24 + 12;
+                            const uint8_t * scales_in1 = &q4_ptr[b].scales[offset1];
+
+                            constexpr uint32_t kmask1 = 0x3f3f3f3f;
+                            constexpr uint32_t kmask2 = 0x0f0f0f0f;
+                            constexpr uint32_t kmask3 = 0x03030303;
+                            constexpr uint8_t  scales_size = 12;
+
+                            uint32_t sm[3];
+                            memcpy(sm, scales_in, scales_size);
+
+                            uint32_t sm1[3];
+                            memcpy(sm1, scales_in1, scales_size);
+
+                            const uint32_t mins_0_3 = sm[1] & kmask1;
+                            const uint32_t mins_4_7 = ((sm[2] >> 4) & kmask2) | (((sm[1] >> 6) & kmask3) << 4);
+
+                            const uint32_t mins_0_3_1 = sm1[1] & kmask1;
+                            const uint32_t mins_4_7_1 = ((sm1[2] >> 4) & kmask2) | (((sm1[1] >> 6) & kmask3) << 4);
+
+                            svuint32_t mins_u32_temp = svzip1_u32(svdup_n_u32(mins_0_3), svdup_n_u32(mins_4_7));
+                            svuint32_t mins_u32_temp_1 = svzip1_u32(svdup_n_u32(mins_0_3_1), svdup_n_u32(mins_4_7_1));
+
+                            /* reinterpret u32 → u8 */
+                            svuint8_t mins_u8 = svreinterpret_u8_u32(mins_u32_temp);
+                            svuint8_t mins_u8_1 = svreinterpret_u8_u32(mins_u32_temp_1);
+
+                            /* widen u8 → u16->u32 (lower half only) */
+                            svuint32_t mins_u16 = svunpklo_u32(svunpklo_u16(mins_u8));
+                            svuint32_t mins_u16_1 = svunpklo_u32(svunpklo_u16(mins_u8_1));
+
+                            q4sb_mins_0 = svreinterpret_s32_u32(mins_u16);
+                            q4sb_mins_1 = svreinterpret_s32_u32(mins_u16_1);
+
+                            uint32_t scales_u32_0 = sm[0] & kmask1;
+                            uint32_t scales_u32_1 = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
+                            uint32_t scales_u32_2 = sm1[0] & kmask1;
+                            uint32_t scales_u32_3 = (sm1[2] & kmask2) | (((sm1[0] >> 6) & kmask3) << 4);
+
+                            svuint32_t S01 = svdup_n_u32(scales_u32_0);
+                            svuint32_t S23 = svdup_n_u32(scales_u32_1);
+                            svuint32_t R01 = svdup_n_u32(scales_u32_2);
+                            svuint32_t R23 = svdup_n_u32(scales_u32_3);
+
+                            svint8_t S01_b = svreinterpret_s8_u32(S01);
+                            svint8_t S23_b = svreinterpret_s8_u32(S23);
+                            svint8_t R01_b = svreinterpret_s8_u32(R01);
+                            svint8_t R23_b = svreinterpret_s8_u32(R23);
+
+                            svint32_t S01_d = svunpklo_s32(svunpklo_s16(svzip1_s8(S01_b, S01_b)));
+                            svint32_t R01_d = svunpklo_s32(svunpklo_s16(svzip1_s8(R01_b, R01_b)));
+                            svint32_t S23_d = svunpklo_s32(svunpklo_s16(svzip1_s8(S23_b, S23_b)));
+                            svint32_t R23_d = svunpklo_s32(svunpklo_s16(svzip1_s8(R23_b, R23_b)));
+
+                            block_scale_0 = svtbl_s32(svzip1_s32(S01_d, R01_d), idx);
+                            block_scale_1 = svtbl_s32(svzip2_s32(S01_d, R01_d), idx);
+                            block_scale_2 = svtbl_s32(svzip1_s32(S23_d, R23_d), idx);
+                            block_scale_3 = svtbl_s32(svzip2_s32(S23_d, R23_d), idx);
+                        }
+
+                        const int8_t * q8_base_1 = q8_ptr[b].qs + sb * 256;
+
+                        // Load 32-byte per row pair, 1 subblock each time
+                        // predicate for activating higher lanes for 16 int8 elements
+                        const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                        // predicate for activating lower lanes for  16 int8 elements
+                        const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                        svint8_t q8_qs_0 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 0), svld1_s8(pl16, q8_base_1 + 112));
+                        svint8_t q8_qs_2 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 32), svld1_s8(pl16, q8_base_1 + 144));
+                        svint8_t q8_qs_4 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 64), svld1_s8(pl16, q8_base_1 + 176));
+                        svint8_t q8_qs_6 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 96), svld1_s8(pl16, q8_base_1 + 208));
+
+                        svint8_t q8_qs_1 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 16), svld1_s8(pl16, q8_base_1 + 128));
+                        svint8_t q8_qs_3 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 48), svld1_s8(pl16, q8_base_1 + 160));
+                        svint8_t q8_qs_5 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 80), svld1_s8(pl16, q8_base_1 + 192));
+                        svint8_t q8_qs_7 = svadd_s8_x(svptrue_b8(), svld1_s8(ph16, q8_base_1 + 112), svld1_s8(pl16, q8_base_1 + 224));
+
+                        // Q4s columns iterated in pairs (01, 23, 45, 67)
+                        for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+
+                            sb_acc_0 = svdup_n_s32(0);
+                            sb_acc_2 = svdup_n_s32(0);
+
+                            svuint8_t q4_qs_cp_00 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 0);
+                            svuint8_t q4_qs_cp_01 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 64);
+                            svuint8_t q4_qs_cp_02 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 128);
+                            svuint8_t q4_qs_cp_03 = svld1rq_u8(svptrue_b8(), q4_ptr[b].qs + sb * QK_K + 16 * cp + 192);
+
+                            svint8_t q4_nibbles_00 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_00, m4b_1), 4));
+                            svint8_t q4_nibbles_01 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_01, m4b_1), 4));
+                            svint8_t q4_nibbles_02 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_02, m4b_1), 4));
+                            svint8_t q4_nibbles_03 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_u8_m(ph16, q4_qs_cp_03, m4b_1), 4));
+
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_00, q8_qs_0);
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_01, q8_qs_2);
+
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_02, q8_qs_4);
+                            sb_acc_0 = svmmla_s32(sb_acc_0, q4_nibbles_03, q8_qs_6);
+
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_00, q8_qs_1);
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_01, q8_qs_3);
+
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_02, q8_qs_5);
+                            sb_acc_2 = svmmla_s32(sb_acc_2, q4_nibbles_03, q8_qs_7);
+
+                            if(cp == 0) {
+                                acc_00 = svmla_s32_m(svptrue_b32(), acc_00, sb_acc_0, block_scale_0);
+                                acc_44 = svmla_s32_m(svptrue_b32(), acc_44, sb_acc_2, block_scale_0);
+                            }
+                            if(cp == 1) {
+                                acc_11 = svmla_s32_m(svptrue_b32(), acc_11, sb_acc_0, block_scale_1);
+                                acc_55 = svmla_s32_m(svptrue_b32(), acc_55, sb_acc_2, block_scale_1);
+                            }
+                            if(cp == 2) {
+                                acc_22 = svmla_s32_m(svptrue_b32(), acc_22, sb_acc_0, block_scale_2);
+                                acc_66 = svmla_s32_m(svptrue_b32(), acc_66, sb_acc_2, block_scale_2);
+                            }
+                            if(cp == 3) {
+                                acc_33 = svmla_s32_m(svptrue_b32(), acc_33, sb_acc_0, block_scale_3);
+                                acc_77 = svmla_s32_m(svptrue_b32(), acc_77, sb_acc_2, block_scale_3);
+                            }
+                        }
+
+                        bias_acc_00 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_00, svdup_n_s32(bsums_arr32[sb][0]), q4sb_mins_0);
+                        bias_acc_00 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_00, svdup_n_s32(bsums_arr32[sb][1]), q4sb_mins_1);
+
+                        bias_acc_22 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_22, svdup_n_s32(bsums_arr32[sb][2]), q4sb_mins_0);
+                        bias_acc_22 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_22, svdup_n_s32(bsums_arr32[sb][3]), q4sb_mins_1);
+
+                        bias_acc_44 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_44, svdup_n_s32(bsums_arr32[sb][4]), q4sb_mins_0);
+                        bias_acc_44 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_44, svdup_n_s32(bsums_arr32[sb][5]), q4sb_mins_1);
+
+                        bias_acc_66 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_66, svdup_n_s32(bsums_arr32[sb][6]), q4sb_mins_0);
+                        bias_acc_66 = svmla_s32_m(svptrue_pat_b32(SV_VL8), bias_acc_66, svdup_n_s32(bsums_arr32[sb][7]), q4sb_mins_1);
+                    }  // for sb
+
+
+                    acc_00 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_00, svext_s32(acc_00, acc_00, 4));
+                    acc_11 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_11, svext_s32(acc_11, acc_11, 4));
+                    acc_22 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_22, svext_s32(acc_22, acc_22, 4));
+                    acc_33 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_33, svext_s32(acc_33, acc_33, 4));
+                    acc_44 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_44, svext_s32(acc_44, acc_44, 4));
+                    acc_55 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_55, svext_s32(acc_55, acc_55, 4));
+                    acc_66 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_66, svext_s32(acc_66, acc_66, 4));
+                    acc_77 = svadd_s32_z(svptrue_pat_b32(SV_VL4), acc_77, svext_s32(acc_77, acc_77, 4));
+
+                    svint32_t reorder_acc_01 = svtbl_s32( svzip1_s32( svtrn1_s32(acc_00, acc_11), svtrn1_s32(acc_22, acc_33)), idx1);
+                    svint32_t reorder_acc_23 = svtbl_s32( svzip1_s32( svtrn2_s32(acc_00, acc_11), svtrn2_s32(acc_22, acc_33)), idx1);
+
+                    svint32_t reorder_acc_45 = svtbl_s32( svzip1_s32( svtrn1_s32(acc_44, acc_55), svtrn1_s32(acc_66, acc_77)), idx1);
+                    svint32_t reorder_acc_67 = svtbl_s32( svzip1_s32( svtrn2_s32(acc_44, acc_55), svtrn2_s32(acc_66, acc_77)), idx1);
+
+                    // Broadcast q8 scalar
+                    svfloat32_t q8_d = svdup_f32(q8_ptr[b].d[0]);
+
+                    svfloat32_t q4_dmin_temp = svcvt_f32_f16_x(svptrue_b32(), svzip1_f16( svld1_f16(svptrue_pat_b16(SV_VL8), (const __fp16 *)q4_ptr[b].dmin), svdup_f16(0)));
+
+                    svfloat32_t q4_d_temp = svcvt_f32_f16_x(svptrue_b32(), svzip1_f16( svld1_f16(svptrue_pat_b16(SV_VL8), (const __fp16 *)q4_ptr[b].d), svdup_f16(0)));
+
+                    svfloat32_t scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    svfloat32_t dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_01 = svmls_f32_m(svptrue_b32(), acc_f32_01, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_00), dmins1);
+                    acc_f32_01 = svmla_f32_m(svptrue_b32(), acc_f32_01, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_01), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[1]);
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_23 = svmls_f32_m(svptrue_b32(), acc_f32_23, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_22), dmins1);
+                    acc_f32_23 = svmla_f32_m(svptrue_b32(), acc_f32_23, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_23), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[2]);
+
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_45 = svmls_f32_m(svptrue_b32(), acc_f32_45, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_44), dmins1);
+                    acc_f32_45 = svmla_f32_m(svptrue_b32(), acc_f32_45, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_45), scale1);
+
+                    q8_d = svdup_f32(q8_ptr[b].d[3]);
+
+                    scale1 = svmul_f32_x(svptrue_b32(), q4_d_temp, q8_d);
+                    dmins1 = svmul_f32_x(svptrue_b32(), q4_dmin_temp, q8_d);
+
+                    acc_f32_67 = svmls_f32_m(svptrue_b32(), acc_f32_67, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), bias_acc_66), dmins1);
+                    acc_f32_67 = svmla_f32_m(svptrue_b32(), acc_f32_67, svcvt_f32_s32_m(svdup_n_f32(0), svptrue_b32(), reorder_acc_67), scale1);
+
+                }  // for b
+
+                // With the previous reorder, the tile is already in the correct memory layout.
+                // Predicate for exactly 4 lanes
+                svbool_t pg4 = svptrue_pat_b32(SV_VL4);
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    int row = y * q8_k_blocklen + i;
+                    for (int j = 0; j < 2; j++) {
+                        int col    = x * ncols_interleaved + j * 4;
+                        int offset = row * bs + col;
+
+                        if (i == 0 && j == 0) {
+                            // acc_f32_0 → lower half of acc_f32_01
+                            svst1_f32(pg4, s + offset, acc_f32_01);
+                        } else if (i == 0 && j == 1) {
+                            // acc_f32_1 → upper half of acc_f32_01
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_01, acc_f32_01, 4));
+                        } else if (i == 1 && j == 0) {
+                            // acc_f32_2
+                            svst1_f32(pg4, s + offset, acc_f32_23);
+                        } else if (i == 1 && j == 1) {
+                            // acc_f32_3
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_23, acc_f32_23, 4));
+                        } else if (i == 2 && j == 0) {
+                            // acc_f32_4
+                            svst1_f32(pg4, s + offset, acc_f32_45);
+                        } else if (i == 2 && j == 1) {
+                            // acc_f32_5
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_45, acc_f32_45, 4));
+                        } else if (i == 3 && j == 0) {
+                            // acc_f32_6
+                            svst1_f32(pg4, s + offset, acc_f32_67);
+                        } else if (i == 3 && j == 1) {
+                            // acc_f32_7
+                            svst1_f32(pg4, s + offset, svext_f32(acc_f32_67, acc_f32_67, 4));
+                        }
+                    }
+                }
+            }  // for x
+        }  // for y
+        return;
+    }
+#endif  // SVE compile-time end
+
 #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     constexpr int    q8_k_blocklen = 4;
     const uint8x16_t m4b           = vdupq_n_u8(0x0f);
@@ -3474,6 +3972,208 @@ void ggml_gemm_q5_K_8x8_q8_K(int                        n,
     ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemm_q6_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    col_groups    = ncols_interleaved / 4;
+    constexpr int    acc_size      = q8_k_blocklen * col_groups;  // 4 rows, 2 column groups
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo       = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi       = vdupq_n_u8(0x30);
+    const int8x16_t  m32s          = vdupq_n_s8(32);
+
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                float32x4_t q6_d_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));
+                float32x4_t q6_d_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));
+                float32x4_t q8_d_0123 = vld1q_f32(q8_ptr[b].d);
+
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 0);
+                sbd_scale_0123[1] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 1);
+                sbd_scale_0123[2] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 2);
+                sbd_scale_0123[3] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 3);
+
+                int32x4_t acc_s32[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    acc_s32[i] = vdupq_n_s32(0);
+                }
+
+                int16_t q6_scales[8 * 16];
+                for (int i = 0; i < 16; i++) {
+                    int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                    vst1q_s16(q6_scales + i * 8, scales);
+                }
+
+                for (int half = 0; half < 2; half++) {
+                    const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                    const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                    for (int sb = 0; sb < QK_K / 64; sb++) {
+                        int32x4_t acc_lo[acc_size];
+                        int32x4_t acc_hi[acc_size];
+                        for (int i = 0; i < acc_size; i++) {
+                            acc_lo[i] = vdupq_n_s32(0);
+                            acc_hi[i] = vdupq_n_s32(0);
+                        }
+
+                        const int8_t * q8_base_l = q8_ptr[b].qs + half * 512 + sb * 64;
+                        const int8_t * q8_base_h = q8_ptr[b].qs + half * 512 + 256 + sb * 64;
+
+                        // 4 rows * 16 elements per scale
+                        // 4 reads of 16 bytes each
+                        constexpr int reads_per_sb = 4;
+                        int8x16_t     q8_l[reads_per_sb];
+                        int8x16_t     q8_h[reads_per_sb];
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            q8_l[k] = vld1q_s8(q8_base_l + 16 * k);
+                            q8_h[k] = vld1q_s8(q8_base_h + 16 * k);
+                        }
+
+                        const int ql_off_base = sb * QK_K / 2;
+                        const int qh_off_base = ql_off_base & 255;
+
+                        uint8x16_t q6_ql_0123[reads_per_sb];
+                        uint8x16_t q6_ql_4567[reads_per_sb];
+                        uint8x16_t q6_qh_0123[reads_per_sb];
+                        uint8x16_t q6_qh_4567[reads_per_sb];
+
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            q6_ql_0123[k] = vld1q_u8(ql_base + ql_off_base + k * 32);
+                            q6_ql_4567[k] = vld1q_u8(ql_base + ql_off_base + k * 32 + 16);
+                            q6_qh_0123[k] = vld1q_u8(qh_base + qh_off_base + k * 32);
+                            q6_qh_4567[k] = vld1q_u8(qh_base + qh_off_base + k * 32 + 16);
+                        }
+
+                        if (sb > 1) {
+                            for (int k = 0; k < reads_per_sb; k++) {
+                                q6_qh_0123[k] = vshrq_n_u8(q6_qh_0123[k], 2);
+                                q6_qh_4567[k] = vshrq_n_u8(q6_qh_4567[k], 2);
+                            }
+                        }
+
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            // q = (ql | qh) - 32
+                            const uint8x16_t hbit_lo_0123 = vandq_u8(q6_qh_0123[k], mask_lo);
+                            const uint8x16_t hbit_hi_0123 = vandq_u8(q6_qh_0123[k], mask_hi);
+                            const uint8x16_t hbit_lo_4567 = vandq_u8(q6_qh_4567[k], mask_lo);
+                            const uint8x16_t hbit_hi_4567 = vandq_u8(q6_qh_4567[k], mask_hi);
+
+                            const int8x16_t q6_0123_lo = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)), m32s);
+                            const int8x16_t q6_0123_hi = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)), m32s);
+
+                            acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q6_0123_lo, q8_l[k], 0);  //  0..3  r0 c0123
+                            acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q6_0123_lo, q8_l[k], 1);  //  0..3  r1 c0123
+                            acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q6_0123_lo, q8_l[k], 2);  //  0..3  r2 c0123
+                            acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q6_0123_lo, q8_l[k], 3);  //  0..3  r3 c0123
+
+                            acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q6_0123_hi, q8_h[k], 0);  // 64..67 r0 c0123
+                            acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q6_0123_hi, q8_h[k], 1);  // 64..67 r1 c0123
+                            acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q6_0123_hi, q8_h[k], 2);  // 64..67 r2 c0123
+                            acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q6_0123_hi, q8_h[k], 3);  // 64..67 r3 c0123
+
+                            const int8x16_t q6_4567_lo = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)), m32s);
+                            const int8x16_t q6_4567_hi = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)), m32s);
+
+                            acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q6_4567_lo, q8_l[k], 0);  //  0..3  r0 c4567
+                            acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q6_4567_lo, q8_l[k], 1);  //  0..3  r1 c4567
+                            acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q6_4567_lo, q8_l[k], 2);  //  0..3  r2 c4567
+                            acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q6_4567_lo, q8_l[k], 3);  //  0..3  r3 c4567
+
+                            acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q6_4567_hi, q8_h[k], 0);  // 64..67 r0 c4567
+                            acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q6_4567_hi, q8_h[k], 1);  // 64..67 r1 c4567
+                            acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q6_4567_hi, q8_h[k], 2);  // 64..67 r2 c4567
+                            acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q6_4567_hi, q8_h[k], 3);  // 64..67 r3 c4567
+                        }
+
+                        // Scale and bias
+                        const int scale_idx_l = half * 8 + sb;
+                        const int scale_idx_h = half * 8 + sb + 4;
+
+                        for (int g = 0; g < col_groups; g++) {
+                            const int16x4_t scales_l16  = vld1_s16(q6_scales + scale_idx_l * 8 + g * 4);
+                            const int16x4_t scales_h16  = vld1_s16(q6_scales + scale_idx_h * 8 + g * 4);
+                            const int32x4_t scale_vec_l = vmovl_s16(scales_l16);
+                            const int32x4_t scale_vec_h = vmovl_s16(scales_h16);
+                            const int       acc_offset  = g * q8_k_blocklen;
+
+                            for (int row = 0; row < q8_k_blocklen; row++) {
+                                const int idx = row * 2 + g;
+                                acc_s32[idx]  = vmlaq_s32(acc_s32[idx], acc_lo[acc_offset + row], scale_vec_l);
+                                acc_s32[idx]  = vmlaq_s32(acc_s32[idx], acc_hi[acc_offset + row], scale_vec_h);
+                            }
+                        }
+                    }
+                }
+
+                // Finally we apply the superblock scales
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    const int       idx0     = 2 * row;
+                    const int       idx1     = 2 * row + 1;
+                    const int32x4_t acc_0123 = acc_s32[idx0];
+                    const int32x4_t acc_4567 = acc_s32[idx1];
+
+                    acc_f32[idx0] = vmlaq_f32(acc_f32[idx0], vcvtq_f32_s32(acc_0123), sbd_scale_0123[row]);
+                    acc_f32[idx1] = vmlaq_f32(acc_f32[idx1], vcvtq_f32_s32(acc_4567), sbd_scale_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q6_K_8x8_q8_K(int                        n,
                              float * GGML_RESTRICT      s,
                              size_t                     bs,

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -268,9 +268,9 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
                            _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 
-static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
-    return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -782,6 +782,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accum1 = _mm256_setzero_ps();
     __m256 accum2 = _mm256_setzero_ps();
+
     for (; ib + 1 < nb; ib += 2) {
         const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
         const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
@@ -795,10 +796,10 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
         const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
-                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
-                _mm256_cvtepi32_ps(p_2), accum2);
+        const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
+        const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
+        accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
     }
 
     sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
@@ -830,7 +831,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
         int sumi1 = 0;
         int sumi2 = 0;
         for (int j = 0; j < QK_MXFP4/2; ++j) {
@@ -3817,4 +3818,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/binary-ops.cpp

@@ -59,11 +59,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
     GGML_ASSERT(nb00 == sizeof(src0_t));
 
     const auto [ir0, ir1] = get_thread_range(params, src0);
-    const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
-
-    if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
-        GGML_ASSERT(ggml_are_same_shape(src0, src1));
-    }
+    const bool is_src1_contiguous_rows = ggml_is_contiguous_rows(src1);
 
 #ifdef GGML_USE_ACCELERATE
     vDSP_fn_t vDSP_op = nullptr;
@@ -94,7 +90,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
         const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
         const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
-        if (is_src1_contiguous) {
+        if (is_src1_contiguous_rows) {
             // src1 is broadcastable across src0 and dst in i1, i2, i3
             const int64_t nr0 = ne00 / ne10;
 

ggml/src/ggml-cpu/common.h

@@ -6,8 +6,8 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 
-#define GGML_FA_TILE_Q  32
-#define GGML_FA_TILE_KV 16
+#define GGML_FA_TILE_Q  64
+#define GGML_FA_TILE_KV 64
 
 #ifdef __cplusplus
 

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -24,6 +24,9 @@ struct ggml_compute_params {
     void * wdata;
 
     struct ggml_threadpool * threadpool;
+
+    // use reference implementation
+    bool use_ref;
 };
 
 

ggml/src/ggml-cpu/ggml-cpu.c

@@ -5,7 +5,6 @@
 #include "ggml-backend.h"
 #include "traits.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "quants.h"
 #include "ggml-threading.h"
@@ -76,6 +75,9 @@
 // precomputed f32 table for f16 (256 KB) (simd-mappings.h)
 float ggml_table_f32_f16[1 << 16];
 
+// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
+float ggml_table_f32_e8m0_half[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int sve_cnt;
@@ -2867,12 +2869,20 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
+                        const int64_t neq2 = node->src[0]->ne[2]; // number of query heads
                         const int64_t DK = node->src[1]->ne[0];
                         const int64_t DV = node->src[2]->ne[0];
 
                         // Tiled flash attention scratch (tile sizes defined in common.h)
-                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
-                        cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + K_f32 + padding
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV + GGML_FA_TILE_KV*DK)*n_tasks;
+
+                        // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
+                        // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
+                        size_t n_chunks = n_tasks;
+                        size_t decode   = sizeof(float)*(neq2*n_chunks*(2+DV) + n_tasks*(DK + 2*DV));
+
+                        cur += MAX(prefill, decode);
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {
@@ -2929,14 +2939,19 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     set_numa_thread_affinity(state->ith);
 
     struct ggml_compute_params params = {
-        /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
-        /*.wsize     =*/ cplan->work_size,
-        /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ tp,
+        /*.ith        =*/ state->ith,
+        /*.nth        =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
+        /*.wsize      =*/ cplan->work_size,
+        /*.wdata      =*/ cplan->work_data,
+        /*.threadpool =*/ tp,
+        /*.use_ref    =*/ cplan->use_ref,
     };
 
-    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
@@ -2963,7 +2978,11 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
-    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
 
     ggml_barrier(state->threadpool);
 
@@ -3673,6 +3692,11 @@ void ggml_cpu_init(void) {
                 ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
+            // initialize E8M0 half table (256 entries)
+            for (int i = 0; i < (1 << 8); ++i) {
+                ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
+            }
+
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
 
             GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -105,6 +105,8 @@ struct ggml_backend_cpu_context {
 
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
+
+    bool                use_ref;  // use reference implementation
 };
 
 static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
@@ -143,6 +145,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
 
     cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
     cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cpu_plan->cplan.use_ref             = cpu_ctx->use_ref;
 
     return cpu_plan;
 }
@@ -182,6 +185,7 @@ static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, s
 
     cplan.abort_callback      = cpu_ctx->abort_callback;
     cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cplan.use_ref             = cpu_ctx->use_ref;
 
     return ggml_graph_compute(cgraph, &cplan);
 }
@@ -223,6 +227,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     ctx->work_size           = 0;
     ctx->abort_callback      = NULL;
     ctx->abort_callback_data = NULL;
+    ctx->use_ref             = false;
 
     ggml_backend_t cpu_backend = new ggml_backend {
         /* .guid    = */ ggml_backend_cpu_guid(),
@@ -270,6 +275,13 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_
     ctx->abort_callback_data = abort_callback_data;
 }
 
+void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->use_ref = use_ref;
+}
+
 // CPU backend - device
 
 struct ggml_backend_cpu_device_context {
@@ -646,6 +658,9 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
     if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
         return (void *)ggml_is_numa;
     }
+    if (strcmp(name, "ggml_backend_cpu_set_use_ref") == 0) {
+        return (void *)ggml_backend_cpu_set_use_ref;
+    }
 
     // threadpool - TODO:  move to ggml-base
     if (strcmp(name, "ggml_threadpool_new") == 0) {

ggml/src/ggml-cpu/ops.cpp

@@ -3,6 +3,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "binary-ops.h"
+#include "simd-gemm.h"
 #include "ggml.h"
 #include "unary-ops.h"
 #include "vec.h"
@@ -2096,10 +2097,14 @@ static void ggml_compute_forward_gelu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2113,10 +2118,14 @@ static void ggml_compute_forward_gelu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2135,10 +2144,14 @@ static void ggml_compute_forward_gelu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2152,10 +2165,14 @@ static void ggml_compute_forward_gelu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2276,10 +2293,14 @@ static void ggml_compute_forward_gelu_erf_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2293,10 +2314,14 @@ static void ggml_compute_forward_gelu_erf_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2315,10 +2340,14 @@ static void ggml_compute_forward_gelu_erf_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2332,10 +2361,14 @@ static void ggml_compute_forward_gelu_erf_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2379,10 +2412,14 @@ static void ggml_compute_forward_gelu_quick_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2396,10 +2433,14 @@ static void ggml_compute_forward_gelu_quick_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2418,10 +2459,14 @@ static void ggml_compute_forward_gelu_quick_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2435,10 +2480,14 @@ static void ggml_compute_forward_gelu_quick_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2482,10 +2531,14 @@ static void ggml_compute_forward_silu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2499,10 +2552,14 @@ static void ggml_compute_forward_silu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -2521,10 +2578,14 @@ static void ggml_compute_forward_silu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2538,10 +2599,14 @@ static void ggml_compute_forward_silu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
@@ -7629,8 +7694,7 @@ static void ggml_compute_forward_pad_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-    GGML_ASSERT( dst->nb[0] == sizeof(float));
+    assert(dst->nb[0] == sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -8042,12 +8106,14 @@ void ggml_compute_forward_top_k(
     }
 }
 
-// ggml_compute_forward_flash_attn_ext
-
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
         ggml_tensor * dst,
-        int ir0, int ir1) {
+        int ir0, int ir1,
+        int64_t ic_start, int64_t ic_end,
+        float * partials, int64_t partial_stride) {
+
+    const bool write_partials = (partials != nullptr);
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -8124,7 +8190,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
     int ith = params->ith;
 
-    // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
         const int iq3 = ir/(neq2*neq1);
@@ -8165,7 +8230,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
 
-        for (int64_t ic = 0; ic < nek1; ++ic) {
+        for (int64_t ic = ic_start; ic < ic_end; ++ic) {
             const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
@@ -8238,8 +8303,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             }
         }
 
-        // sinks
-        if (sinks) {
+        // sinks - apply only on the first kv-chunk
+        if (sinks && ic_start == 0) {
             const float s = ((float *)((char *) sinks->data))[h];
 
             float ms = 1.0f;
@@ -8247,6 +8312,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
             if (s > M) {
                 ms = expf(M - s);
+                M = s;
                 ggml_vec_scale_f32(DV, VKQ32, ms);
             } else {
                 vs = expf(s - M);
@@ -8255,20 +8321,26 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             S = S*ms + vs;
         }
 
-        // V /= S
-        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
-        ggml_vec_scale_f32(DV, VKQ32, S_inv);
+        if (write_partials) {
+            // Write M, S, VKQ to partials for later reduction
+            // partials layout: [M, S, VKQ[DV]] per query head
+            float * partial = partials + ir * partial_stride;
+            partial[0] = M;
+            partial[1] = S;
+            memcpy(partial + 2, VKQ32, DV * sizeof(float));
+        } else {
+            // V /= S
+            const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
+            ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
-        // dst indices
-        const int i1 = iq1;
-        const int i2 = iq2;
-        const int i3 = iq3;
+            // dst indices
+            const int i1 = iq1;
+            const int i2 = iq2;
+            const int i3 = iq3;
 
-        // original
-        //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
-
-        // permute(0, 2, 1, 3)
-        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+        }
     }
 }
 
@@ -8318,10 +8390,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     GGML_ASSERT(k->type == v->type);
     const ggml_type kv_type = k->type;
 
-    const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type);
-    const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float;
-    const ggml_vec_dot_t    kv_vec_dot    = kv_type_traits_cpu->vec_dot;
-    const size_t kv_type_size = ggml_type_size(kv_type);
 
     // broadcast factors
     const int64_t rk2 = neq2/nek2;
@@ -8353,8 +8421,6 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
     static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
 
-    GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ");
-
     int ir = ir0;
     while (ir < ir1) {
         // q indices for the start of this tile
@@ -8381,18 +8447,20 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         }
 
         // Per-thread scratch layout:
-        // Q_q:    Q_TILE_SZ * DK (converted Q tile in KV type)
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile — F32 for GEMM, KV type for scalar)
         // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
         // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
         // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
-        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion)
-        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32);
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile)
+        // K_f32:  KV_TILE_SZ * DK (F32 buffer for K tile — GEMM path)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + KV_TILE_SZ*DK + CACHE_LINE_SIZE_F32);
 
         void  * Q_q    = base;
         float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
         float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
         float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
-        float * V32    = VKQ32 + Q_TILE_SZ * DV;  // F32 buffer for V tile
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;
+        float * K_f32  = V32 + KV_TILE_SZ * DV;
 
         memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
         memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
@@ -8405,28 +8473,38 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
         const int iv3 = iq3 / rv3;
         const int iv2 = iq2 / rv2;
 
-        for (int tq = 0; tq < tile_rows; tq++) {
-            const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
-            kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK);
-        }
-        // Zero-pad remaining rows
-        for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
-            memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size);
+        {
+            float * Q_f32 = (float *)Q_q;
+            for (int tq = 0; tq < tile_rows; tq++) {
+                const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+                memcpy(Q_f32 + tq * DK, pq, DK * sizeof(float));
+            }
+            for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+                memset(Q_f32 + tq * DK, 0, DK * sizeof(float));
+            }
         }
 
+        memset(K_f32, 0, DK * KV_TILE_SZ * sizeof(float));
+        memset(V32,   0, KV_TILE_SZ * DV * sizeof(float));
+
         for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, nek1 - ic);
 
             // skip the tile entirely if all the masks are -inf
             if (mask) {
                 bool can_skip = true;
                 for (int tq = 0; tq < tile_rows; tq++) {
                     const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    for (int tk = 0; tk < kv_tile; tk++) {
                         mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
                         if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
                             can_skip = false;
                         }
                     }
+                    // Pad remaining mask entries with -inf
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
 
                 if (can_skip) {
@@ -8434,13 +8512,32 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 }
             }
 
-            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                const void * q_row = (const char *)Q_q + tq * DK * kv_type_size;
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3);
-                    float s;
-                    kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1);
-                    KQ[tq * KV_TILE_SZ + tk] = s * scale;
+            // Pack K tile transposed: K_f32[dk][kv] so KV_TILE is contiguous (SIMD dim)
+            // Zero-pad the last tile so the GEMM always operates on KV_TILE_SZ columns
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * k_data = (const char *)k->data + (ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
+                if (kv_type == GGML_TYPE_F16) {
+                    const ggml_fp16_t * k_f16 = (const ggml_fp16_t *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = GGML_CPU_FP16_TO_FP32(k_f16[dk]);
+                    }
+                } else {
+                    const float * k_f32_src = (const float *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = k_f32_src[dk];
+                    }
+                }
+            }
+            memset(KQ, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+            simd_gemm(KQ, (const float *)Q_q, K_f32, Q_TILE_SZ, DK, KV_TILE_SZ);
+            ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, scale);
+
+            // Set padded KQ entries to -inf so softmax gives them zero weight
+            if (kv_tile < KV_TILE_SZ) {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        KQ[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
                 }
             }
 
@@ -8480,33 +8577,22 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
                 S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
             }
 
-            // Convert V tile to F32 first (if F16), then do MAD
-            // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster.
-            // TODO: on ARM, native f16 should be faster
-            if (kv_type == GGML_TYPE_F16) {
-                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                    const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                    ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV);
-                }
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p);
-                    }
-                }
-            } else {
-                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
-                    if (skip[tq]) continue;
-                    float * vkq_row = VKQ32 + tq * DV;
-                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
-                        const float p = KQ[tq * KV_TILE_SZ + tk];
-                        const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
-                        ggml_vec_mad_f32(DV, vkq_row, v_row, p);
-                    }
+            // V accumulation: VKQ32 += softmax(KQ) * V
+            // Pack V tile to contiguous F32, zero-padded
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
+                if (kv_type == GGML_TYPE_F16) {
+                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                } else {
+                    memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
                 }
             }
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                if (skip[tq]) {
+                    memset(KQ + tq * KV_TILE_SZ, 0, KV_TILE_SZ * sizeof(float));
+                }
+            }
+            simd_gemm(VKQ32, KQ, V32, Q_TILE_SZ, KV_TILE_SZ, DV);
         }
 
         // sinks (apply only to valid rows in the tile)
@@ -8546,6 +8632,78 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     }
 }
 
+// Reduction function: combines partial results across KV chunks
+// Partials layout in wdata: [n_q_heads][n_chunks][2 + DV]
+static void ggml_flash_attn_ext_reduce_partials(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        const int64_t n_chunks,
+        const int64_t chunk_size) {
+
+    const ggml_tensor * q = dst->src[0];
+    const ggml_tensor * k = dst->src[1];
+    const ggml_tensor * v = dst->src[2];
+
+    const int64_t DK        = k->ne[0];
+    const int64_t DV        = v->ne[0];
+    const int64_t nek1      = k->ne[1];
+    const int64_t n_q_heads = q->ne[2];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t wdata_per_thread = DK + 2*DV + CACHE_LINE_SIZE_F32;
+    float *       thread_wdata     = (float *) params->wdata + ith * wdata_per_thread;
+
+    const int64_t partials_offset  = nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
+    const int64_t partial_size     = 2 + DV;
+    const float * partials_base    = (const float *) params->wdata + partials_offset;
+
+    // Output layout
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const size_t  nb1 = dst->nb[1];
+
+    // Each thread reduces a subset of query heads
+    for (int64_t q_head = ith; q_head < n_q_heads; q_head += nth) {
+        float   M_final   = -INFINITY;
+        float   S_final   = 0.0f;
+        float * VKQ_final = thread_wdata;
+        memset(VKQ_final, 0, DV * sizeof(float));
+
+        // Combine partials from all chunks
+        for (int64_t chunk_idx = 0; chunk_idx < n_chunks; ++chunk_idx) {
+            const int64_t ic_start = chunk_idx * chunk_size;
+            if (ic_start >= nek1) continue;
+
+            const float * partial   = partials_base + (q_head * n_chunks + chunk_idx) * partial_size;
+            const float   M_chunk   = partial[0];
+            const float   S_chunk   = partial[1];
+            const float * VKQ_chunk = partial + 2;
+
+            if (S_chunk == 0.0f) continue;
+
+            const float M_new     = fmaxf(M_final, M_chunk);
+            const float scale_old = expf(M_final - M_new);
+            const float scale_new = expf(M_chunk - M_new);
+
+            for (int64_t d = 0; d < DV; ++d) {
+                VKQ_final[d] = VKQ_final[d] * scale_old + VKQ_chunk[d] * scale_new;
+            }
+            S_final = S_final * scale_old + S_chunk * scale_new;
+            M_final = M_new;
+        }
+
+        // Normalize and write to output
+        if (S_final != 0.0f) {
+            const float S_inv = 1.0f / S_final;
+            ggml_vec_scale_f32(DV, VKQ_final, S_inv);
+        }
+        // iq1=0, iq3=0 for decode
+        memcpy((char *) dst->data + (0*ne2*ne1 + q_head + 0*ne1)*nb1, VKQ_final, nb1);
+    }
+}
+
 static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8567,6 +8725,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const int64_t DV = nev0;
     const int64_t N  = neq1;
 
+
     GGML_ASSERT(ne0 == DV);
     GGML_ASSERT(ne2 == N);
 
@@ -8587,60 +8746,92 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int64_t nr = neq1*neq2*neq3;
-
-    // rows per thread
     const int ith = params->ith;
     const int nth = params->nth;
 
-    // disable for NUMA
-    const bool disable_chunking = ggml_is_numa();
+    // When use_ref is set, force the vec-only reference implementation (no tiling, no KV-chunking)
+    const bool use_ref = params->use_ref;
 
-    // 4x chunks per thread
-    int nth_scaled = nth * 4;
-    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
-    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
-
-    if (nth == 1 || nchunk < nth || disable_chunking) {
-        nchunk = nth;
-    }
-
-    if (ith == 0) {
-        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-        ggml_threadpool_chunk_set(params->threadpool, nth);
-    }
-
-    ggml_barrier(params->threadpool);
-
-    // The number of elements in each chunk
-    const int64_t dr = (nr + nchunk - 1) / nchunk;
-
-    static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
-    static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
     const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
-    const bool use_tiled = (q->type == GGML_TYPE_F32 &&
-                            kv_is_f32_or_f16 &&
-                            k->type == v->type &&
-                            nek1 % KV_TILE_SZ == 0 &&
-                            neq1 >= Q_TILE_SZ);  // Only use tiled for batch >= tile size
+    const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;
 
-    // The first chunk comes from our thread_id, the rest will get auto-assigned.
-    int current_chunk = ith;
+    if (use_split_kv_path) {
+        const int64_t chunk_size = (nek1 + nth - 1) / nth;
 
-    while (current_chunk < nchunk) {
-        const int64_t ir0 = dr * current_chunk;
-        const int64_t ir1 = MIN(ir0 + dr, nr);
+        // Partials buffer layout: [q_head][kv_chunk][M, S, VKQ]
+        const int64_t partial_size  = 2 + DV;
+        float *       partials_base = (float *) params->wdata + nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
 
-        if (use_tiled) {
-            ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+        const int64_t ic_start = ith * chunk_size;
+        const int64_t ic_end   = std::min(ic_start + chunk_size, nek1);
+
+        const int64_t partial_stride = nth * partial_size;
+        float *       chunk_partials = partials_base + ith * partial_size;
+
+        if (ic_start < nek1) {
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(
+                    params, dst, q_head, q_head + 1, ic_start, ic_end,
+                    chunk_partials, partial_stride);
+            }
         } else {
-            ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                float * q_partials = chunk_partials + q_head * partial_stride;
+                q_partials[0] = -INFINITY;  // M
+                q_partials[1] = 0.0f;       // S
+            }
         }
 
-        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        ggml_barrier(params->threadpool);
+        ggml_flash_attn_ext_reduce_partials(params, dst, nth, chunk_size);
+    } else {
+
+        // total rows in q
+        const int64_t nr = neq1*neq2*neq3;
+
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        if (ith == 0) {
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+        static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+        bool use_tiled = !use_ref &&
+                               (q->type == GGML_TYPE_F32 &&
+                                kv_is_f32_or_f16 &&
+                                k->type == v->type &&
+                                neq1 >= Q_TILE_SZ);
+#ifdef GGML_SIMD
+        use_tiled &= (DV % GGML_F32_EPR == 0);
+#endif
+        int current_chunk = ith;
+
+        while (current_chunk < nchunk) {
+            const int64_t ir0 = dr * current_chunk;
+            const int64_t ir1 = MIN(ir0 + dr, nr);
+
+            if (use_tiled) {
+                ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+            } else {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, nek1, nullptr, 0);
+            }
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
     }
 }
 

ggml/src/ggml-cpu/repack.cpp

@@ -256,6 +256,200 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
     ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
 }
 
+template <int M, int N>
+static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0f;
+        }
+
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                const int base_h = base_l + 64;
+
+                const int scale_idx_l = base_l / 16;
+                const int scale_idx_h = base_h / 16;
+
+                const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                const int qh_half_l = (base_l / 128) * 32;
+                const int qh_half_h = (base_h / 128) * 32;
+
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                    int sumi_l = 0;
+                    int sumi_h = 0;
+
+                    for (int i = 0; i < blocklen; i++) {
+                        const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                        const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
+                        const int qh_chunk_l  = qh_idx_l / blocklen;
+                        const int qh_pos_l    = qh_idx_l % blocklen;
+                        const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                        const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
+                        const int qh_chunk_h  = qh_idx_h / blocklen;
+                        const int qh_pos_h    = qh_idx_h % blocklen;
+                        const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                        const int8_t a_l = a_ptr[l].qs[base_l + i];
+                        const int8_t a_h = a_ptr[l].qs[base_h + i];
+
+                        sumi_l += q_l * a_l;
+                        sumi_h += q_h * a_h;
+                    }
+
+                    sumf[j] +=
+                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+        }
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+template <int M, int N>
+static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
+    const int     q8_half_stride    = 512;
+    const int     q8_low_high_step  = 256;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+
+    float sumf[4][8];
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0f;
+                }
+            }
+
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                    const int base_h = base_l + 64;
+
+                    const int scale_idx_l = base_l / 16;
+                    const int scale_idx_h = base_h / 16;
+
+                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                    const int qh_half_l = (base_l / 128) * 32;
+                    const int qh_half_h = (base_h / 128) * 32;
+
+                    const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4);
+
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                            int sumi_l = 0;
+                            int sumi_h = 0;
+
+                            for (int i = 0; i < blocklen; i++) {
+                                const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                                const int qh_idx_l   = qh_half_l + ((base_l + i) % 32);
+                                const int qh_chunk_l = qh_idx_l / blocklen;
+                                const int qh_pos_l   = qh_idx_l % blocklen;
+                                const int qh_offset_l =
+                                    qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                                const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                                const int qh_idx_h   = qh_half_h + ((base_h + i) % 32);
+                                const int qh_chunk_h = qh_idx_h / blocklen;
+                                const int qh_pos_h   = qh_idx_h % blocklen;
+                                const int qh_offset_h =
+                                    qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                                const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * blocklen + i];
+                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step];
+
+                                sumi_l += q_l * q8_l;
+                                sumi_h += q_h * q8_h;
+                            }
+
+                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
+                                          a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
 extern "C" {
 
 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -704,94 +898,12 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int                        n,
 }
 
 
+void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    constexpr int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-    UNUSED(nr);
-
-    float sumf[8];
-
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0f;
-        }
-
-        for (int l = 0; l < nb; l++) {
-
-
-            for (int k = 0; k < 16; k++) {
-                // k = 0.. 7 weights 0-63 low, 64-127 high
-                // k = 8..15 weights 128-191 low, 192-255 high
-                const int base_l = (k / 8) * 128 + (k % 8) * 8;
-                const int base_h = base_l + 64;
-
-                const int scale_idx_l = base_l / 16;
-                const int scale_idx_h = base_h / 16;
-
-                // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
-                const int qh_shift_l = ((base_l % 128) / 32) * 2;
-                const int qh_shift_h = ((base_h % 128) / 32) * 2;
-
-                // qh_half: offset to the correct 32-byte half (0 or 32)
-                const int qh_half_l = (base_l / 128) * 32;
-                const int qh_half_h = (base_h / 128) * 32;
-
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    // Interleaved scales
-                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
-                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
-
-                    int sumi_l = 0;
-                    int sumi_h = 0;
-
-                    for (int i = 0; i < blocklen; i++) {
-                        const int ql_pos = k * 64 + j * 8 + i;
-                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
-                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
-
-                        // qh indexing with 8-byte interleaving (like q5_K)
-                        const int qh_byte_l   = qh_half_l + ((base_l + i) % 32);
-                        const int qh_chunk_l  = qh_byte_l / 8;
-                        const int qh_pos_l    = qh_byte_l % 8;
-                        const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
-                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
-
-                        const int qh_byte_h   = qh_half_h + ((base_h + i) % 32);
-                        const int qh_chunk_h  = qh_byte_h / 8;
-                        const int qh_pos_h    = qh_byte_h % 8;
-                        const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
-                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
-
-                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
-                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
-
-                        const int8_t a_l = a_ptr[l].qs[base_l + i];
-                        const int8_t a_h = a_ptr[l].qs[base_h + i];
-
-                        sumi_l += q_l * a_l;
-                        sumi_h += q_h * a_h;
-                    }
-
-                    sumf[j] +=
-                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-        }
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1485,109 +1597,12 @@ void ggml_gemm_q5_K_8x8_q8_K_generic(int                        n,
     }
 }
 
-void ggml_gemm_q6_K_8x8_q8_K_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK_K;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen          = 8;
+void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
 
-    assert(n % qk == 0);
-    assert(nr % 4 == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-
-    float sumf[4][8];
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
-
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0f;
-                }
-            }
-
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < 16; k++) {
-                    // k = 0.. 7 weights 0-63 low, 64-127 high
-                    // k = 8..15 weights 128-191 low, 192-255 high
-                    const int base_l = (k / 8) * 128 + (k % 8) * 8;
-                    const int base_h = base_l + 64;
-
-                    const int scale_idx_l = base_l / 16;
-                    const int scale_idx_h = base_h / 16;
-
-                    // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
-                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
-                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
-
-                    // qh_half: offset to the correct 32-byte half (0 or 32)
-                    const int qh_half_l = (base_l / 128) * 32;
-                    const int qh_half_h = (base_h / 128) * 32;
-
-                    // Activation base indices for q8_Kx4 interleaved format
-                    // Layout: 128-value halves (k/8), then 8-value sub-blocks (k%8) with stride 32
-                    const int q8_base = (k / 8) * 512 + (k % 8) * 32;
-
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            // Interleaved scales
-                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
-                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
-
-                            int sumi_l = 0;
-                            int sumi_h = 0;
-
-                            for (int i = 0; i < blocklen; i++) {
-                                const int ql_pos = k * 64 + j * 8 + i;
-                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
-                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
-
-                                const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
-                                const int qh_chunk_l  = qh_idx_l / 8;
-                                const int qh_pos_l    = qh_idx_l % 8;
-                                const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
-                                const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
-
-                                const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
-                                const int qh_chunk_h  = qh_idx_h / 8;
-                                const int qh_pos_h    = qh_idx_h % 8;
-                                const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
-                                const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
-
-                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
-                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
-
-                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * 8 + i];
-                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * 8 + i + 256];
-
-                                sumi_l += q_l * q8_l;
-                                sumi_h += q_h * q8_h;
-                            }
-
-                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
-                                          a_ptr[l].d[m];
-                        }
-                    }
-                }
-            }
-
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+   ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }
 
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1901,9 +1916,10 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
         int src_offset = (i / 8) * blck_size_interleave;
         int dst_offset = i * blck_size_interleave;
 
+        // buffer large enough for the max interleave block size (8 bytes)
         uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&elems, &in[src_id].qs[src_offset], blck_size_interleave);
+        memcpy(&out.qs[dst_offset], &elems, blck_size_interleave);
     }
 
     // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K
@@ -2097,18 +2113,18 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
     }
 
     const int end_ls = QK_K * 4 / blck_size_interleave;
-    // Interleave Q6_K quants by taking 8 bytes at a time
+    // Interleave Q6_K quants by taking blck_size_interleave bytes at a time
     for (int i = 0; i < end_ls; ++i) {
         int src_id     = i % n_blocks;
         int src_offset = (i / n_blocks) * blck_size_interleave;
         int dst_offset = i * blck_size_interleave;
 
         uint64_t elem_ls;
-        memcpy(&elem_ls, &in[src_id].ql[src_offset], sizeof(uint64_t));
-        memcpy(&out.ql[dst_offset], &elem_ls, sizeof(uint64_t));
+        memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave);
+        memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave);
     }
 
-    // Interleave high bits using same 8-byte pattern as low bits
+    // Interleave high bits using same chunk size as low bits
     const int end_hs = end_ls / 2;
     for (int i = 0; i < end_hs; ++i) {
         int src_id     = i % n_blocks;
@@ -2116,8 +2132,8 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
         int dst_offset = i * blck_size_interleave;
 
         uint64_t elem_hs;
-        memcpy(&elem_hs, &in[src_id].qh[src_offset], sizeof(uint64_t));
-        memcpy(&out.qh[dst_offset], &elem_hs, sizeof(uint64_t));
+        memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave);
+        memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave);
     }
 
     // The below logic is designed so as to unpack and rearrange scales in Q6_K
@@ -2262,7 +2278,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,
 
 static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
     constexpr int nrows_interleaved = 8;
 
     block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
@@ -2511,6 +2527,10 @@ template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size);
 }
@@ -2575,6 +2595,10 @@ template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -2634,6 +2658,10 @@ template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -3043,6 +3071,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
 
     // instance for Q6_K
+    static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;
 
     // instance for Q2
@@ -3107,6 +3136,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q6_K_8x8_q8_K;
             }
         }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q6_K_8x4_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {

ggml/src/ggml-cpu/repack.h

@@ -112,6 +112,7 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -122,6 +123,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -142,6 +144,7 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -152,6 +155,7 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);

ggml/src/ggml-cpu/simd-gemm.h

@@ -0,0 +1,136 @@
+#pragma once
+
+// Computes C[M x N] += A[M x K] * B[K x N]
+
+#include "simd-mappings.h"
+
+// TODO: add support for sizeless vector types
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+
+// TODO: untested on avx512
+// These are in units of GGML_F32_EPR
+#if defined(__AVX512F__) || defined (__ARM_NEON__)
+    static constexpr int GEMM_RM = 4;
+    static constexpr int GEMM_RN = 4; // 16+4+1 = 25/32
+#elif defined(__AVX2__) || defined(__AVX__)
+    static constexpr int GEMM_RM = 6;
+    static constexpr int GEMM_RN = 2; // 12+2+1 = 15/16
+#else
+    static constexpr int GEMM_RM = 2;
+    static constexpr int GEMM_RN = 2;
+#endif
+
+template <int RM, int RN>
+static inline void simd_gemm_ukernel(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    GGML_F32_VEC acc[RM][RN];
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            acc[i][r] = GGML_F32_VEC_LOAD(C + i * N + r * KN);
+        }
+    }
+
+    for (int64_t kk = 0; kk < K; kk++) {
+        GGML_F32_VEC Bv[RN];
+        for (int r = 0; r < RN; r++) {
+            Bv[r] = GGML_F32_VEC_LOAD(B + kk * N + r * KN);
+        }
+        for (int64_t i = 0; i < RM; i++) {
+            GGML_F32_VEC p = GGML_F32_VEC_SET1(A[i * K + kk]);
+            for (int r = 0; r < RN; r++) {
+                acc[i][r] = GGML_F32_VEC_FMA(acc[i][r], Bv[r], p);
+            }
+        }
+    }
+
+    for (int64_t i = 0; i < RM; i++) {
+        for (int r = 0; r < RN; r++) {
+            GGML_F32_VEC_STORE(C + i * N + r * KN, acc[i][r]);
+        }
+    }
+}
+
+// C[M x N] += A[M x K] * B[K x N]
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    static constexpr int KN = GGML_F32_EPR;
+
+    int64_t ii = 0;
+    for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<GEMM_RM, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<GEMM_RM, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            for (int64_t i = 0; i < GEMM_RM; i++) {
+                float a = C[i * N + jj];
+                for (int64_t kk = 0; kk < K; kk++) {
+                    a += A[i + kk] * B[kk * N + jj];
+                }
+                C[i * N + jj] = a;
+            }
+        }
+
+        A += GEMM_RM * K;
+        C += GEMM_RM * N;
+    }
+
+    // Tail rows: one at a time
+    for (; ii < M; ii++) {
+        int64_t jj = 0;
+        for (; jj + GEMM_RN * KN <= N; jj += GEMM_RN * KN) {
+            simd_gemm_ukernel<1, GEMM_RN>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj + KN <= N; jj += KN) {
+            simd_gemm_ukernel<1, 1>(C + jj, A, B + jj, K, N);
+        }
+        for (; jj < N; jj++) {
+            float a = C[jj];
+            for (int64_t kk = 0; kk < K; kk++) {
+                a += A[kk] * B[kk * N + jj];
+            }
+            C[jj] = a;
+        }
+
+        A += K;
+        C += N;
+    }
+}
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic pop
+#endif
+
+#else // scalar path
+
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    for (int64_t i = 0; i < M; i++) {
+        for (int64_t j = 0; j < N; j++) {
+            float sum = C[i * N + j];
+            for (int64_t kk = 0; kk < K; kk++) {
+                sum += A[i * K + kk] * B[kk * N + j];
+            }
+            C[i * N + j] = sum;
+        }
+    }
+}
+
+#endif // GGML_SIMD

ggml/src/ggml-cpu/simd-mappings.h

@@ -116,6 +116,17 @@ extern "C" {
 // defined in ggml-cpu.c, initialized in ggml_cpu_init()
 extern float ggml_table_f32_f16[1 << 16];
 
+// precomputed f32 table for e8m0 half (1 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_e8m0_half[1 << 8];
+
+// Use lookup table for E8M0 on x86 (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
+#else
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
+#endif
+
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
@@ -1149,6 +1160,14 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     float32x4_t tmp = x[0] + vec_reve(x[0]);        \
     res = tmp[0] + tmp[1];                          \
 }
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
+{                                                \
+    float32x4_t v = vec_add(vec_add(s0, s1),     \
+                            vec_add(s2, s3));    \
+    v = vec_add(v, vec_sld(v, v, 8));            \
+    v = vec_add(v, vec_sld(v, v, 4));            \
+    res += (ggml_float)vec_extract(v, 0);        \
+}
 
 #define GGML_F32_VEC        GGML_F32x4
 #define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
@@ -1198,6 +1217,24 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_F16_VEC_MUL            GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
 
+// BF16 s390x
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR  8
+
+#define GGML_BF16x8         __vector unsigned short
+#define GGML_BF16x8_ZERO    vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC      GGML_BF16x8
+#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
+#define GGML_BF16_TO_F32_LO(v) ((float32x4_t) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((float32x4_t) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_FMA_LO(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
 #elif defined(__riscv_v_intrinsic)
 
 // compatible with vlen >= 128

ggml/src/ggml-cpu/unary-ops.cpp

@@ -111,7 +111,7 @@ template <float (*op)(float), typename src0_t, typename dst_t>
 static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_contiguous_rows(src0) && ggml_is_contiguous_rows(dst) && ggml_are_same_shape(src0, dst));
 
     GGML_TENSOR_UNARY_OP_LOCALS
 

ggml/src/ggml-cpu/vec.cpp

@@ -236,8 +236,7 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
     vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
     sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
 
-#endif
-#if defined(__POWER9_VECTOR__)
+#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
     const int np = (n & ~(GGML_BF16_STEP - 1));
     if (np > 0) {
         GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};

ggml/src/ggml-cuda/CMakeLists.txt

@@ -64,7 +64,7 @@ if (CUDAToolkit_FOUND)
         FetchContent_Declare(
             CCCL
             GIT_REPOSITORY https://github.com/nvidia/cccl.git
-            GIT_TAG        v3.2.0-rc2
+            GIT_TAG        v3.2.0
             GIT_SHALLOW    TRUE
         )
 

ggml/src/ggml-cuda/binbcast.cu

@@ -39,13 +39,16 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
                                    const uint3            ne11,
                                    const uint3            ne12,
                                    const uint3            ne13,
-                                   /*int s0, */ const int s1,
+                                 /*const int              s0,*/
+                                   const int              s1,
                                    const int              s2,
                                    const int              s3,
-                                   /*int s00,*/ const int s01,
+                                   const int              s00,
+                                   const int              s01,
                                    const int              s02,
                                    const int              s03,
-                                   /*int s10,*/ const int s11,
+                                   const int              s10,
+                                   const int              s11,
                                    const int              s12,
                                    const int              s13,
                                    src1_ptrs... src1s) {
@@ -72,11 +75,11 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
     for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
         const uint32_t i10 = fastmodulo(i0, ne10);
 
-        float result = src0_row ? (float) src0_row[i0] : 0.0f;
+        float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
         if constexpr (sizeof...(src1_ptrs) > 0) {
-            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
         } else {
-            result = bin_op(result, (float)src1[i_src1 + i10]);
+            result = bin_op(result, (float)src1[i_src1 + i10*s10]);
         }
 
         dst_row[i0] = (dst_t) result;
@@ -101,13 +104,16 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
                                            const uint3            ne11,
                                            const uint3            ne12,
                                            const uint3            ne13,
-                                           /*int s0, */ const int s1,
+                                         /*const int              s0,*/
+                                           const int              s1,
                                            const int              s2,
                                            const int              s3,
-                                           /*int s00,*/ const int s01,
+                                           const int              s00,
+                                           const int              s01,
                                            const int              s02,
                                            const int              s03,
-                                           /*int s10,*/ const int s11,
+                                           const int              s10,
+                                           const int              s11,
                                            const int              s12,
                                            const int              s13,
                                            src1_ptrs... src1s) {
@@ -135,11 +141,11 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
 
     const int i10 = fastmodulo(i0, ne10);
 
-    float result = src0_row ? (float) src0_row[i0] : 0.0f;
+    float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
     if constexpr (sizeof...(src1_ptrs) > 0) {
-        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
     } else {
-        result = bin_op(result, (float)src1[i_src1 + i10]);
+        result = bin_op(result, (float)src1[i_src1 + i10*s10]);
     }
 
     dst_row[i0] = (dst_t) result;
@@ -179,7 +185,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         cnb[3] *= cne[3];
     };
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
         for (int i = 0; i < 4; i++) {
             if (nr[i] != 1) {
                 break;
@@ -221,7 +227,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         size_t nb12 = cnb1[2];
         size_t nb13 = cnb1[3];
 
-        size_t s0 = nb0 / sizeof(dst_t);
+      //size_t s0 = nb0 / sizeof(dst_t);
         size_t s1 = nb1 / sizeof(dst_t);
         size_t s2 = nb2 / sizeof(dst_t);
         size_t s3 = nb3 / sizeof(dst_t);
@@ -251,10 +257,6 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
         GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
-        GGML_ASSERT(s0 == 1);
-        GGML_ASSERT(s00 == 1);
-        GGML_ASSERT(s10 == 1);
-
         const int block_size = 128;
 
         int64_t hne0 = std::max(ne0 / 2LL, 1LL);
@@ -284,31 +286,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
                 k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
                     ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
                 k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
                     <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
                                                            ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
-                                                           /* s0, */ s1, s2, s3,
-                                                           /* s00,*/ s01, s02, s03,
-                                                           /* s10,*/ s11, s12, s13);
+                                                         /*s0,*/ s1,  s2,  s3,
+                                                           s00, s01, s02, s03,
+                                                           s10, s11, s12, s13);
             }
         } else {
             const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
             if constexpr (sizeof...(I) > 0) {
                 k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1, s2,  s3,
+                    s00 ,s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
                 k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13);
             }
         }
     }

ggml/src/ggml-cuda/convert.cu

@@ -7,7 +7,8 @@
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = 2 * (int64_t(blockDim.x)*blockIdx.x + threadIdx.x);
 
@@ -16,23 +17,27 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     }
 
     const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
 
-    const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
+    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+        const int64_t i02 = dm.y;
+        const int64_t i03 = dm.x;
 
-    const int64_t ib = ibx0 + i00/qk; // block index
-    const int64_t iqs = (i00%qk)/qr; // quant index
-    const int64_t iybs = i00 - i00%qk; // y block start index
-    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+        const int64_t ibx0 = i03*s03 + i02*s02 + i01*s01;
 
-    // dequantize
-    float2 v;
-    dequantize_kernel(vx, ib, iqs, v);
+        const int64_t ib = ibx0 + i00/qk; // block index
+        const int64_t iqs = (i00%qk)/qr; // quant index
+        const int64_t iybs = i00 - i00%qk; // y block start index
+        const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
-    const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
-    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+        // dequantize
+        float2 v;
+        dequantize_kernel(vx, ib, iqs, v);
+
+        const int64_t iy0 = (i0203*ne01 + i01)*ne00 + iybs + iqs;
+        y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+        y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+    }
 }
 
 template <bool need_check>
@@ -485,9 +490,11 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE), ne01, (int)std::min(ne0203, (int64_t)65535));
     dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -612,7 +619,8 @@ static void dequantize_row_mxfp4_cuda(const void * vx, dst_t * y, const int64_t
 
 template <typename src_t, typename dst_t>
 static __global__ void convert_unary(
-        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
+        const int64_t ne0203, const uint3 ne02,
         const int64_t s01, const int64_t s02, const int64_t s03) {
     const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -621,23 +629,29 @@ static __global__ void convert_unary(
     }
 
     const int64_t i01 = blockIdx.y;
-    const int64_t i02 = blockIdx.z % ne02;
-    const int64_t i03 = blockIdx.z / ne02;
 
     const src_t * x = (const src_t *) vx;
 
-    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
-    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+    for (int64_t i0203 = blockIdx.z; i0203 < ne0203; i0203 += gridDim.z) {
+        const uint2 dm = fast_div_modulo((uint32_t)i0203, ne02);
+        const int64_t i02 = dm.y;
+        const int64_t i03 = dm.x;
+
+        const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+        const int64_t iy = (i0203*ne01 + i01)*ne00 + i00;
+        y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
+    }
 }
 
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * vx, dst_t * y,
         const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
         const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
-    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    const int64_t ne0203 = ne02*ne03;
+    const uint3 ne02_fdv = init_fastdiv_values(ne02);
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, (int)std::min(ne0203, (int64_t)65535));
     convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
-        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+        (vx, y, ne00, ne01, ne0203, ne02_fdv, s01, s02, s03);
 }
 
 template <typename src_t, typename dst_t>

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -63,11 +63,19 @@ static __global__ void flash_attn_ext_f16(
     constexpr int frag_m = ncols == 8 ? 32 : 16;
     constexpr int frag_n = ncols == 8 ?  8 : 16;
     static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+#if defined(GGML_USE_HIP)
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::row_major> frag_a_K;
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_a_V;
+    typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_b;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, _Float16>                          frag_c_VKQ;
+#else
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::row_major> frag_a_K;
     typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::col_major> frag_a_V;
     typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, half, wmma::col_major> frag_b;
     typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
     typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+#endif
 
     constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
     constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
@@ -126,6 +134,19 @@ static __global__ void flash_attn_ext_f16(
 
     __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
     half2 * VKQ2 = (half2 *) VKQ;
+
+#if defined(GGML_USE_HIP)
+    const _Float16 * K_h_f16  = reinterpret_cast<const _Float16 *>(K_h);
+    const _Float16 * V_h_f16  = reinterpret_cast<const _Float16 *>(V_h);
+    _Float16       * KQ_f16   = reinterpret_cast<_Float16 *>(KQ);
+    _Float16       * VKQ_f16  = reinterpret_cast<_Float16 *>(VKQ);
+#else
+    const half * K_h_f16  = K_h;
+    const half * V_h_f16  = V_h;
+    half       * KQ_f16   = KQ;
+    half       * VKQ_f16  = VKQ;
+#endif
+
 #pragma unroll
     for (int j0 = 0; j0 < ncols; j0 += nwarps) {
         const int j = j0 + threadIdx.y;
@@ -160,7 +181,7 @@ static __global__ void flash_attn_ext_f16(
     for (int i0 = 0; i0 < D; i0 += 16) {
 #pragma unroll
         for (int j0 = 0; j0 < ncols; j0 += frag_n) {
-            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ_f16 + j0*D_padded + i0, D_padded);
         }
     }
 
@@ -180,7 +201,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
             for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
                 frag_a_K K_a;
-                wmma::load_matrix_sync(K_a, K_h + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+                wmma::load_matrix_sync(K_a, K_h_f16 + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
@@ -310,7 +331,7 @@ static __global__ void flash_attn_ext_f16(
                 const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
                 wmma::load_matrix_sync(
                     KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
-                    KQ + j0*(kqar*kqs_padded) + k,
+                    KQ_f16 + j0*(kqar*kqs_padded) + k,
                     kqar*kqs_padded);
             }
         }
@@ -328,7 +349,7 @@ static __global__ void flash_attn_ext_f16(
                 const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
 
                 frag_a_V v_a;
-                wmma::load_matrix_sync(v_a, V_h + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+                wmma::load_matrix_sync(v_a, V_h_f16 + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
 #pragma unroll
                 for (int j = 0; j < ncols/frag_n; ++j) {
                     wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
@@ -344,7 +365,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
             for (int j0 = 0; j0 < ncols; j0 += frag_n) {
                 wmma::store_matrix_sync(
-                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    KQ_f16 + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
                     VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
                     D_padded, wmma::mem_col_major);
             }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2279,13 +2279,19 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
     if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        if (ne2 == 1) {
+        static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
+        if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                if (ne2 <= 4) {
+                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                    return;
+                }
             } else {
-                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                if (GGML_CUDA_CC_IS_AMD(cc)) {
+                    ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                    return;
+                }
             }
-            return;
         }
 
         if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
@@ -2866,6 +2872,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
     const std::string ffn_moe_down_bias_prefix = "ffn_moe_down_biased";
     const std::string nemotron_h_block_out_prefix = "nemotron_h_block_out";
     const std::string mamba2_y_add_d_prefix = "mamba2_y_add_d";
+    const std::string delta_net_prefix = "dnet_add";
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -2896,7 +2903,8 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
             strncmp(node->name, ffn_moe_up_bias_prefix.c_str(), ffn_moe_up_bias_prefix.size()) != 0 &&
             strncmp(node->name, ffn_moe_down_bias_prefix.c_str(), ffn_moe_down_bias_prefix.size()) != 0 &&
             strncmp(node->name, nemotron_h_block_out_prefix.c_str(), nemotron_h_block_out_prefix.size()) != 0 &&
-            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0) {
+            strncmp(node->name, mamba2_y_add_d_prefix.c_str(), mamba2_y_add_d_prefix.size()) != 0 &&
+            strncmp(node->name, delta_net_prefix.c_str(), delta_net_prefix.size()) != 0) {
             // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
             // by means of matching node names. See
             // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
@@ -2973,8 +2981,7 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         }
     }
 
-    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
-        memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+    if (memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
         return false;
     }
 
@@ -3635,11 +3642,13 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         n_fuse++;
 
                         if (n_fuse > 1) {
+                            ggml_tensor fused_add_node;
+                            memcpy(&fused_add_node, node, sizeof(ggml_tensor));
                             for (int j = 0; j < n_fuse - 1; ++j) {
-                                node->src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+                                fused_add_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
                             }
-                            cgraph->nodes[i + n_fuse - 1]->data = node->data;
-                            ggml_cuda_op_fused_add(*cuda_ctx, node, n_fuse);
+                            fused_add_node.data = cgraph->nodes[i + n_fuse - 1]->data;
+                            ggml_cuda_op_fused_add(*cuda_ctx, &fused_add_node, n_fuse);
                             i += n_fuse - 1;
 
                             continue;
@@ -4537,6 +4546,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
                 case GGML_UNARY_OP_TRUNC:
+                    // TODO: should become:
+                    //return ggml_is_contiguous_rows(op->src[0]);
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -4815,8 +4826,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
-        case GGML_OP_ACC:
             return true;
+        case GGML_OP_ACC:
+            // TODO: extend support like so:
+            //return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]);
+            return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
         case GGML_OP_SUM:
             return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_TOP_K:
@@ -4829,8 +4843,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_PAD:
             return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_PAD:
+            return true;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
@@ -5049,16 +5064,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
-            // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
-            // PR: https://github.com/ggml-org/llama.cpp/pull/19042
-            if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
-#ifdef _WIN32
-                _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
-#else
-                setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
-#endif // _WIN32
-            }
-
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
             const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 

ggml/src/ggml-cuda/mmq.cuh

@@ -2715,14 +2715,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
 #pragma unroll
         for (int l = 0; l < QR2_XXS; ++l) {
-            const int * grid_pos = (const int *) (iq2xxs_grid + aux8[l]);
-            const int signs_packed = ksigns_iq2xs[(aux32 >> (7*l)) & 0x7F];
+            const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]];
+            const uint32_t signs = unpack_ksigns(aux32 >> (7 * l));
 
-            const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-            const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-            const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
@@ -2733,12 +2733,12 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         }
 
-        const int ls = aux32 >> 28;
+        const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
         const float d = bxi->d;
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
-        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = (ls*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0   + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #else
-        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4;
+        x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)  || defined(AMD_WMMA_AVAILABLE)
     }
 }
@@ -2776,11 +2776,14 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 
     #pragma unroll
         for (int l = 0; l < QR2_XS; ++l) {
-            const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l] & 0x000001FF));
-            const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l] >> 9));
+            const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF];
+            const uint32_t signs = unpack_ksigns(q2[l] >> 9);
 
-            const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
@@ -2904,11 +2907,13 @@ template <int mmq_y, bool need_check> static __device__ __forceinline__ void loa
 #pragma unroll
         for (int l = 0; l < QR3_XXS; ++l) {
             const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
+            const uint32_t signs = unpack_ksigns(aux32 >> (7*l));
 
-            const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l)) & 0x7F));
+            const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
-            const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-            const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+            const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;
@@ -3697,13 +3702,20 @@ static __global__ void mul_mat_q(
          tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
-
 template <ggml_type type, int mmq_x, bool need_check>
-static __global__ void mul_mat_q_stream_k_fixup(
-        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
-        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
-        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst,
-        const int ncols_max) {
+static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
+                                                const int32_t * expert_bounds,
+                                                float * __restrict__ dst,
+                                                const float * __restrict__ tmp_last_tile,
+                                                const int    ncols_x,
+                                                const int    nrows_x,
+                                                const int    ncols_dst,
+                                                const size_t stride_col_dst,
+                                                const int    nchannels_y,
+                                                const size_t stride_channel_dst,
+                                                const int    nsamples_y,
+                                                const size_t stride_sample_dst,
+                                                const int    ncols_max) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     ITER_K          = get_iter_k(type);

ggml/src/ggml-cuda/mmvf.cu

@@ -4,26 +4,48 @@
 #include "mmvf.cuh"
 #include "convert.cuh"
 
-template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false, bool is_multi_token_id = false>
 static __global__ void mul_mat_vec_f(
         const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
-        const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
+        const int ncols2, const uint3 nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
         const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const int ids_stride) {
     const int row         = blockIdx.x;
+    // for MUL_MAT_ID - blockIdx.y = n_expert_used, blockIdx.z = ncols_dst (tokens)
     const int channel_dst = blockIdx.y;
-    const int channel_x   = ids ? ids[channel_dst]          : fastdiv((uint32_t) channel_dst, channel_ratio);
-    const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
-    const int sample_dst  = blockIdx.z;
+    const int tid         = threadIdx.x;
+
+    int token_idx;
+    int channel_x;
+    int channel_y;
+    int sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        token_idx  = ids ? blockIdx.z                                          : 0;
+        channel_x  = ids ? ids[blockIdx.y + token_idx * ids_stride]            : fastdiv((uint32_t) channel_dst, channel_ratio);
+        channel_y  = ids ? fastmodulo(blockIdx.y, nchannels_y)                 : channel_dst;
+        sample_dst = ids ? 0                                                   : blockIdx.z;
+    }
+
     const int sample_x    = fastdiv((uint32_t) sample_dst, sample_ratio);
     const int sample_y    = sample_dst;
-    const int tid         = threadIdx.x;
 
     constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
     x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
     y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
     dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
+    if constexpr (is_multi_token_id) {
+        y   += token_idx*stride_col_y2*2;
+        dst += token_idx*stride_col_dst;
+    }
 
     bool use_gate = false;
     bool use_bias = false;
@@ -56,8 +78,10 @@ static __global__ void mul_mat_vec_f(
     if (use_gate) {
         gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
     }
+
+    const int channel_bias = ids ? channel_x : channel_dst;
+
     if constexpr (has_fusion) {
-        const int channel_bias = ids ? channel_x : channel_dst;
         if (use_bias) {
             x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
         }
@@ -349,36 +373,36 @@ static __global__ void mul_mat_vec_f(
     }
 }
 
-template<typename T, typename type_acc, int ncols_dst, int block_size>
+template<typename T, typename type_acc, int ncols_dst, int block_size, bool is_multi_token_id = false>
 static void mul_mat_vec_f_switch_fusion(
         const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
-        const int64_t ncols, const int64_t nrows,
+        const int64_t ncols, const uint3 nchannels_y,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const int ids_stride, const cudaStream_t stream) {
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
        }
     }
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 
 }
 
-template <typename T, typename type_acc, int ncols_dst>
+template <typename T, typename type_acc, int ncols_dst, bool is_multi_token_id = false>
 void launch_mul_mat_vec_f_cuda(
         const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows,
@@ -386,12 +410,13 @@ void launch_mul_mat_vec_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        const int64_t nsamples_or_ntokens, const int64_t ids_stride, cudaStream_t stream) {
     GGML_ASSERT(ncols        % 2 == 0);
     GGML_ASSERT(stride_row   % 2 == 0);
     GGML_ASSERT(stride_col_y % 2 == 0);
     GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
     GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const uint3 nchannels_y_fd   = ids ? init_fastdiv_values(nchannels_y) : make_uint3(0, 0, 0);
     const uint3 channel_ratio_fd = ids ? make_uint3(0, 0, 0) : init_fastdiv_values(nchannels_dst / nchannels_x);
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
@@ -415,56 +440,56 @@ void launch_mul_mat_vec_f_cuda(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
 
     const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
-    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_or_ntokens);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case   64: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case   96: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  128: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  160: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  192: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  224: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  256: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -480,55 +505,88 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        const int64_t ids_stride, cudaStream_t stream) {
+
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst, true>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
+    if (has_ids) {
+        // Single-token MUL_MAT_ID path
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
     switch (ncols_dst) {
         case 1:
             launch_mul_mat_vec_f_cuda<T, type_acc, 1>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 2:
             launch_mul_mat_vec_f_cuda<T, type_acc, 2>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 3:
             launch_mul_mat_vec_f_cuda<T, type_acc, 3>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 4:
             launch_mul_mat_vec_f_cuda<T, type_acc, 4>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 5:
             launch_mul_mat_vec_f_cuda<T, type_acc, 5>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 6:
             launch_mul_mat_vec_f_cuda<T, type_acc, 6>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 7:
             launch_mul_mat_vec_f_cuda<T, type_acc, 7>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 8:
             launch_mul_mat_vec_f_cuda<T, type_acc, 8>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -544,21 +602,21 @@ static void mul_mat_vec_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        enum ggml_prec prec, cudaStream_t stream) {
+        const int64_t ids_stride, enum ggml_prec prec, cudaStream_t stream) {
 
     if constexpr(std::is_same_v<T, half>) {
         if (prec == GGML_PREC_DEFAULT) {
             mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
                 (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             return;
         }
     }
     mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
         (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
 }
 
 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
@@ -573,7 +631,7 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const size_t ts_src1 = ggml_type_size(src1->type);
     const size_t ts_dst  = ggml_type_size(dst->type);
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVF_MAX_BATCH_SIZE);
     GGML_ASSERT(ne13 == ne3);
 
     GGML_ASSERT(        nb00       == ts_src0);
@@ -626,29 +684,31 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const int64_t ncols_dst          = ids ? ne2  : ne1;
     const int64_t nchannels_y        = ids ? ne11 : ne12;
     const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -695,19 +755,19 @@ void ggml_cuda_op_mul_mat_vec_f(
             const float * src0_d = (const float *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));

ggml/src/ggml-cuda/mmvf.cuh

@@ -1,5 +1,7 @@
 #include "common.cuh"
 
+#define MMVF_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVF kernels.
+
 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
     const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 

ggml/src/ggml-cuda/mmvq.cu

@@ -137,15 +137,15 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-// tell the compiler to use as many registers as it wants, see nwarps definition below
-template <ggml_type type, int ncols_dst, bool has_fusion>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
-        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst) {
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const uint32_t ids_stride) {
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
@@ -162,11 +162,25 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
     const uint32_t channel_dst = blockIdx.y;
-    const uint32_t channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-    const uint32_t channel_y   = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-    const uint32_t sample_dst  = blockIdx.z;
+
+    uint32_t token_idx = 0;
+    uint32_t channel_x;
+    uint32_t channel_y;
+    uint32_t sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+        sample_dst = blockIdx.z;
+    }
+
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
@@ -188,11 +202,11 @@ static __global__ void mul_mat_vec_q(
         active_glu    = fusion.glu_op;
     }
 
-    const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
     float x_biases[ncols_dst]    = { 0.0f };
     float gate_biases[ncols_dst] = { 0.0f };
     if constexpr (has_fusion) {
+        const uint32_t channel_bias = ids ? channel_x : channel_dst;
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
             // 1. Hide latency by prefetching bias and gate here
@@ -222,6 +236,9 @@ static __global__ void mul_mat_vec_q(
     float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    if constexpr (is_multi_token_id) {
+        y += token_idx*stride_col_y;
+    }
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -275,6 +292,10 @@ static __global__ void mul_mat_vec_q(
 
     dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
 
+    if constexpr (is_multi_token_id) {
+        dst += token_idx*stride_col_dst;
+    }
+
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
@@ -335,40 +356,41 @@ static __global__ void mul_mat_vec_q(
 }
 
 static std::pair<dim3, dim3> calc_launch_params(
-        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
         const int warp_size, const mmvq_parameter_table_id table_id) {
     const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
-    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
     const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst>
+template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
         const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
+        const uint32_t ids_stride, cudaStream_t stream) {
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
         }
     }
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }
 
 template <ggml_type type>
@@ -379,7 +401,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
     GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
@@ -393,8 +415,19 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
+        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+             dims.first, dims.second, 0, ids_stride, stream);
+        return;
+    }
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
@@ -402,7 +435,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
@@ -410,7 +443,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
@@ -418,7 +451,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
@@ -426,7 +459,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
@@ -434,7 +467,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
@@ -442,7 +475,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
@@ -450,7 +483,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
@@ -458,7 +491,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -474,127 +507,127 @@ static void mul_mat_vec_q_switch_type(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
     switch (type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -622,7 +655,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVQ_MAX_BATCH_SIZE);
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -693,11 +726,13 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
+
     mul_mat_vec_q_switch_type(
         src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,               stream);
+        ne03,              ne3,           s03, s13,              s3,               ids_stride, stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -726,7 +761,7 @@ void ggml_cuda_op_mul_mat_vec_q(
     ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
         src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
-        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }

ggml/src/ggml-cuda/pad.cu

@@ -7,7 +7,7 @@ __device__ __forceinline__ int64_t wrap_around(int64_t coord, int64_t size) {
     return (coord + size) % size;
 }
 
-static __global__ void pad_f32(const float * src, float * dst,
+static __global__ void pad_f32(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
                                const int lp0, const int rp0, const int lp1, const int rp1,
                                const int lp2, const int rp2, const int lp3, const int rp3,
                                const int ne0, const int ne1, const int ne2, const int ne3,
@@ -34,11 +34,8 @@ static __global__ void pad_f32(const float * src, float * dst,
             const int64_t i01  = i1 - lp1;
             const int64_t i02  = i2 - lp2;
             const int64_t i03  = i3 - lp3;
-            const int64_t ne02 = ne2 - lp2 - rp2;
-            const int64_t ne01 = ne1 - lp1 - rp1;
-            const int64_t ne00 = ne0 - lp0 - rp0;
 
-            const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
+            const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
 
             dst[dst_idx] = src[src_idx];
         } else {
@@ -57,21 +54,21 @@ static __global__ void pad_f32(const float * src, float * dst,
         const int64_t i02 = wrap_around(i2 - lp2, ne02);
         const int64_t i03 = wrap_around(i3 - lp3, ne03);
 
-        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
+        const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
 
         dst[dst_idx] = src[src_idx];
     }
 }
 
 
-static void pad_f32_cuda(const float * src, float * dst,
+static void pad_f32_cuda(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
     const int lp0, const int rp0, const int lp1, const int rp1,
     const int lp2, const int rp2, const int lp3, const int rp3,
     const int ne0, const int ne1, const int ne2, const int ne3,
     const bool circular, cudaStream_t stream) {
     int  num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
     dim3 gridDim(num_blocks, ne1, ne2 * ne3);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst,
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, s00, s01, s02, s03, dst,
                                                          lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
                                                          ne0, ne1, ne2, ne3, circular);
 }
@@ -82,9 +79,10 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     float *             dst_d  = (float *) dst->data;
     cudaStream_t        stream = ctx.stream();
 
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
 
     const int32_t lp0      = ((const int32_t *) (dst->op_params))[0];
     const int32_t rp0      = ((const int32_t *) (dst->op_params))[1];
@@ -96,7 +94,12 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int32_t rp3      = ((const int32_t *) (dst->op_params))[7];
     const int32_t circular = ((const int32_t *) (dst->op_params))[8];
 
-    pad_f32_cuda(src0_d, dst_d,
+    const size_t s00 = nb00 / ggml_type_size(src0->type);
+    const size_t s01 = nb01 / ggml_type_size(src0->type);
+    const size_t s02 = nb02 / ggml_type_size(src0->type);
+    const size_t s03 = nb03 / ggml_type_size(src0->type);
+
+    pad_f32_cuda(src0_d, s00, s01, s02, s03, dst_d,
                  lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
                  dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
                  (bool) circular, stream);

ggml/src/ggml-cuda/rope.cu

@@ -43,10 +43,15 @@ static __device__ void rope_yarn(
 template <bool forward, bool has_ff, typename T, typename D>
 static __global__ void rope_norm(const T *            x,
                                  D *                  dst,
-                                 const int            ne0,
-                                 const int            ne1,
+                                 const int            ne00,
+                                 const int            ne01,
+                                 const int            ne02,
+                                 const int            s01,
+                                 const int            s02,
+                                 const int            s03,
                                  const int            s1,
                                  const int            s2,
+                                 const int            s3,
                                  const int            n_dims,
                                  const int32_t *      pos,
                                  const float          freq_scale,
@@ -59,23 +64,23 @@ static __global__ void rope_norm(const T *            x,
                                  const int            set_rows_stride) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
         return;
     }
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
-
-    int       idst = row_dst * ne0 + i0;
-    const int ix   = channel_x*s2 + row_x*s1 + i0;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
 
+    int       idst = i0 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 + i1 * s01 + i2 * s02 + i3 * s03;
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
     if (set_rows_stride != 0) {
-        idst = row_x * ne0 + i0;
-        idst += row_indices[channel_x] * set_rows_stride;
+        idst = i1 * s1 + i0;
+        idst += row_indices[i2] * set_rows_stride;
     }
 
     const auto & store_coaelsced = [&](float x0, float x1) {
@@ -92,7 +97,7 @@ static __global__ void rope_norm(const T *            x,
         return;
     }
 
-    const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
 
@@ -110,10 +115,15 @@ static __global__ void rope_norm(const T *            x,
 template <bool forward, bool has_ff, typename T, typename D>
 static __global__ void rope_neox(const T *            x,
                                  D *                  dst,
-                                 const int            ne0,
-                                 const int            ne1,
+                                 const int            ne00,
+                                 const int            ne01,
+                                 const int            ne02,
+                                 const int            s01,
+                                 const int            s02,
+                                 const int            s03,
                                  const int            s1,
                                  const int            s2,
+                                 const int            s3,
                                  const int            n_dims,
                                  const int32_t *      pos,
                                  const float          freq_scale,
@@ -126,23 +136,24 @@ static __global__ void rope_neox(const T *            x,
                                  const int            set_rows_stride) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
         return;
     }
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
 
-    int       idst = row_dst * ne0 + i0 / 2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
     if (set_rows_stride != 0) {
-        idst = row_x * ne0 + i0 / 2;
-        idst += row_indices[channel_x] * set_rows_stride;
+        idst = i1 * s1 + i0 / 2;
+        idst += row_indices[i2] * set_rows_stride;
     }
 
     if (i0 >= n_dims) {
@@ -152,7 +163,7 @@ static __global__ void rope_neox(const T *            x,
         return;
     }
 
-    const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
 
@@ -168,24 +179,42 @@ static __global__ void rope_neox(const T *            x,
     dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }
 
-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_multi(
-        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
-        const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections, const bool is_imrope) {
-    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+template <bool forward, bool has_ff, typename T>
+static __global__ void rope_multi(const T *            x,
+                                  T *                  dst,
+                                  const int            ne00,
+                                  const int            ne01,
+                                  const int            ne02,
+                                  const int            s01,
+                                  const int            s02,
+                                  const int            s03,
+                                  const int            s1,
+                                  const int            s2,
+                                  const int            s3,
+                                  const int            n_dims,
+                                  const int32_t *      pos,
+                                  const float          freq_scale,
+                                  const float          ext_factor,
+                                  const float          attn_factor,
+                                  const rope_corr_dims corr_dims,
+                                  const float          theta_scale,
+                                  const float *        freq_factors,
+                                  const mrope_sections sections,
+                                  const bool           is_imrope) {
+    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);
 
-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
         return;
     }
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
 
-    const int idst = row_dst*ne0 + i0/2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
     if (i0 >= n_dims) {
         dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
@@ -200,27 +229,24 @@ static __global__ void rope_multi(
 
     float theta_base = 0.0;
     if (is_imrope) {
-        if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
-            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
-            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
-        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
-            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {         // h
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {  // w
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {  // t
+            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
         } else {
-            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
         }
     } else {
         if (sector < sections.v[0]) {
-            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
-        }
-        else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-        }
-        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
-        }
-        else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
         }
     }
 
@@ -238,37 +264,53 @@ static __global__ void rope_multi(
     dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }
 
-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_vision(
-        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims,
-        const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
-        const float theta_scale, const float * freq_factors, const mrope_sections sections) {
+template <bool forward, bool has_ff, typename T>
+static __global__ void rope_vision(const T *            x,
+                                   T *                  dst,
+                                   const int            ne00,
+                                   const int            ne01,
+                                   const int            ne02,
+                                   const int            s01,
+                                   const int            s02,
+                                   const int            s03,
+                                   const int            s1,
+                                   const int            s2,
+                                   const int            s3,
+                                   const int            n_dims,
+                                   const int32_t *      pos,
+                                   const float          freq_scale,
+                                   const float          ext_factor,
+                                   const float          attn_factor,
+                                   const rope_corr_dims corr_dims,
+                                   const float          theta_scale,
+                                   const float *        freq_factors,
+                                   const mrope_sections sections) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
         return;
     }
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;
 
-    const int idst = row_dst*ne0 + i0/2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
     const int sect_dims = sections.v[0] + sections.v[1];
-    const int sec_w = sections.v[1] + sections.v[0];
-    const int sector = (i0 / 2) % sect_dims;
+    const int sec_w     = sections.v[1] + sections.v[0];
+    const int sector    = (i0 / 2) % sect_dims;
 
     float theta_base = 0.0;
     if (sector < sections.v[0]) {
         const int p = sector;
-        theta_base = pos[channel_x]*powf(theta_scale, p);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base  = pos[i2] * powf(theta_scale, p);
+    } else if (sector >= sections.v[0] && sector < sec_w) {
         const int p = sector - sections.v[0];
-        theta_base = pos[channel_x + ne2]*powf(theta_scale, p);
+        theta_base  = pos[i2 + ne02] * powf(theta_scale, p);
     }
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -288,10 +330,15 @@ static __global__ void rope_vision(
 template <bool forward, typename T, typename D>
 static void rope_norm_cuda(const T *            x,
                            D *                  dst,
-                           const int            ne0,
-                           const int            ne1,
+                           const int            ne00,
+                           const int            ne01,
+                           const int            ne02,
+                           const int            s01,
+                           const int            s02,
+                           const int            s03,
                            const int            s1,
                            const int            s2,
+                           const int            s3,
                            const int            n_dims,
                            const int            nr,
                            const int32_t *      pos,
@@ -304,31 +351,36 @@ static void rope_norm_cuda(const T *            x,
                            const int64_t *      row_indices,
                            const int            set_rows_stride,
                            cudaStream_t         stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+    GGML_ASSERT(ne00 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nr, n_blocks_x, 1);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
     if (freq_factors == nullptr) {
         rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     } else {
         rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     }
 }
 
 template <bool forward, typename T, typename D>
 static void rope_neox_cuda(const T *            x,
                            D *                  dst,
-                           const int            ne0,
-                           const int            ne1,
+                           const int            ne00,
+                           const int            ne01,
+                           const int            ne02,
+                           const int            s01,
+                           const int            s02,
+                           const int            s03,
                            const int            s1,
                            const int            s2,
+                           const int            s3,
                            const int            n_dims,
                            const int            nr,
                            const int32_t *      pos,
@@ -341,55 +393,92 @@ static void rope_neox_cuda(const T *            x,
                            const int64_t *      row_indices,
                            const int            set_rows_stride,
                            cudaStream_t         stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+    GGML_ASSERT(ne00 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nr, n_blocks_x, 1);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
     if (freq_factors == nullptr) {
         rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     } else {
         rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
     }
 }
 
-template<bool forward, typename T>
-static void rope_multi_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, const bool is_imrope, cudaStream_t stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+template <bool forward, typename T>
+static void rope_multi_cuda(const T *            x,
+                            T *                  dst,
+                            const int            ne00,
+                            const int            ne01,
+                            const int            ne02,
+                            const int            s01,
+                            const int            s02,
+                            const int            s03,
+                            const int            s1,
+                            const int            s2,
+                            const int            s3,
+                            const int            n_dims,
+                            const int            nr,
+                            const int32_t *      pos,
+                            const float          freq_scale,
+                            const float          freq_base,
+                            const float          ext_factor,
+                            const float          attn_factor,
+                            const rope_corr_dims corr_dims,
+                            const float *        freq_factors,
+                            const mrope_sections sections,
+                            const bool           is_imrope,
+                            cudaStream_t         stream) {
+    GGML_ASSERT(ne00 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nr, n_blocks_x, 1);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
     if (freq_factors == nullptr) {
         rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     } else {
         rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     }
 }
 
-template<bool forward, typename T>
-static void rope_vision_cuda(
-        const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
-        const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+template <bool forward, typename T>
+static void rope_vision_cuda(const T *            x,
+                             T *                  dst,
+                             const int            ne00,
+                             const int            ne01,
+                             const int            ne02,
+                             const int            s01,
+                             const int            s02,
+                             const int            s03,
+                             const int            s1,
+                             const int            s2,
+                             const int            s3,
+                             const int            n_dims,
+                             const int            nr,
+                             const int32_t *      pos,
+                             const float          freq_scale,
+                             const float          freq_base,
+                             const float          ext_factor,
+                             const float          attn_factor,
+                             const rope_corr_dims corr_dims,
+                             const float *        freq_factors,
+                             const mrope_sections sections,
+                             cudaStream_t         stream) {
+    GGML_ASSERT(ne00 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
     const dim3 block_nums(nr, n_blocks_x, 1);
     // break down (head_dim, heads, seq) into (CUDA_ROPE_BLOCK_SIZE, x, heads * seq)
     // where x ~= ceil(head_dim / CUDA_ROPE_BLOCK_SIZE);
@@ -398,11 +487,11 @@ static void rope_vision_cuda(
 
     if (freq_factors == nullptr) {
         rope_vision<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections);
     } else {
         rope_vision<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
             attn_factor, corr_dims, theta_scale, freq_factors, sections);
     }
 }
@@ -445,6 +534,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
 
     const size_t s01 = src0->nb[1] / ggml_type_size(src0->type);
     const size_t s02 = src0->nb[2] / ggml_type_size(src0->type);
+    const size_t s03 = src0->nb[3] / ggml_type_size(src0->type);
+
+    const size_t s1 = dst->nb[1] / ggml_type_size(dst->type);
+    const size_t s2 = dst->nb[2] / ggml_type_size(dst->type);
+    const size_t s3 = dst->nb[3] / ggml_type_size(dst->type);
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -495,57 +589,63 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
     // compute
     if (is_neox) {
         if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
-            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                  freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                  set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                 freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                 set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
-                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                set_rows_stride, stream);
         } else {
             GGML_ABORT("fatal error");
         }
     } else if (is_mrope && !is_vision) {
         if (src0->type == GGML_TYPE_F32) {
-            rope_multi_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
+            rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, stream);
         } else if (src0->type == GGML_TYPE_F16) {
-            rope_multi_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
+            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, stream);
         } else {
             GGML_ABORT("fatal error");
         }
     } else if (is_vision) {
         if (src0->type == GGML_TYPE_F32) {
-            rope_vision_cuda<forward>(
-                (const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+            rope_vision_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                      s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                      corr_dims, freq_factors, sections, stream);
         } else if (src0->type == GGML_TYPE_F16) {
-            rope_vision_cuda<forward>(
-                (const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+            rope_vision_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                      s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                      corr_dims, freq_factors, sections, stream);
         } else {
             GGML_ABORT("fatal error");
         }
     } else {
         if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
-            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                  freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                  set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                 freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                 set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
-                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                set_rows_stride, stream);
         } else {
             GGML_ABORT("fatal error");
         }

ggml/src/ggml-cuda/vecdotq.cuh

@@ -94,6 +94,15 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
 #endif
 }
 
+static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
+    // v is a 7 bit int, with the 8th sign being encodable as popcnt
+    // with xor we can "correct" the bit instead of having to mask
+    const uint32_t p = __popc(v) & 1;
+    const uint32_t s = v ^ p << 7;
+    // broadcast over uint to allow for 0x08040201 / 0x80402010 as selectors
+    return s * 0x01010101;
+}
+
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
@@ -905,22 +914,22 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     int sumi = 0;
 #pragma unroll
     for (int k0 = 0; k0 < 8; k0 += 2) {
-        const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]);
-        const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F];
+        const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[k0/2]];
+        const uint32_t signs = unpack_ksigns(aux32 >> (7 * k0 / 2));
 
-        const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
-        const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0);
         sumi = ggml_cuda_dp4a(grid0, u0, sumi);
 
-        const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
-        const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1);
         sumi = ggml_cuda_dp4a(grid1, u1, sumi);
     }
 
-    const int ls = aux32 >> 28;
-    sumi = (ls*sumi + sumi/2)/4;
+    const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
+    sumi = sumi * ls / 8;           // (sumi * scale + sumi / 2) / 4
     const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
     return d * sumi;
 }
@@ -942,13 +951,15 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     int sumi1 = 0;
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
-        const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF));
-        const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l0/2] >> 9));
-
-        const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+        const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l0/2] & 0x1FF];
+        const uint32_t signs = unpack_ksigns(q2[l0/2] >> 9);
 
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         if (l0 < 4) {
@@ -1028,13 +1039,16 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #pragma unroll
     for (int l0 = 0; l0 < 8; l0 += 2) {
         const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]);
+        const uint32_t signs = unpack_ksigns(aux32 >> (7*l0/2));
 
-        const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F));
-
-        const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
-        const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+        const int signs0 = __vcmpne4(signs & 0x08040201, 0);
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
 
         const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+
+        const int signs1 = __vcmpne4(signs & 0x80402010, 0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
         const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
 
         sumi = ggml_cuda_dp4a(grid_l, u0, sumi);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1935,11 +1935,6 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
         return false;
     }
 
-    // TODO: add support for non-contigiuos tensors
-    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
-        return false;
-    }
-
     return true;
 }
 
@@ -1991,6 +1986,25 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
     return true;
 }
 
+static bool ggml_hexagon_supported_sum_rows(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (!hex_supported_src0_type(src0->type)) {
+        return false;
+    }
+    if (!hex_supported_dst_type(dst->type)) {
+        return false;
+    }
+
+    // TODO: add support for non-contigiuos tensors
+    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session * sess,
                                                const struct ggml_tensor *          op) {
     const struct ggml_tensor * src0 = op->src[0];
@@ -2111,6 +2125,26 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
     return true;
 }
 
+static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0]; // values
+    const struct ggml_tensor * dst  = op;         // indices
+
+    if (src0->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (dst->type != GGML_TYPE_I32) {
+        return false;
+    }
+
+    if (src0->ne[0] > (16*1024)) {
+        // reject tensors with huge rows for now
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const int32_t * op_params = &op->op_params[0];
 
@@ -2278,6 +2312,9 @@ static inline size_t init_binary_req(htp_general_req * req, dspqueue_buffer * bu
         case GGML_OP_SUB:
             req->op = HTP_OP_SUB;
             break;
+        case GGML_OP_DIV:
+            req->op = HTP_OP_DIV;
+            break;
         default:
             GGML_ABORT("ggml-hex: binary : unsupported op: %d\n", t->op);
             break;
@@ -2316,6 +2353,17 @@ static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer *
     return n_bufs;
 }
 
+static inline size_t init_argsort_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    req->op = HTP_OP_ARGSORT;
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 template <bool _is_src0_constant>
 static inline size_t init_binary_id_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
     switch (t->op) {
@@ -2370,6 +2418,16 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
             supported = true;
             break;
 
+        case GGML_OP_SQR:
+            req->op   = HTP_OP_SQR;
+            supported = true;
+            break;
+
+        case GGML_OP_SQRT:
+            req->op   = HTP_OP_SQRT;
+            supported = true;
+            break;
+
         case GGML_OP_UNARY:
             if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
                 req->op   = HTP_OP_UNARY_SILU;
@@ -2387,6 +2445,9 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
             } else if (ggml_get_glu_op(t) == GGML_GLU_OP_SWIGLU_OAI) {
                 req->op   = HTP_OP_GLU_SWIGLU_OAI;
                 supported = true;
+            } else if (ggml_get_glu_op(t) == GGML_GLU_OP_GEGLU) {
+                req->op   = HTP_OP_GLU_GEGLU;
+                supported = true;
             }
             break;
 
@@ -2411,6 +2472,17 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
     return n_bufs;
 }
 
+static inline size_t init_sum_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+    req->op = HTP_OP_SUM_ROWS;
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 static inline size_t init_rope_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
     memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
     req->op = HTP_OP_ROPE;
@@ -2519,6 +2591,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
             case GGML_OP_MUL:
             case GGML_OP_ADD:
             case GGML_OP_SUB:
+            case GGML_OP_DIV:
                 ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
                 break;
             case GGML_OP_ADD_ID:
@@ -2528,6 +2601,13 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
             case GGML_OP_SCALE:
                 ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                 break;
+            case GGML_OP_SQR:
+            case GGML_OP_SQRT:
+                ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
+                break;
+            case GGML_OP_SUM_ROWS:
+                ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
+                break;
             case GGML_OP_UNARY:
                 if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
                         (ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
@@ -2536,7 +2616,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 break;
             case GGML_OP_GLU:
                 if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
-                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI)) {
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
                     ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                 }
                 break;
@@ -2564,6 +2645,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
                 break;
 
+            case GGML_OP_ARGSORT:
+                ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
+                break;
+
             default:
                 GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
         }
@@ -2916,6 +3001,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_MUL:
         case GGML_OP_ADD:
         case GGML_OP_SUB:
+        case GGML_OP_DIV:
             supp = ggml_hexagon_supported_binary(sess, op);
             break;
 
@@ -2928,6 +3014,15 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_unary(sess, op);
             break;
 
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            supp = ggml_hexagon_supported_unary(sess, op);
+            break;
+
+        case GGML_OP_SUM_ROWS:
+            supp = ggml_hexagon_supported_sum_rows(sess, op);
+            break;
+
         case GGML_OP_SOFT_MAX:
             supp = ggml_hexagon_supported_softmax(sess, op);
             break;
@@ -2943,7 +3038,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_GLU:
             {
                 const auto glu_op = ggml_get_glu_op(op);
-                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI)) {
+                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
                     supp = ggml_hexagon_supported_activations(sess, op);
                 }
                 break;
@@ -2968,6 +3063,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_cpy(sess, op);
             break;
 
+        case GGML_OP_ARGSORT:
+            supp = ggml_hexagon_supported_argsort(sess, op);
+            break;
+
         default:
             break;
     }

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -6,6 +6,7 @@ include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
 include_directories(
     ${HEXAGON_SDK_ROOT}/incs
     ${HEXAGON_SDK_ROOT}/incs/stddef
+    ${CMAKE_CURRENT_SOURCE_DIR}/../../../include
     ${CMAKE_CURRENT_SOURCE_DIR}/../..
     ${CMAKE_CURRENT_SOURCE_DIR}/..
     ${CMAKE_CURRENT_SOURCE_DIR}
@@ -21,6 +22,7 @@ add_library(${HTP_LIB} SHARED
     matmul-ops.c
     binary-ops.c
     unary-ops.c
+    sum-rows-ops.c
     softmax-ops.c
     act-ops.c
     rope-ops.c
@@ -28,6 +30,7 @@ add_library(${HTP_LIB} SHARED
     set-rows-ops.c
     get-rows-ops.c
     cpy-ops.c
+    argsort-ops.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

ggml/src/ggml-hexagon/htp/act-ops.c

@@ -410,7 +410,7 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
             // gelu = x * sigmoid(1.702 * x) // current implementation
             hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0);
             hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
         }
 
         dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -516,7 +516,7 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
 
             // silu = x * sigmoid(x)
             hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
         }
 
         dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -541,6 +541,143 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
          ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }
 
+static const float GELU_COEF_A     = 0.044715f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
+                                       const struct htp_tensor * src1,
+                                       struct htp_tensor *       dst,
+                                       const int32_t *           op_params,
+                                       struct htp_spad *         src0_spad,
+                                       struct htp_spad *         src1_spad,
+                                       struct htp_spad *         dst_spad,
+                                       uint32_t                  nth,
+                                       uint32_t                  ith,
+                                       uint32_t                  src0_nrows_per_thread,
+                                       dma_queue *               dma_queue) {
+    htp_act_preamble3;
+
+    size_t src0_row_size = nb01;
+    size_t src1_row_size = nb11;
+    size_t dst_row_size  = nb1;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    const bool src1_valid = src1->ne[0];
+    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
+    if (!src1_valid) {
+        const int32_t swapped = op_params[1];
+        data_src1             = data_src0;
+        src1_row_size         = src0_row_size;
+
+        const size_t nc_in_bytes = nc * SIZEOF_FP32;
+        data_src0 += swapped ? nc_in_bytes : 0;
+        data_src1 += swapped ? 0 : nc_in_bytes;
+    }
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
+    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
+    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+
+    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
+    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
+    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
+    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
+
+    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    if (BLOCK == 0) {
+        FARF(ERROR,
+             "geglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+             src0_spad->size_per_thread, src0_row_size_aligned);
+        return;
+    }
+
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);
+
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)),
+            src1_row_size_aligned, src1_row_size, block_size);
+    }
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
+        float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;
+
+        for (uint32_t ib = 0; ib < block_size; ib++) {
+            const uint8_t * src0_spad_ptr = (const uint8_t *)(src0_spad + ib * (src0_row_size_aligned / sizeof(float)));
+            const uint8_t * src1_spad_ptr = (const uint8_t *)(src1_spad + ib * (src1_row_size_aligned / sizeof(float)));
+            uint8_t *       dst_spad_ptr  = (uint8_t *)(dst_spad + ib * (dst_row_size_aligned / sizeof(float)));
+
+            // geglu tanh implementation
+            // geglu(x, g) = gelu(x) * g
+            // gelu(x) = 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)))
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, src0_spad_ptr, nc);                       // res = x*x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, GELU_COEF_A, nc);   // res = res * GELU_COEF_A
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 1.0f, nc);          // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, SQRT_2_OVER_PI, nc); // res = result * SQRT_2_OVER_PI
+            hvx_tanh_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, nc);         // res = tanh(res)
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, 1.0f, nc);           // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 0.5f, nc);          // res = res + 0.5f
+            hvx_mul_f32_aaa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, src1_spad_ptr, nc);       // res = res * g
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size,
+                                   dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)),
+                                       src0_row_size_aligned, src0_row_size, pref_block_size);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)),
+                                       src1_row_size_aligned, src1_row_size, pref_block_size);
+        }
+    }
+
+    dma_queue_flush(dma_queue);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "geglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
 static void unary_silu_f32(unsigned int n, unsigned int i, void * data) {
     struct htp_ops_context * octx = (struct htp_ops_context *) data;
     unary_silu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
@@ -559,6 +696,12 @@ static void glu_swiglu_oai_f32(unsigned int n, unsigned int i, void * data) {
                                    &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
 }
 
+static void glu_geglu_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    glu_geglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
+                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
+}
+
 static int execute_op_activations_f32(struct htp_ops_context * octx) {
     int err = HTP_STATUS_OK;
 
@@ -593,6 +736,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
             act_op_func = unary_gelu_f32;
             op_type     = "gelu-f32";
             break;
+
+        case HTP_OP_GLU_GEGLU:
+            act_op_func = glu_geglu_f32;
+            op_type     = "geglu-f32";
+            break;
         default:
             FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
             return HTP_STATUS_NO_SUPPORT;

ggml/src/ggml-hexagon/htp/argsort-ops.c

@@ -0,0 +1,281 @@
+#include <string.h>
+#include <stdlib.h>
+#include <math.h>
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "ggml.h"
+
+#include "hvx-utils.h"
+#include "hex-dma.h"
+
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+#ifndef MIN
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+struct htp_argsort_context {
+    struct htp_ops_context * octx;
+    uint32_t                 nrows_per_thread;
+};
+
+static inline bool all_greater_f32(HVX_Vector x, HVX_Vector y)
+{
+    const HVX_Vector one  = Q6_V_vsplat_R(1);
+    const HVX_Vector zero = Q6_V_vzero();
+
+    HVX_VectorPred pred = Q6_Q_vcmp_gt_VsfVsf(x, y);
+    HVX_Vector matches = Q6_V_vmux_QVV(pred, one, zero);
+    HVX_Vector sum = hvx_vec_reduce_sum_i32(matches);
+    return hvx_vec_get_i32(sum) == 32;
+}
+
+// Sorts values and mirrors swaps to indices.
+static void quicksort_values_indices_asc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+    while (i <= j) {
+        // Vectorized scan for i
+        while (i <= j) {
+            // Check if we have at least one full vector
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    // If all elements are < pivot, we can skip this whole block
+                    i += 32;
+                    continue;
+                }
+            }
+
+            // Scalar fallback / cleanup
+            if (values[i] < pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j
+        while (i <= j) {
+            if (j - 32 >= i) {
+                // Load 32 elements ending at j.
+                // Since we want `values[j] > pivot`, let's load from j-31 to j.
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] > pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_asc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_asc(values, indices, i, right);
+}
+
+static void quicksort_values_indices_desc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+
+    while (i <= j) {
+        // Vectorized scan for i (values[i] > pivot)
+        while (i <= j) {
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    i += 32;
+                    continue;
+                }
+            }
+
+            if (values[i] > pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j (values[j] < pivot)
+        while (i <= j) {
+            if (j - 32 >= i) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] < pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_desc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_desc(values, indices, i, right);
+}
+
+static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_argsort_context * actx = (struct htp_argsort_context *)data;
+    struct htp_ops_context * octx = actx->octx;
+
+    // Unpack context
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * dst = &octx->dst;
+
+    // Scratchpad memory
+    uint8_t * spad = octx->src0_spad.data + octx->src0_spad.size_per_thread * i;
+
+    // Dimensions
+    uint32_t ne00 = src0->ne[0];
+    uint32_t ne01 = src0->ne[1];
+    uint32_t ne02 = src0->ne[2];
+    uint32_t ne03 = src0->ne[3];
+
+    uint32_t nb01 = src0->nb[1];
+    //uint32_t nb02 = src0->nb[2];
+    //uint32_t nb03 = src0->nb[3];
+
+    uint32_t nb1 = dst->nb[1];
+    //uint32_t nb2 = dst->nb[2];
+    //uint32_t nb3 = dst->nb[3];
+
+    // Sort order
+    enum ggml_sort_order order = (enum ggml_sort_order) octx->op_params[0];
+
+    // Rows to process
+    uint32_t total_rows = ne01 * ne02 * ne03;
+    uint32_t rows_per_thread = actx->nrows_per_thread;
+    uint32_t start_row = rows_per_thread * i;
+    uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);
+
+    // Scratchpad layout:
+    // We need space for one row of float data (values) and one row of int32 indices.
+    // values: ne00 * sizeof(float)
+    // indices: ne00 * sizeof(int32_t)
+    // Padded to 128 bytes.
+
+    size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
+    float * values_buf = (float *) spad;
+    int32_t * indices_buf = (int32_t *) (spad + values_size);
+
+    for (uint32_t r = start_row; r < end_row; r++) {
+        uint32_t src_offset = r * nb01;
+        uint32_t dst_offset = r * nb1;
+
+        uint8_t * src_ptr = (uint8_t *) src0->data + src_offset;
+        uint8_t * dst_ptr = (uint8_t *) dst->data  + dst_offset;
+
+        hex_l2fetch(src_ptr, ne00 * sizeof(float), ne00 * sizeof(float), 1);
+        hvx_copy_f32_au((uint8_t*)values_buf, src_ptr, ne00);
+
+        // Initialize indices
+        for (uint32_t j = 0; j < ne00; j++) {
+            indices_buf[j] = j;
+        }
+
+        // Sort values and mirror swaps to indices
+        if (order == GGML_SORT_ORDER_ASC) {
+            quicksort_values_indices_asc(values_buf, indices_buf, 0, ne00 - 1);
+        } else {
+            quicksort_values_indices_desc(values_buf, indices_buf, 0, ne00 - 1);
+        }
+
+        // Copy indices back to DDR
+        hvx_copy_f32_ua(dst_ptr, (const uint8_t *) indices_buf, ne00);
+    }
+}
+
+int op_argsort(struct htp_ops_context * octx) {
+    // Check supported types
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    // Allocate scratchpad
+    // We need 1 row of float + 1 row of int32 per thread.
+    uint32_t ne00 = octx->src0.ne[0];
+    size_t values_size  = hex_round_up(ne00 * sizeof(float), 128);
+    size_t indices_size = hex_round_up(ne00 * sizeof(int32_t), 128);
+    size_t spad_per_thread = values_size + indices_size;
+
+    // Make sure we round up to 256 for alignment requirements
+    spad_per_thread = hex_round_up(spad_per_thread, 256);
+
+    size_t total_spad_size = spad_per_thread * octx->n_threads;
+
+    if (octx->ctx->vtcm_size < total_spad_size) {
+        FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src0_spad.size = total_spad_size;
+    octx->src0_spad.size_per_thread = spad_per_thread;
+
+    FARF(HIGH, "argsort: %ux%ux%ux%u -> %ux%ux%ux%u (0x%x, 0x%x)",
+         octx->src0.ne[0], octx->src0.ne[1], octx->src0.ne[2], octx->src0.ne[3],
+         octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
+         octx->src0.data, octx->dst.data);
+
+    uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    uint32_t n_jobs = MIN(total_rows, octx->n_threads);
+
+    struct htp_argsort_context actx;
+    actx.octx = octx;
+    actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
+
+    // Run jobs
+    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
+
+    return HTP_STATUS_OK;
+}

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -17,121 +17,6 @@
 #include "htp-msg.h"
 #include "htp-ops.h"
 
-static inline HVX_Vector hvx_load_f32_to_f16(const HVX_Vector * restrict src, const HVX_Vector zero) {
-    HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(src[0], zero);  // 32 elements
-    HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(src[1], zero);  // 32 elements
-    return Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
-}
-
-// Dot product of FP32 and FP16 vectors, accumulating to float
-static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict y, const void * restrict x, unsigned int n, float s) {
-    const HVX_Vector * restrict vy = (const HVX_Vector * restrict) y; // fp32
-    const HVX_Vector * restrict vx = (const HVX_Vector * restrict) x; // fp16
-
-    uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16; // leftover elements
-
-    const HVX_Vector zero = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum = Q6_V_vsplat_R(0);
-
-    uint32_t i = 0;
-
-    #pragma unroll(4)
-    for (i = 0; i < nvec; i++) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x_hf  = vx[i];
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
-    }
-
-    if (nloe) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x_hf  = vx[i];
-
-        // Zero-out unused elements
-        // Note that we need to clear both x and y because they may contain NANs
-        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        x_hf = Q6_V_vand_QV(bmask, x_hf);
-        y_hf = Q6_V_vand_QV(bmask, y_hf);
-
-        HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
-
-        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
-    }
-
-    rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
-    hvx_vec_store_u(r, 4, Q6_Vsf_equals_Vqf32(rsum));
-}
-
-// Dot product of FP32 and FP16 vectors, accumulating to float
-static inline void hvx_dot_f32_f16_aa_rx2(float * restrict r,
-                                          const void * restrict y,
-                                          const void * restrict x0,
-                                          const void * restrict x1,
-                                          unsigned int n,
-                                          float        s) {
-    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp32
-    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x0;  // fp16
-    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
-
-    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
-
-    const HVX_Vector zero  = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
-
-    uint32_t i = 0;
-
-    #pragma unroll(2)
-    for (i = 0; i < nvec; i++) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-        // Load x (fp16)
-        HVX_Vector x0_hf = vx0[i];
-        HVX_Vector x1_hf = vx1[i];
-
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
-    }
-
-    if (nloe) {
-        // Load y (fp32) and convert into fp16
-        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
-
-        // Load x (fp16)
-        HVX_Vector x0_hf = vx0[i];
-        HVX_Vector x1_hf = vx1[i];
-
-        // Zero-out unused elements
-        // Note that we need to clear both x and y because they may contain NANs
-        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        x0_hf                = Q6_V_vand_QV(bmask, x0_hf);
-        x1_hf                = Q6_V_vand_QV(bmask, x1_hf);
-        y_hf                 = Q6_V_vand_QV(bmask, y_hf);
-
-        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
-        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
-
-        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
-        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
-    }
-
-    HVX_Vector rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32x2(rsum0, rsum1));
-    hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
-}
-
 // Dot product of two F16 vectors, accumulating to float
 static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict x, const void * restrict y, unsigned int n, float s) {
     const HVX_Vector * restrict vx = (const HVX_Vector * restrict) x; // fp16
@@ -140,8 +25,7 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     uint32_t nvec = n / VLEN_FP16; // num full fp16 hvx vectors
     uint32_t nloe = n % VLEN_FP16; // leftover elements
 
-    const HVX_Vector zero = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum = Q6_V_vsplat_R(0);
+    HVX_Vector rsum = Q6_V_vsplat_R(0);
 
     uint32_t i = 0;
 
@@ -156,11 +40,10 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
     }
 
     if (nloe) {
-        HVX_Vector y_hf = vy[i];
-
         // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        HVX_Vector      x_hf = Q6_V_vand_QV(bmask, vx[i]);
+        HVX_Vector y_hf = Q6_V_vand_QV(bmask, vy[i]);
+        HVX_Vector x_hf = Q6_V_vand_QV(bmask, vx[i]);
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
@@ -181,12 +64,11 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
     const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp16
 
-    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
-    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
+    uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;  // leftover elements
 
-    const HVX_Vector zero  = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
-    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
+    HVX_Vector rsum0 = Q6_V_vsplat_R(0);
+    HVX_Vector rsum1 = Q6_V_vsplat_R(0);
 
     uint32_t i = 0;
 
@@ -204,12 +86,11 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     }
 
     if (nloe) {
-        HVX_Vector y_hf = vy[i];
-
         // Load x (fp16) and zero-out unused elements
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
-        HVX_Vector     x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
-        HVX_Vector     x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+        HVX_Vector x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
+        HVX_Vector x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+        HVX_Vector y_hf  = Q6_V_vand_QV(bmask, vy[i]);
 
         HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
         HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
@@ -222,7 +103,7 @@ static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
     hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
 }
 
-// MAD: y (F32) += x (F16) * s (float)
+// MAD: y (F32) += x (F16) * s (F32)
 static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, int n, float s) {
     const HVX_Vector * restrict ptr_x = (const HVX_Vector *) x;
     HVX_Vector * restrict ptr_y = (HVX_Vector *) y;
@@ -259,15 +140,125 @@ static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict
     }
 }
 
+// MAD: y (F32) += x0 (F16) * s0 (F32) + x1 (F16) * s1 (F32)
+static inline void hvx_mad_f32_f16_aa_rx2(float * restrict y,
+                                          const void * restrict x0,
+                                          const void * restrict x1,
+                                          float s0,
+                                          float s1,
+                                          int   n) {
+    const HVX_Vector * restrict ptr_x0 = (const HVX_Vector *) x0;
+    const HVX_Vector * restrict ptr_x1 = (const HVX_Vector *) x1;
+    HVX_Vector * restrict ptr_y        = (HVX_Vector *) y;
+
+    uint32_t nvec = n / VLEN_FP16;  // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;  // leftover elements
+
+    HVX_Vector S0 = hvx_vec_splat_f16(s0);
+    HVX_Vector S1 = hvx_vec_splat_f16(s1);
+
+    uint32_t i = 0;
+    #pragma unroll(2)
+    for (i = 0; i < nvec; ++i) {
+        // Multiply x * s -> pair of F32 vectors
+        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
+        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
+
+        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
+        HVX_Vector xs_p_hi = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
+
+        ptr_y[i * 2]     = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_lo, ptr_y[i * 2]));
+        ptr_y[i * 2 + 1] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs_p_hi, ptr_y[i * 2 + 1]));
+    }
+
+    if (nloe) {
+        HVX_VectorPair xs0_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x0[i]), S0);
+        HVX_VectorPair xs1_p = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(ptr_x1[i]), S1);
+
+        HVX_Vector xs_p_lo = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xs0_p), Q6_V_lo_W(xs1_p));
+        HVX_Vector xs      = xs_p_lo;
+        i = 2 * i;  // index for ptr_y
+
+        if (nloe >= 32) {
+            ptr_y[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
+            nloe -= 32; ++i;
+            xs = Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_hi_W(xs0_p), Q6_V_hi_W(xs1_p));
+        }
+
+        if (nloe) {
+            HVX_Vector xy = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(xs, ptr_y[i]));
+            hvx_vec_store_a(&ptr_y[i], nloe * 4, xy);
+        }
+    }
+}
+
 #define FLASH_ATTN_BLOCK_SIZE 128
 
-static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, int nth) {
+struct htp_fa_context {
+    const struct htp_ops_context * octx;
+
+    struct fastdiv_values src0_div21;
+    struct fastdiv_values src0_div1;
+
+    struct fastdiv_values broadcast_rk2;
+    struct fastdiv_values broadcast_rk3;
+    struct fastdiv_values broadcast_rv2;
+    struct fastdiv_values broadcast_rv3;
+
+    struct fastdiv_values src3_div2;
+    struct fastdiv_values src3_div3;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+
+    uint32_t n_blocks;
+
+    size_t size_q_row_padded;
+    size_t size_k_row_padded;
+    size_t size_v_row_padded;
+
+    size_t size_k_block;
+    size_t size_v_block;
+    size_t size_m_block;
+
+    bool is_q_fp32;
+};
+
+static inline void hvx_scale_vec_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, HVX_Vector vs) {
+    assert((size_t) dst % 128 == 0);
+    assert((size_t) src % 128 == 0);
+
+    const HVX_Vector * restrict vsrc = (const HVX_Vector * restrict) src;
+    HVX_Vector * restrict vdst       = (HVX_Vector * restrict) dst;
+
+    const uint32_t nvec = n / VLEN_FP32;
+    const uint32_t nloe = n % VLEN_FP32;
+
+    uint32_t i = 0;
+    #pragma unroll(4)
+    for (; i < nvec; ++i) {
+        vdst[i] = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs));
+    }
+    if (nloe) {
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs);
+        hvx_vec_store_a(&vdst[i], nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v));
+    }
+}
+
+static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_fa_context * factx = (struct htp_fa_context *) data;
+    const struct htp_ops_context * octx = factx->octx;
     const struct htp_tensor * q = &octx->src0;
     const struct htp_tensor * k = &octx->src1;
     const struct htp_tensor * v = &octx->src2;
     const struct htp_tensor * mask  = (octx->src3.data) ? &octx->src3 : NULL;
     const struct htp_tensor * sinks = (octx->src4.data) ? &octx->src4 : NULL;
-    struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * dst = &octx->dst;
 
     const uint32_t neq0 = q->ne[0];
     const uint32_t neq1 = q->ne[1];
@@ -304,18 +295,6 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     const uint32_t nb2 = dst->nb[2];
     const uint32_t nb3 = dst->nb[3];
 
-    float scale         = 1.0f;
-    float max_bias      = 0.0f;
-    float logit_softcap = 0.0f;
-
-    memcpy(&scale,         (float *) octx->op_params + 0, sizeof(float));
-    memcpy(&max_bias,      (float *) octx->op_params + 1, sizeof(float));
-    memcpy(&logit_softcap, (float *) octx->op_params + 2, sizeof(float));
-
-    if (logit_softcap != 0) {
-        scale /= logit_softcap;
-    }
-
     // total rows in q
     const uint32_t nr = neq1*neq2*neq3;
 
@@ -331,18 +310,8 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     const uint32_t DV = nev0;
 
     const size_t size_q_row = DK * ((q->type == HTP_TYPE_F32) ? 4 : 2);
-    const size_t size_q_row_padded = hex_round_up(size_q_row, 128);
-
     const size_t size_k_row = DK * sizeof(__fp16);
     const size_t size_v_row = DV * sizeof(__fp16);
-    const size_t size_m_row = FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16); // Treat block as one row for mask
-
-    const size_t size_k_row_padded = hex_round_up(size_k_row, 128);
-    const size_t size_v_row_padded = hex_round_up(size_v_row, 128);
-
-    const size_t size_k_block = size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    const size_t size_v_block = size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    const size_t size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
 
     // Scratchpad buffers for Q, K, V, Mask, and VKQ32 accumulator
     uint8_t * spad_q = octx->src0_spad.data + octx->src0_spad.size_per_thread * ith;
@@ -351,31 +320,28 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     uint8_t * spad_m = octx->src3_spad.data + octx->src3_spad.size_per_thread * ith;
     uint8_t * spad_a = octx->dst_spad.data  + octx->dst_spad.size_per_thread  * ith;
 
-    const uint32_t n_head = neq2;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+    const HVX_Vector logit_cap = hvx_vec_splat_f32(factx->logit_softcap);
 
     for (uint32_t ir = ir0; ir < ir1; ++ir) {
-        const uint32_t iq3 = fastdiv(ir, &octx->src0_div21);
-        const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &octx->src0_div1);
+        const uint32_t iq3 = fastdiv(ir, &factx->src0_div21);
+        const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &factx->src0_div1);
         const uint32_t iq1 = (ir - iq3*neq2*neq1 - iq2 * neq1);
 
-        const uint32_t ik3 = fastdiv(iq3, &octx->broadcast_rk3);
-        const uint32_t ik2 = fastdiv(iq2, &octx->broadcast_rk2);
+        const uint32_t ik3 = fastdiv(iq3, &factx->broadcast_rk3);
+        const uint32_t ik2 = fastdiv(iq2, &factx->broadcast_rk2);
 
-        const uint32_t iv3 = fastdiv(iq3, &octx->broadcast_rv3);
-        const uint32_t iv2 = fastdiv(iq2, &octx->broadcast_rv2);
+        const uint32_t iv3 = fastdiv(iq3, &factx->broadcast_rv3);
+        const uint32_t iv2 = fastdiv(iq2, &factx->broadcast_rv2);
 
         // Fetch Q row
         const uint8_t * q_row_ptr = (const uint8_t *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3);
-        dma_queue_push(dma, dma_make_ptr(spad_q, q_row_ptr), size_q_row_padded, nbq1, size_q_row, 1);
+        dma_queue_push(dma, dma_make_ptr(spad_q, q_row_ptr), factx->size_q_row_padded, nbq1, size_q_row, 1);
 
         const uint32_t h = iq2; // head index
-        const float slope = (max_bias > 0.0f) ? (h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1)) : 1.0f;
+        const float slope = (factx->max_bias > 0.0f) ? (h < factx->n_head_log2 ? powf(factx->m0, h + 1) : powf(factx->m1, 2*(h - factx->n_head_log2) + 1)) : 1.0f;
 
-        float S = 0.0f;      // sum
-        float M = -INFINITY; // maximum KQ value
+        HVX_Vector S_vec = hvx_vec_splat_f32(0.0f);
+        HVX_Vector M_vec = hvx_vec_splat_f32(-INFINITY);
 
         // Clear accumulator
         hvx_splat_f32_a(spad_a, 0, DV);
@@ -383,40 +349,42 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
 
         const __fp16 * mp_base = NULL;
         if (mask) {
-            const uint32_t im2 = fastmodulo(iq2, mask->ne[2], &octx->src3_div2);
-            const uint32_t im3 = fastmodulo(iq3, mask->ne[3], &octx->src3_div3);
+            const uint32_t im2 = fastmodulo(iq2, mask->ne[2], &factx->src3_div2);
+            const uint32_t im3 = fastmodulo(iq3, mask->ne[3], &factx->src3_div3);
             mp_base = (const __fp16 *) ((const uint8_t *) mask->data + iq1*mask->nb[1] + im2*mask->nb[2] + im3*mask->nb[3]);
         }
 
-        const uint32_t n_blocks = (nek1 + FLASH_ATTN_BLOCK_SIZE - 1) / FLASH_ATTN_BLOCK_SIZE;
-
         // Prefetch first two blocks
-        for (uint32_t ib = 0; ib < MIN(n_blocks, 2); ++ib) {
+        for (uint32_t ib = 0; ib < MIN(factx->n_blocks, 2); ++ib) {
             const uint32_t ic_start = ib * FLASH_ATTN_BLOCK_SIZE;
             const uint32_t current_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - ic_start);
 
             // K
             const uint8_t * k_src = (const uint8_t *) k->data + (ic_start*nbk1 + ik2*nbk2 + ik3*nbk3);
-            uint8_t * k_dst = spad_k + (ib % 2) * size_k_block;
-            dma_queue_push(dma, dma_make_ptr(k_dst, k_src), size_k_row_padded, nbk1, size_k_row, current_block_size);
+            uint8_t * k_dst = spad_k + (ib % 2) * factx->size_k_block;
+            dma_queue_push(dma, dma_make_ptr(k_dst, k_src), factx->size_k_row_padded, nbk1, size_k_row, current_block_size);
 
             // V
             const uint8_t * v_src = (const uint8_t *) v->data + (ic_start*nbv1 + iv2*nbv2 + iv3*nbv3);
-            uint8_t * v_dst = spad_v + (ib % 2) * size_v_block;
-            dma_queue_push(dma, dma_make_ptr(v_dst, v_src), size_v_row_padded, nbv1, size_v_row, current_block_size);
+            uint8_t * v_dst = spad_v + (ib % 2) * factx->size_v_block;
+            dma_queue_push(dma, dma_make_ptr(v_dst, v_src), factx->size_v_row_padded, nbv1, size_v_row, current_block_size);
 
             // Mask
             if (mask) {
                 const uint8_t * m_src = (const uint8_t *) (mp_base + ic_start);
-                uint8_t * m_dst = spad_m + (ib % 2) * size_m_block;
+                uint8_t * m_dst = spad_m + (ib % 2) * factx->size_m_block;
                 // Mask is 1D contiguous for this row
                 dma_queue_push(dma, dma_make_ptr(m_dst, m_src), current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
             }
         }
 
-        const uint8_t * q_ptr_vtcm = dma_queue_pop(dma).dst;
+        uint8_t * q_ptr_vtcm = dma_queue_pop(dma).dst;
+        if (factx->is_q_fp32) {
+            hvx_copy_f16_f32_aa(q_ptr_vtcm, q_ptr_vtcm, DK);  // inplace convert f32 to f16
+        }
 
-        for (uint32_t ib = 0; ib < n_blocks; ++ib) {
+        const HVX_Vector slope_vec = hvx_vec_splat_f16(slope);
+        for (uint32_t ib = 0; ib < factx->n_blocks; ++ib) {
             const uint32_t ic_start = ib * FLASH_ATTN_BLOCK_SIZE;
             const uint32_t current_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - ic_start);
 
@@ -428,8 +396,6 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             // Inner loop processing the block from VTCM
             uint32_t ic = 0;
 
-            const bool is_q_fp32 = (q->type == HTP_TYPE_F32);
-
             // Process in blocks of 32 (VLEN_FP32)
             static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 <= 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
             HVX_Vector_x4 scores_x4;
@@ -437,22 +403,18 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
                 float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
-                for (int j = 0; j < VLEN_FP32; j += 2) {
+                for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
                     const uint32_t cur_ic = ic + j;
-                    const uint8_t * k_ptr = k_base + cur_ic * size_k_row_padded;
-                    if (is_q_fp32) {
-                        hvx_dot_f32_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
-                    } else {
-                        hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
-                    }
+                    const uint8_t * k_ptr = k_base + cur_ic * factx->size_k_row_padded;
+                    hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + factx->size_k_row_padded, DK, factx->scale);
                 }
 
                 HVX_Vector scores = *(HVX_Vector *) scores_arr;
 
                 // 2. Softcap
-                if (logit_softcap != 0.0f) {
+                if (factx->logit_softcap != 0.0f) {
                     scores = hvx_vec_tanh_f32(scores);
-                    scores = Q6_Vqf32_vmpy_VsfVsf(scores, hvx_vec_splat_f32(logit_softcap));
+                    scores = Q6_Vqf32_vmpy_VsfVsf(scores, logit_cap);
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
@@ -460,70 +422,59 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 if (mask) {
                     const __fp16 * mp = m_base + ic;
                     HVX_Vector m_vals_f16 = *(const HVX_UVector *) mp;
-
-                    HVX_Vector one_f16 = Q6_Vh_vsplat_R(0x3c00);
-                    HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), one_f16);
-
-                    HVX_Vector m_vals_f32 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(m_vals_f32_pair));
-
-                    HVX_Vector slope_vec = hvx_vec_splat_f32(slope);
-                    HVX_Vector add_val = Q6_Vqf32_vmpy_VsfVsf(m_vals_f32, slope_vec);
-                    scores = Q6_Vqf32_vadd_VsfVsf(scores, Q6_Vsf_equals_Vqf32(add_val));
+                    HVX_VectorPair m_vals_f32_pair = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(m_vals_f16), slope_vec);
+                    HVX_Vector add_val = Q6_V_lo_W(m_vals_f32_pair);
+                    scores = Q6_Vqf32_vadd_Vqf32Vsf(add_val, scores);
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
                 scores_x4.v[iv] = scores;
-                v_max = Q6_Vsf_vmax_VsfVsf(scores, v_max);
+                v_max = hvx_vec_reduce_max2_f32(scores, v_max); // All lanes have block max
             }
 
             {
                 // 4. Online Softmax Update
-                v_max = hvx_vec_reduce_max_f32(v_max);
-                float m_block = hvx_vec_get_f32(v_max);
-                float M_old = M;
-                float M_new = (m_block > M) ? m_block : M;
-                M = M_new;
+                HVX_Vector M_new_vec = Q6_Vsf_vmax_VsfVsf(v_max, M_vec);
+                HVX_Vector diff_vec  = Q6_Vqf32_vsub_VsfVsf(M_vec, M_new_vec);
+                HVX_Vector ms_vec    = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(diff_vec));
+                M_vec = M_new_vec;
 
-                const float ms = expf(M_old - M_new);
-                hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
+                hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
-                HVX_Vector M_new_vec = hvx_vec_splat_f32(M_new);
                 HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
                 for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
                     HVX_Vector scores = scores_x4.v[iv];
-                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
+                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_vec);
                     HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
 
                     p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
 
                     // 5. Accumulate V
                     float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
-                    *(HVX_Vector*)p_arr = P;
+                    *(HVX_Vector *) p_arr = P;
 
-                    for (int j = 0; j < VLEN_FP32; ++j) {
-                        const uint32_t cur_ic = ic2 + j;
-                        const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
-                        hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    for (uint32_t j = 0; j < VLEN_FP32; j += 2) {
+                        const uint32_t  cur_ic = ic2 + j;
+                        const uint8_t * v_ptr  = v_base + cur_ic * factx->size_v_row_padded;
+                        hvx_mad_f32_f16_aa_rx2(VKQ32, v_ptr, v_ptr + factx->size_v_row_padded, p_arr[j], p_arr[j + 1], DV);
                     }
                 }
 
                 p_sum_vec = hvx_vec_reduce_sum_f32(p_sum_vec);
-                S = S * ms + hvx_vec_get_f32(p_sum_vec);
+                S_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(S_vec, ms_vec)), p_sum_vec));
             }
 
+            // Sync scalars for leftover/next block if needed
+            float M = hvx_vec_get_f32(M_vec);
+            float S = hvx_vec_get_f32(S_vec);
+
             // Leftover
             for (; ic < current_block_size; ++ic) {
                 float s_val;
-                const uint8_t * k_ptr = k_base + ic * size_k_row_padded;
-
-                if (is_q_fp32) {
-                    hvx_dot_f32_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
-                } else {
-                    hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
-                }
-
-                if (logit_softcap != 0.0f) {
-                    s_val = logit_softcap * tanhf(s_val);
+                const uint8_t * k_ptr = k_base + ic * factx->size_k_row_padded;
+                hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, factx->scale);
+                if (factx->logit_softcap != 0.0f) {
+                    s_val = factx->logit_softcap * tanhf(s_val);
                 }
 
                 if (mask) {
@@ -532,37 +483,42 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 }
 
                 const float Mold = M;
-                float ms = 1.0f;
                 float vs = 1.0f;
 
                 if (s_val > M) {
                     M = s_val;
-                    ms = expf(Mold - M);
-                    hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
+                    HVX_Vector diff_vec = hvx_vec_splat_f32(Mold - M);
+                    HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
+                    hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
+
+                    float ms = hvx_vec_get_f32(ms_vec);
+                    S = S * ms + vs;
                 } else {
-                    vs = expf(s_val - M);
+                    HVX_Vector diff_vec = hvx_vec_splat_f32(s_val - M);
+                    vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
+                    S += vs;
                 }
 
-                const uint8_t * v_ptr = v_base + ic * size_v_row_padded;
+                const uint8_t * v_ptr = v_base + ic * factx->size_v_row_padded;
 
                 hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, vs);
-
-                S = S * ms + vs;
             }
+            M_vec = hvx_vec_splat_f32(M);
+            S_vec = hvx_vec_splat_f32(S);
 
             // Issue DMA for next+1 block (if exists)
-            if (ib + 2 < n_blocks) {
+            if (ib + 2 < factx->n_blocks) {
                 const uint32_t next_ib = ib + 2;
                 const uint32_t next_ic_start = next_ib * FLASH_ATTN_BLOCK_SIZE;
                 const uint32_t next_block_size = MIN(FLASH_ATTN_BLOCK_SIZE, nek1 - next_ic_start);
 
                 // K
                 const uint8_t * k_src = (const uint8_t *) k->data + (next_ic_start*nbk1 + ik2*nbk2 + ik3*nbk3);
-                dma_queue_push(dma, dma_make_ptr(k_base, k_src), size_k_row_padded, nbk1, size_k_row, next_block_size);
+                dma_queue_push(dma, dma_make_ptr(k_base, k_src), factx->size_k_row_padded, nbk1, size_k_row, next_block_size);
 
                 // V
                 const uint8_t * v_src = (const uint8_t *) v->data + (next_ic_start*nbv1 + iv2*nbv2 + iv3*nbv3);
-                dma_queue_push(dma, dma_make_ptr(v_base, v_src), size_v_row_padded, nbv1, size_v_row, next_block_size);
+                dma_queue_push(dma, dma_make_ptr(v_base, v_src), factx->size_v_row_padded, nbv1, size_v_row, next_block_size);
 
                 // Mask
                 if (mask) {
@@ -573,20 +529,26 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
         }
 
         // sinks
+        float M = hvx_vec_get_f32(M_vec);
+        float S = hvx_vec_get_f32(S_vec);
+
         if (sinks) {
             const float s = ((float *)((char *) sinks->data))[h];
 
-            float ms = 1.0f;
             float vs = 1.0f;
 
             if (s > M) {
-                ms = expf(M - s);
-                hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
-            } else {
-                vs = expf(s - M);
-            }
+                HVX_Vector diff_vec = hvx_vec_splat_f32(M - s);
+                HVX_Vector ms_vec   = hvx_vec_exp_f32(diff_vec);
+                hvx_scale_vec_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms_vec);
 
-            S = S * ms + vs;
+                float ms = hvx_vec_get_f32(ms_vec);
+                S = S * ms + vs;
+            } else {
+                HVX_Vector diff_vec = hvx_vec_splat_f32(s - M);
+                vs = hvx_vec_get_f32(hvx_vec_exp_f32(diff_vec));
+                S += vs;
+            }
         }
 
         const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
@@ -609,53 +571,73 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
     }
 }
 
-static void htp_flash_attn_ext_job(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = data;
-    flash_attn_ext_f16_thread(octx, i, n);
-}
-
 int op_flash_attn_ext(struct htp_ops_context * octx) {
     const struct htp_tensor * q = &octx->src0;
     const struct htp_tensor * k = &octx->src1;
     const struct htp_tensor * v = &octx->src2;
-    const struct htp_tensor * mask = (octx->src3.type != HTP_TYPE_COUNT) ? &octx->src3 : NULL;
-    struct htp_tensor * dst = &octx->dst;
+    const struct htp_tensor * mask = (octx->src3.data) ? &octx->src3 : NULL;
+    const struct htp_tensor * dst = &octx->dst;
 
     // Check support
-    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) ||
-        k->type != HTP_TYPE_F16 ||
-        v->type != HTP_TYPE_F16) {
+    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) || k->type != HTP_TYPE_F16 || v->type != HTP_TYPE_F16) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    octx->src0_div21 = init_fastdiv_values(q->ne[2] * q->ne[1]);
-    octx->src0_div1  = init_fastdiv_values(q->ne[1]);
+    struct htp_fa_context factx;
+    factx.octx = octx;
 
-    octx->broadcast_rk2 = init_fastdiv_values(q->ne[2]/k->ne[2]);
-    octx->broadcast_rk3 = init_fastdiv_values(q->ne[3]/k->ne[3]);
-    octx->broadcast_rv2 = init_fastdiv_values(q->ne[2]/v->ne[2]);
-    octx->broadcast_rv3 = init_fastdiv_values(q->ne[3]/v->ne[3]);
+    factx.src0_div21 = init_fastdiv_values(q->ne[2] * q->ne[1]);
+    factx.src0_div1  = init_fastdiv_values(q->ne[1]);
+
+    factx.broadcast_rk2 = init_fastdiv_values(q->ne[2]/k->ne[2]);
+    factx.broadcast_rk3 = init_fastdiv_values(q->ne[3]/k->ne[3]);
+    factx.broadcast_rv2 = init_fastdiv_values(q->ne[2]/v->ne[2]);
+    factx.broadcast_rv3 = init_fastdiv_values(q->ne[3]/v->ne[3]);
 
     if (mask) {
-        octx->src3_div2 = init_fastdiv_values(mask->ne[2]);
-        octx->src3_div3 = init_fastdiv_values(mask->ne[3]);
+        factx.src3_div2 = init_fastdiv_values(mask->ne[2]);
+        factx.src3_div3 = init_fastdiv_values(mask->ne[3]);
     }
 
-    size_t size_q_row_padded = hex_round_up(q->ne[0] * (q->type == HTP_TYPE_F32 ? 4 : 2), 128);
-    size_t size_k_row_padded = hex_round_up(k->ne[0] * sizeof(__fp16), 128);
-    size_t size_v_row_padded = hex_round_up(v->ne[0] * sizeof(__fp16), 128);
+    factx.is_q_fp32 = (q->type == HTP_TYPE_F32);
+    factx.size_q_row_padded = hex_round_up(q->ne[0] * (factx.is_q_fp32 ? 4 : 2), 128);
+    factx.size_k_row_padded = hex_round_up(k->ne[0] * sizeof(__fp16), 128);
+    factx.size_v_row_padded = hex_round_up(v->ne[0] * sizeof(__fp16), 128);
 
-    size_t size_q_block = size_q_row_padded * 1; // single row for now
-    size_t size_k_block = size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    size_t size_v_block = size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
-    size_t size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
+    size_t size_q_block = factx.size_q_row_padded * 1; // single row for now
+    factx.size_k_block = factx.size_k_row_padded * FLASH_ATTN_BLOCK_SIZE;
+    factx.size_v_block = factx.size_v_row_padded * FLASH_ATTN_BLOCK_SIZE;
+    factx.size_m_block = hex_round_up(FLASH_ATTN_BLOCK_SIZE * sizeof(__fp16), 128);
+
+    factx.n_blocks = (k->ne[1] + FLASH_ATTN_BLOCK_SIZE - 1) / FLASH_ATTN_BLOCK_SIZE;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) octx->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) octx->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) octx->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
+
+    factx.scale = scale;
+    factx.max_bias = max_bias;
+    factx.logit_softcap = logit_softcap;
+
+    uint32_t n_head = q->ne[2];
+    factx.n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+    factx.m0 = powf(2.0f, -(max_bias       ) / factx.n_head_log2);
+    factx.m1 = powf(2.0f, -(max_bias / 2.0f) / factx.n_head_log2);
 
     size_t size_vkq_acc = hex_round_up(v->ne[0] * sizeof(float), 128); // VKQ32
 
     octx->src0_spad.size_per_thread = size_q_block * 1;
-    octx->src1_spad.size_per_thread = size_k_block * 2;
-    octx->src2_spad.size_per_thread = size_v_block * 2;
-    octx->src3_spad.size_per_thread = mask ? size_m_block * 2 : 0;
+    octx->src1_spad.size_per_thread = factx.size_k_block * 2;
+    octx->src2_spad.size_per_thread = factx.size_v_block * 2;
+    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * 2 : 0;
     octx->dst_spad.size_per_thread  = size_vkq_acc;
 
     octx->src0_spad.size = octx->src0_spad.size_per_thread * octx->n_threads;
@@ -677,7 +659,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size;
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        worker_pool_run_func(octx->ctx->worker_pool, htp_flash_attn_ext_job, octx, octx->n_threads);
+        worker_pool_run_func(octx->ctx->worker_pool, flash_attn_ext_f16_thread, &factx, octx->n_threads);
     }
 
     return HTP_STATUS_OK;

ggml/src/ggml-hexagon/htp/htp-msg.h

@@ -42,32 +42,36 @@ enum htp_data_type {
     HTP_TYPE_COUNT
 };
 
-// These values are manually translated over to HTP
-// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
+// Do not reorder first 4 (used as an index)
 enum htp_op {
-    HTP_OP_MUL            = 0,
-    HTP_OP_ADD            = 1,
-    HTP_OP_SUB            = 2,
-    HTP_OP_DIV            = 3,
-    HTP_OP_MUL_MAT        = 4,
-    HTP_OP_MUL_MAT_ID     = 5,
-    HTP_OP_RMS_NORM       = 6,
-    HTP_OP_UNARY_SILU     = 7,
-    HTP_OP_UNARY_GELU     = 8,
-    HTP_OP_GLU_SWIGLU     = 9,
-    HTP_OP_GLU_SWIGLU_OAI = 10,
-    HTP_OP_SOFTMAX        = 11,
-    HTP_OP_ADD_ID         = 12,
-    HTP_OP_ROPE           = 13,
-    HTP_OP_FLASH_ATTN_EXT = 14,
-    HTP_OP_SET_ROWS       = 15,
-    HTP_OP_SCALE          = 16,
-    HTP_OP_GET_ROWS       = 17,
-    HTP_OP_CPY            = 18,
+    HTP_OP_MUL = 0,
+    HTP_OP_ADD = 1,
+    HTP_OP_SUB = 2,
+    HTP_OP_DIV = 3,
+    HTP_OP_MUL_MAT,
+    HTP_OP_MUL_MAT_ID,
+    HTP_OP_RMS_NORM,
+    HTP_OP_UNARY_SILU,
+    HTP_OP_UNARY_GELU,
+    HTP_OP_GLU_SWIGLU,
+    HTP_OP_GLU_SWIGLU_OAI,
+    HTP_OP_GLU_GEGLU,
+    HTP_OP_SOFTMAX,
+    HTP_OP_ADD_ID,
+    HTP_OP_ROPE,
+    HTP_OP_FLASH_ATTN_EXT,
+    HTP_OP_SET_ROWS,
+    HTP_OP_GET_ROWS,
+    HTP_OP_SCALE,
+    HTP_OP_CPY,
+    HTP_OP_ARGSORT,
+    HTP_OP_SQR,
+    HTP_OP_SQRT,
+    HTP_OP_SUM_ROWS,
     INVALID
 };
 
-static inline size_t htp_type_block_size(uint32_t t) {
+static inline size_t htp_t_block_size(uint32_t t) {
     switch (t) {
         case HTP_TYPE_F32:
             return 1;
@@ -103,22 +107,6 @@ static inline size_t htp_type_nbytes(uint32_t t) {
     return 0;
 }
 
-static const char * htp_type_name(uint32_t t) {
-    switch (t) {
-        case HTP_TYPE_F32:
-            return "fp32";
-        case HTP_TYPE_F16:
-            return "fp16";
-        case HTP_TYPE_Q4_0:
-            return "q4_0";
-        case HTP_TYPE_Q8_0:
-            return "q8_0";
-        case HTP_TYPE_MXFP4:
-            return "mxfp4";
-    }
-    return 0;
-}
-
 // Internal types
 #define QK_Q4_0x4x2  256  // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
 #define QK_Q8_0x4x2  256  // 4x Q8_0 blocks concat with next 4x Q8_0 blocks

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -64,25 +64,12 @@ struct htp_ops_context {
     struct fastdiv_values broadcast_rv2;
     struct fastdiv_values broadcast_rv3;
 
-    struct fastdiv_values mm_div_ne12_ne1; // fastdiv values for ne12 * ne1
-    struct fastdiv_values mm_div_ne1;      // fastdiv values for ne1
-    struct fastdiv_values mm_div_r2;       // fastdiv values for ne12 / ne02
-    struct fastdiv_values mm_div_r3;       // fastdiv values for ne13 / ne03
-
     struct fastdiv_values set_rows_div_ne12; // fastdiv values for ne12
     struct fastdiv_values set_rows_div_ne11; // fastdiv values for ne11
 
     struct fastdiv_values get_rows_div_ne10;      // fastdiv values for ne10
     struct fastdiv_values get_rows_div_ne10_ne11; // fastdiv values for ne10 * ne11
 
-    struct fastdiv_values cpy_div_ne01; // fastdiv values for ne01
-    struct fastdiv_values cpy_div_ne02; // fastdiv values for ne02
-    struct fastdiv_values cpy_div_ne03; // fastdiv values for ne03
-
-    struct fastdiv_values cpy_rshp_div_n0;       // fastdiv values for ne00
-    struct fastdiv_values cpy_rshp_div_n1n0;     // fastdiv values for ne00*ne01
-    struct fastdiv_values cpy_rshp_div_n2n1n0;   // fastdiv values for ne00*ne01*ne02
-
     uint32_t flags;
 };
 
@@ -90,6 +77,7 @@ int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
+int op_sum_rows(struct htp_ops_context * octx);
 int op_activations(struct htp_ops_context * octx);
 int op_softmax(struct htp_ops_context * octx);
 int op_add_id(struct htp_ops_context * octx);
@@ -98,5 +86,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx);
 int op_set_rows(struct htp_ops_context * octx);
 int op_get_rows(struct htp_ops_context * octx);
 int op_cpy(struct htp_ops_context * octx);
+int op_argsort(struct htp_ops_context * octx);
 
 #endif /* HTP_OPS_H */

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -46,127 +46,76 @@
 #define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
 #endif
 
-// ADD variants
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \
 
-static inline void hvx_add_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD);
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
+
+// Dispatcher logic
+#define HVX_BINARY_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128)) { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
+        } \
+    } else { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
+        } \
+    } \
 }
 
-static inline void hvx_add_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-// SUB variants
-
-static inline void hvx_sub_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-// MUL variants
-
-static inline void hvx_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-// Dispatchers
-
-static inline void hvx_add_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_add_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_add_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_add_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_add_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_sub_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_sub_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_sub_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_sub_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_sub_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_mul_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_mul_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_mul_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_mul_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_mul_f32_uu(dst, src0, src1, num_elems);
-    }
-}
+HVX_BINARY_DISPATCHER(hvx_add_f32)
+HVX_BINARY_DISPATCHER(hvx_sub_f32)
+HVX_BINARY_DISPATCHER(hvx_mul_f32)
 
 // Mul-Mul Optimized
-
 static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src0 % 128 == 0);
@@ -443,6 +392,68 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
     }
 }
 
+//
+// Square
+//
+
+#define hvx_sqr_loop_body(dst_type, src_type, vec_store)           \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(float);                          \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]);                        \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]);                   \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #undef HVX_OP_ADD
 #undef HVX_OP_SUB
 #undef HVX_OP_MUL
@@ -453,5 +464,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
+#undef DEFINE_HVX_BINARY_OP_VARIANTS
+#undef HVX_BINARY_DISPATCHER
 
 #endif // HVX_ARITH_H

ggml/src/ggml-hexagon/htp/hvx-base.h

@@ -66,6 +66,12 @@ static inline float hvx_vec_get_f32(HVX_Vector v) {
     return x;
 }
 
+static inline int32_t hvx_vec_get_i32(HVX_Vector v) {
+    int32_t __attribute__((aligned(128))) x;
+    hvx_vec_store_a(&x, 4, v);
+    return x;
+}
+
 static inline HVX_Vector hvx_vec_abs_f16(HVX_Vector v) {
     // abs by clearing the fp16 sign bit
     HVX_Vector mask = Q6_Vh_vsplat_R(0x7fff);

ggml/src/ggml-hexagon/htp/hvx-copy.h

@@ -136,8 +136,6 @@ static inline void hvx_copy_f32_uu(uint8_t * restrict dst, const uint8_t * restr
         dst_type * restrict vdst = (dst_type *) dst;                                \
         src_type * restrict vsrc = (src_type *) src;                                \
                                                                                     \
-        const HVX_Vector zero = Q6_V_vsplat_R(0);                                   \
-                                                                                    \
         const uint32_t elem_size = sizeof(__fp16);                                  \
         const uint32_t epv  = 128 / elem_size;                                      \
         const uint32_t nvec = n / epv;                                              \

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -0,0 +1,116 @@
+#ifndef HVX_DIV_H
+#define HVX_DIV_H
+
+#include <HAP_farf.h>
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "hvx-base.h"
+#include "hex-utils.h"
+#include "hvx-inverse.h"
+#include "hvx-arith.h"
+
+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
+#define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store)             \
+    do {                                                                             \
+        dst_type * restrict vdst = (dst_type *) dst;                                 \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                             \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                             \
+                                                                                     \
+        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                   \
+                                                                                     \
+        const uint32_t nvec = n / VLEN_FP32;                                         \
+        const uint32_t nloe = n % VLEN_FP32;                                         \
+                                                                                     \
+        uint32_t i = 0;                                                              \
+                                                                                     \
+        _Pragma("unroll(4)")                                                         \
+        for (; i < nvec; i++) {                                                      \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vdst[i] = res;                                                           \
+        }                                                                            \
+        if (nloe) {                                                                  \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res);                   \
+        }                                                                            \
+    } while(0)
+
+// 3-letter suffix variants
+static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_aau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_auu(dst, src0, src1, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uuu(dst, src0, src1, num_elems);
+        }
+    }
+}
+
+#undef HVX_OP_MUL
+
+#endif // HVX_DIV_H

ggml/src/ggml-hexagon/htp/hvx-sigmoid.h

@@ -91,6 +91,27 @@ static inline HVX_Vector hvx_vec_tanh_f32(HVX_Vector x) {
         }                                                       \
     } while(0)
 
+#define hvx_tanh_loop_body(dst_type, src_type, vec_store)       \
+    do {                                                        \
+        dst_type * restrict vdst = (dst_type *) dst;            \
+        src_type * restrict vsrc = (src_type *) src;            \
+                                                                \
+        const uint32_t epv  = 128 / sizeof(float);              \
+        const uint32_t nvec = n / epv;                          \
+        const uint32_t nloe = n % epv;                          \
+                                                                \
+        uint32_t i = 0;                                         \
+                                                                \
+        _Pragma("unroll(4)")                                    \
+        for (; i < nvec; i++) {                                 \
+             vdst[i] = hvx_vec_tanh_f32(vsrc[i]);               \
+        }                                                       \
+        if (nloe) {                                             \
+             HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]);        \
+             vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
+        }                                                       \
+    } while(0)
+
 static inline void hvx_sigmoid_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
     assert((unsigned long) dst % 128 == 0);
     assert((unsigned long) src % 128 == 0);
@@ -111,4 +132,10 @@ static inline void hvx_sigmoid_f32_uu(uint8_t * restrict dst, const uint8_t * re
     hvx_sigmoid_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }
 
+static inline void hvx_tanh_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_tanh_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
 #endif /* HVX_SIGMOID_H */

ggml/src/ggml-hexagon/htp/hvx-sqrt.h

@@ -12,11 +12,17 @@
 #define RSQRT_ONE_HALF     0x3f000000  // 0.5
 #define RSQRT_THREE_HALVES 0x3fc00000  // 1.5
 
+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
 static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
     //Algorithm :
     //  x2 = input*0.5
     //  y  = * (long *) &input
-    //  y  = 0x5f3759df - (y>>2)
+    //  y  = 0x5f3759df - (y>>1)
     //  y  = y*(threehalfs - x2*y*y)
 
     HVX_Vector rsqrtconst = Q6_V_vsplat_R(RSQRT_CONST);
@@ -57,4 +63,64 @@ static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
     return Q6_Vsf_equals_Vqf32(temp);
 }
 
+// Compute sqrt(x) as x*inv_sqrt(x)
+#define hvx_sqrt_f32_loop_body(dst_type, src_type, vec_store)                \
+    do {                                                                     \
+        dst_type * restrict vdst = (dst_type *) dst;                         \
+        src_type * restrict vsrc = (src_type *) src;                         \
+                                                                             \
+        const uint32_t nvec = n / VLEN_FP32;                                 \
+        const uint32_t nloe = n % VLEN_FP32;                                 \
+                                                                             \
+        uint32_t i = 0;                                                      \
+                                                                             \
+        _Pragma("unroll(4)")                                                 \
+        for (; i < nvec; i++) {                                              \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vdst[i] = sqrt_res;                                              \
+        }                                                                    \
+        if (nloe) {                                                          \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, sqrt_res);      \
+        }                                                                    \
+    } while(0)
+
+static inline void hvx_sqrt_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
+    if ((unsigned long) dst % 128 == 0) {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #endif /* HVX_SQRT_H */

ggml/src/ggml-hexagon/htp/hvx-utils.h

@@ -12,6 +12,7 @@
 #include "hvx-sigmoid.h"
 #include "hvx-sqrt.h"
 #include "hvx-arith.h"
+#include "hvx-div.h"
 #include "hvx-base.h"
 
 #endif /* HVX_UTILS_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -189,7 +189,7 @@ static int vtcm_release_callback(unsigned int rctx, void * state) {
     // otherwise we'll release it once we're done with the current Op.
 
     if (ctx->vtcm_inuse) {
-        ctx->vtcm_needs_release = false;
+        ctx->vtcm_needs_release = true;
         return 0;
     }
 
@@ -440,6 +440,45 @@ static void proc_matmul_req(struct htp_context *     ctx,
     send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
+static void proc_argsort_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[1];
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[0].fd     = bufs[1].fd;
+    rsp_bufs[0].ptr    = bufs[1].ptr;
+    rsp_bufs[0].offset = bufs[1].offset;
+    rsp_bufs[0].size   = bufs[1].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_argsort(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_cpy_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
     struct dspqueue_buffer rsp_bufs[1];
 
@@ -679,6 +718,45 @@ static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * re
     send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
+static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+
+    // We had written to the output buffer, we'd also need to flush it
+    rsp_bufs[0].fd     = bufs[1].fd;
+    rsp_bufs[0].ptr    = bufs[1].ptr;
+    rsp_bufs[0].offset = bufs[1].offset;
+    rsp_bufs[0].size   = bufs[1].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                         DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+
+    // Setup Op context
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_sum_rows(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_activations_req(struct htp_context *     ctx,
                                  struct htp_general_req * req,
                                  struct dspqueue_buffer * bufs,
@@ -951,6 +1029,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             case HTP_OP_MUL:
             case HTP_OP_ADD:
             case HTP_OP_SUB:
+            case HTP_OP_DIV:
                 if (n_bufs != 3) {
                     FARF(ERROR, "Bad binary-req buffer list");
                     continue;
@@ -968,6 +1047,25 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 proc_unary_req(ctx, &req, bufs);
                 break;
 
+            case HTP_OP_SQR:
+            case HTP_OP_SQRT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_unary_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_SUM_ROWS:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_sum_rows_req(ctx, &req, bufs);
+                break;
+
             case HTP_OP_UNARY_SILU:
             case HTP_OP_UNARY_GELU:
                 if (n_bufs != 2) {
@@ -980,6 +1078,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             case HTP_OP_GLU_SWIGLU:
             case HTP_OP_GLU_SWIGLU_OAI:
             case HTP_OP_SOFTMAX:
+            case HTP_OP_GLU_GEGLU:
                 if ((n_bufs != 2) && (n_bufs != 3)) {
                     FARF(ERROR, "Bad act-req buffer list");
                     continue;
@@ -1035,6 +1134,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 proc_cpy_req(ctx, &req, bufs);
                 break;
 
+            case HTP_OP_ARGSORT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad argsort-req buffer list");
+                    continue;
+                }
+                proc_argsort_req(ctx, &req, bufs);
+                break;
+
             default:
                 FARF(ERROR, "Unknown Op %u", req.op);
                 break;

ggml/src/ggml-hexagon/htp/sum-rows-ops.c

@@ -0,0 +1,115 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <string.h>
+#include <math.h>
+
+#include "hex-dma.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+
+#define sum_rows_preamble                       \
+    struct htp_tensor *src0 =  &octx->src0;\
+    struct htp_tensor *dst  = &octx->dst;  \
+                                           \
+    const uint32_t ne00 = src0->ne[0];     \
+    const uint32_t ne01 = src0->ne[1];     \
+    const uint32_t ne02 = src0->ne[2];     \
+    const uint32_t ne03 = src0->ne[3];     \
+                                           \
+    const uint32_t nb00 = src0->nb[0];     \
+    const uint32_t nb01 = src0->nb[1];     \
+    const uint32_t nb02 = src0->nb[2];     \
+    const uint32_t nb03 = src0->nb[3];     \
+                                           \
+    const uint32_t  ne0 = dst->ne[0];      \
+    const uint32_t  ne1 = dst->ne[1];      \
+    const uint32_t  ne2 = dst->ne[2];      \
+    const uint32_t  ne3 = dst->ne[3];      \
+                                           \
+    const uint32_t  nb0 = dst->nb[0];      \
+    const uint32_t  nb1 = dst->nb[1];      \
+    const uint32_t  nb2 = dst->nb[2];      \
+    const uint32_t  nb3 = dst->nb[3];      \
+
+static int sum_rows_thread_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+    sum_rows_preamble;
+
+    const uint32_t src0_nrows_per_thread  = octx->src0_nrows_per_thread;
+    const size_t src0_row_size = nb01;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return HTP_STATUS_OK;
+    }
+
+    int opt_path   = 0;
+    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src = (const uint8_t *) src0->data;
+    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+
+    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
+    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
+
+    for (uint32_t ir = 0; ir < src0_nrows_per_thread; ir++) {
+        const float * restrict src_local = src_th + (ir * ne00);
+
+        if (ir + 1 < src0_nrows_per_thread) {
+            hex_l2fetch(src_local + ne00, src0_row_size, src0_row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            dst_th[ir] = hvx_reduce_sum_f32_a((const uint8_t *) src_local, ne00);
+        } else {
+            dst_th[ir] = hvx_reduce_sum_f32((const uint8_t *) src_local, ne00);
+        }
+    }
+
+    return HTP_STATUS_OK;
+}
+
+static void sum_rows_work_f32(unsigned int n, unsigned int i, void *data) {
+    sum_rows_thread_f32((struct htp_ops_context *) data, n, i);
+}
+
+int op_sum_rows(struct htp_ops_context * octx) {
+    sum_rows_preamble;
+
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+
+    uint32_t n_jobs = MIN(n_threads, src0_nrows);
+    octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_work_f32, octx, n_jobs);
+
+    return HTP_STATUS_OK;
+}
+

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -132,6 +132,56 @@ static void rms_norm_htp_f32(const float * restrict src,
     }
 }
 
+static void sqr_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqr_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
+static void sqrt_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqrt_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
 static void unary_job_f32_per_thread(const struct htp_tensor * src,
                                      struct htp_tensor *       dst,
                                      uint8_t *                 spad,
@@ -181,6 +231,12 @@ static void unary_job_f32_per_thread(const struct htp_tensor * src,
         case HTP_OP_SCALE:
             scale_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
             break;
+        case HTP_OP_SQR:
+            sqr_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
+        case HTP_OP_SQRT:
+            sqrt_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
 
         default:
             break;
@@ -218,6 +274,14 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
             unary_op_func = unary_job_dispatcher_f32;
             op_type       = "scale-f32";
             break;
+        case HTP_OP_SQR:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqr-f32";
+            break;
+        case HTP_OP_SQRT:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqrt-f32";
+            break;
 
         default:
             FARF(ERROR, "Unsupported unary Op %u\n", octx->op);