cmake : add variable to skip installing tests (#19370)

When packaging downstream, there's usually little point in installing
test. The default behaviour remains the same.

.devops/nix/nixpkgs-instances.nix

@@ -4,7 +4,7 @@
   # the module `{ pkgs ... }: { /* config */ }` implicitly uses
   # `_module.args.pkgs` (defined in this case by flake-parts).
   perSystem =
-    { system, ... }:
+    { lib, system, ... }:
     {
       _module.args = {
         # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs
@@ -33,7 +33,7 @@
                 "CUDA EULA"
                 "cuDNN EULA"
               ]
-            ) (p.meta.licenses or [ p.meta.license ]);
+            ) (p.meta.licenses or (lib.toList p.meta.license));
         };
         # Ensure dependencies use ROCm consistently
         pkgsRocm = import inputs.nixpkgs {

.devops/nix/package-gguf-py.nix

@@ -3,6 +3,7 @@
   llamaVersion,
   numpy,
   tqdm,
+  requests,
   sentencepiece,
   pyyaml,
   poetry-core,
@@ -20,6 +21,7 @@ buildPythonPackage {
     tqdm
     sentencepiece
     pyyaml
+    requests
   ];
   src = lib.cleanSource ../../gguf-py;
   pythonImportsCheck = [

.devops/nix/scope.nix

@@ -7,13 +7,6 @@
 
 let
   pythonPackages = python3.pkgs;
-  buildPythonPackage = pythonPackages.buildPythonPackage;
-  numpy = pythonPackages.numpy;
-  tqdm = pythonPackages.tqdm;
-  sentencepiece = pythonPackages.sentencepiece;
-  pyyaml = pythonPackages.pyyaml;
-  poetry-core = pythonPackages.poetry-core;
-  pytestCheckHook = pythonPackages.pytestCheckHook;
 in
 
 # We're using `makeScope` instead of just writing out an attrset
@@ -23,17 +16,18 @@ in
 lib.makeScope newScope (self: {
   inherit llamaVersion;
   gguf-py = self.callPackage ./package-gguf-py.nix {
-    inherit
-      buildPythonPackage
+    inherit (pythonPackages)
       numpy
       tqdm
       sentencepiece
-      poetry-core
       pyyaml
       pytestCheckHook
+      requests
+      buildPythonPackage
+      poetry-core
       ;
   };
-  python-scripts = self.callPackage ./python-scripts.nix { inherit buildPythonPackage poetry-core; };
+  python-scripts = self.callPackage ./python-scripts.nix { inherit (pythonPackages) buildPythonPackage poetry-core; };
   llama-cpp = self.callPackage ./package.nix { };
   docker = self.callPackage ./docker.nix { };
   docker-min = self.callPackage ./docker.nix { interactive = false; };

.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/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-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

CMakeLists.txt

@@ -109,6 +109,7 @@ 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)
@@ -164,29 +165,6 @@ 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")
 

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

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023-2024 The ggml authors
+Copyright (c) 2023-2026 The ggml authors
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

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)

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

@@ -75,6 +75,8 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     ngram-map.cpp
     ngram-map.h
+    ngram-mod.cpp
+    ngram-mod.h
     peg-parser.cpp
     peg-parser.h
     preset.cpp

common/arg.cpp

@@ -3396,7 +3396,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
-        {"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v]",
+        {"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]",
         string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n",
             common_speculative_type_to_str(params.speculative.type).c_str()),
         [](common_params & params, const std::string & value) {
@@ -3410,6 +3410,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
             } else if (value == "ngram-map-k4v") {
                 params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V;
+            } else if (value == "ngram-mod") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD;
             } else {
                 throw std::invalid_argument("unknown speculative decoding type without draft model");
             }

common/common.h

@@ -171,6 +171,7 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
+    COMMON_SPECULATIVE_TYPE_NGRAM_MOD,
     COMMON_SPECULATIVE_TYPE_NGRAM_CACHE,   // self-speculative decoding with 3-level n-gram cache
     COMMON_SPECULATIVE_TYPE_COUNT          // number of types, unknown type
 };
@@ -252,6 +253,8 @@ struct common_params_model {
     std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
+struct common_ngram_mod;
+
 struct common_params_speculative {
     common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
 
@@ -269,6 +272,8 @@ struct common_params_speculative {
     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;
+
     std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding           // NOLINT
     std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding          // NOLINT
 

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/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);
         }
     }
 

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/jinja/value.h

@@ -12,6 +12,7 @@
 #include <set>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 namespace jinja {

common/ngram-map.cpp

@@ -7,6 +7,33 @@
 #include <cstdio>
 #include <sstream>
 
+// prime number used for LCG hash function (32 bit), it is near (sqrt(5) - 1)/2 * 2^32.
+#define LCG_FACTOR 2654435761UL
+
+// Compute the LCG hash of a n-gram of size len at offset start.
+static uint32_t common_ngram_map_hash(const llama_tokens & tokens, size_t start, size_t len) {
+    uint32_t hash = 0;
+    for (size_t i = 0; i < len; ++i) {
+        hash = hash * LCG_FACTOR + tokens[start + i];
+    }
+    return hash;
+}
+
+// Print the values of a sublist of `llama_tokens & inp` to a string in the form [v0, v1, v2, ...].
+static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length) {
+    std::ostringstream oss;
+    oss << '[';
+    for (size_t i = 0; i < length; ++i) {
+        if (i > 0) {
+            oss << ", ";
+        }
+        oss << inp[start + i];
+    }
+    oss << ']';
+    return oss.str();
+}
+
+
 // n-gram simple
 //
 
@@ -20,21 +47,15 @@
  * @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.
@@ -53,9 +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) {
@@ -100,7 +118,99 @@ llama_tokens common_ngram_simple_draft(
 // maximum number of counted values of a ngram map value.
 #define COMMON_NGRAM_MAX_VALUE_COUNT 16380
 
-static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length);
+void common_ngram_map_begin(
+    common_ngram_map & map, const llama_tokens & tokens) {
+    size_t size_begin = tokens.size();
+
+    LOG_DBG("%s: begin, idx_last_draft=%zu, new begin=%zu, #keys=%zu\n", __func__,
+            map.idx_last_check, size_begin, map.keys.size());
+
+    size_t count_map_entries_upd = 0;
+    if (!map.key_map.empty() && size_begin < map.idx_last_check) {
+        if (map.show_key_map_stats) {
+            // Print statistics of hash map map_key.
+            size_t count_nonzero = 0;
+            uint32_t min_idx = UINT32_MAX;
+            uint32_t max_idx = 0;
+            for (size_t i = 0; i < map.key_map.size(); ++i) {
+                uint32_t key_idx = map.key_map[i];
+                if (key_idx != 0) {
+                    ++count_nonzero;
+                    if (key_idx < min_idx) min_idx = key_idx;
+                    if (key_idx > max_idx) max_idx = key_idx;
+                }
+            }
+            if (count_nonzero == 0) {
+                min_idx = 0;
+            }
+            LOG_INF("%s: key_map stats: entries=%zu, min_idx=%u, max_idx=%u, key_map_last_idx=%u\n",
+                    __func__, count_nonzero, min_idx, max_idx, map.key_map_last_idx);
+        }
+
+        // Update the map from hash to key index (clear outdated entries).
+        for (size_t i = 0; i < map.key_map.size(); ++i) {
+            uint32_t key_idx = map.key_map[i];
+            if (key_idx >= map.size_last_begin) {
+                map.key_map[i] = 0;
+                count_map_entries_upd++;
+            }
+        }
+        map.key_map_last_idx = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+    }
+
+    if (size_begin < map.idx_last_check && !map.keys.empty()) {
+        // The next token generation will start at index size_begin.
+        // The tokens between map.size_last_begin and size_begin are no longer valid.
+        //
+        // Refresh map: Remove all entries with index >= map.size_last_begin.
+        size_t count_keys = map.keys.size();
+        size_t count_keys_del = 0;
+        size_t count_values_del = 0;
+        for (int32_t i = map.keys.size() - 1; i >= 0; --i) {
+            common_ngram_map_key & key = map.keys[i];
+            if (key.key_idx >= map.size_last_begin) {
+                // Delete the key.
+                LOG_DBG("%s: delete key %d at index %zu (>= size_last_begin=%zu)\n", __func__, i, key.key_idx, map.size_last_begin);
+                map.keys.erase(map.keys.begin() + i);
+                count_keys_del++;
+                continue;
+            }
+            if (map.key_only) {
+                continue;
+            }
+
+            // Check the indices of the values.
+            for (int16_t j = COMMON_NGRAM_MAX_VALUES - 1; j >= 0; --j) {
+                common_ngram_map_value & value = key.values[j];
+                if (value.value_idx >= map.size_last_begin) {
+                    // Delete the value.
+                    count_values_del++;
+
+                    // Move all values after this value to the left.
+                    for (uint16_t k = j; k < COMMON_NGRAM_MAX_VALUES - 1; ++k) {
+                        key.values[k] = key.values[k + 1];
+                    }
+                    // Clear the last value.
+                    key.values[COMMON_NGRAM_MAX_VALUES - 1].value_idx = 0;
+                    key.values[COMMON_NGRAM_MAX_VALUES - 1].value_num = 0;
+                }
+            }
+            if (key.values[0].value_idx == 0) {
+                // No values left, delete the key.
+                LOG_DBG("%s: delete key %d at index %zu (no values left)\n", __func__, i, key.key_idx);
+                map.keys.erase(map.keys.begin() + i);
+                count_keys_del++;
+            }
+        }
+
+        LOG_INF("%s: refresh map: idx_last_draft=%zu, new begin=%zu, #keys_checked=%zu, #keys_del=%zu, #values_del=%zu, #hashes_upd=%zu\n", __func__,
+                map.idx_last_check, size_begin,
+                count_keys, count_keys_del, count_values_del, count_map_entries_upd);
+    }
+
+    map.idx_last_check = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+    map.size_last_begin = size_begin;
+}
 
 void common_ngram_map_draft(common_ngram_map & map,
         const llama_tokens & inp, llama_token sampled,
@@ -116,6 +226,10 @@ void common_ngram_map_draft(common_ngram_map & map,
     if (cur_len < static_cast<size_t>(2 * n + m)) {
         return;
     }
+    if (cur_len >= static_cast<size_t>(UINT32_MAX)) {
+        // key_map uses uint32_t instead of size_t.
+        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
@@ -134,24 +248,92 @@ void common_ngram_map_draft(common_ngram_map & map,
 
     // search for the key in the map
     size_t match_pos = 0;
-    for (size_t j = cur_len - n - m - 1; j > 0; --j) {
-        bool match = true;
-        for (size_t k = 0; k < n; ++k) {
-            if (inp[j + k] != key_tokens[k]) {
-                match = false;
-                break;
+    if (map.size_last_begin > cur_len) {
+        GGML_ABORT("%s: map.size_last_begin > cur_len: %zu > %zu", __func__, map.size_last_begin, cur_len);
+    }
+    if (!map.key_map.empty()) {
+        // Search for the key in the map key_map from hash of ngrams to index of ngram.
+        uint32_t idx_hash = (common_ngram_map_hash(key_tokens, 0, n) % map.key_map.size());
+        uint32_t idx_key = map.key_map[idx_hash];
+        if (idx_key != 0 && idx_key < cur_len - n - m - 1) {
+            // Check if the key matches the key at idx_key (because of possible collisions).
+            bool match = true;
+            for (size_t k = 0; k < n; ++k) {
+                if (inp[idx_key + k] != key_tokens[k]) {
+                    match = false;
+                    break;
+                }
+            }
+            LOG_DBG("%s: key hash %x -> idx_key %d: match %d\n", __func__, idx_hash, idx_key, match ? 1 : 0);
+            if (match) {
+                match_pos = idx_key;
             }
         }
-        if (match) {
-           match_pos = j;
-           break;
+    }
+    if (match_pos == 0 && map.size_last_begin > (size_t) (n + m + 1)) {
+        // Search for the key in [1, map.size_last_begin - n - m -1], descending.
+        for (size_t j = map.size_last_begin - n - m - 1; j > map.key_map_last_idx; --j) {
+            // Check if the key matches the key.
+            bool match = true;
+            for (size_t k = 0; k < n; ++k) {
+                if (inp[j + k] != key_tokens[k]) {
+                    match = false;
+                    break;
+                }
+            }
+            if (match) {
+               match_pos = j;
+               break;
+            }
+        }
+    }
+    if (match_pos == 0) {
+        // In case of a reasoning chat, the part after size_last_begin may be deleted/reordered later.
+        //
+        // Search in [size_last_begin, cur_len - n - m - 1], descending.
+        for (size_t j = cur_len - n - m - 1; j > map.size_last_begin && j > map.key_map_last_idx; --j) {
+            bool match = true;
+            for (size_t k = 0; k < n; ++k) {
+                if (inp[j + k] != key_tokens[k]) {
+                    match = false;
+                    break;
+                }
+            }
+            if (match) {
+               match_pos = j;
+               break;
+            }
         }
     }
     if (match_pos > 0) {
-        LOG_INF("%s: cur_len = %zu, n = %d, m = %d, sz_tkns = %zu, sampled = %d, match_pos = %zu\n", __func__,
+        LOG_DBG("%s: cur_len = %zu, n = %d, m = %d, sz_tkns = %zu, sampled = %d, match_pos = %zu\n", __func__,
             cur_len, n, m, key_tokens.size(), sampled, match_pos);
     }
 
+    if (!map.key_map.empty()) {
+        // Add hashes of new ngrams in key_map.
+        //
+        // Use the same order as above.
+        if (map.size_last_begin > (size_t) (n + m + 1)) {
+            for (size_t j = map.size_last_begin - n - m - 1; j > map.key_map_last_idx; --j) {
+                // compute hash and store index of ngram at idx j in the map.
+                uint32_t idx_hash = (common_ngram_map_hash(inp, j, n) % map.key_map.size());
+                if (map.key_map[idx_hash] == 0) {
+                    map.key_map[idx_hash] = j; // collisions may occur
+                }
+            }
+        }
+
+        for (size_t j = cur_len - n - m - 1; j > map.size_last_begin && j > map.key_map_last_idx; --j) {
+            // compute hash and store index of ngram at idx j in the map.
+            uint32_t idx_hash = (common_ngram_map_hash(inp, j, n) % map.key_map.size());
+            if (map.key_map[idx_hash] == 0) {
+                map.key_map[idx_hash] = j;
+            }
+        }
+        map.key_map_last_idx = std::max(static_cast<uint32_t>(cur_len - n - m - 1), map.key_map_last_idx);
+    }
+
     if (match_pos == 0) {
         return;
     }
@@ -202,8 +384,8 @@ void common_ngram_map_draft(common_ngram_map & map,
             draft.push_back(inp[match_pos + n + i]);
         }
 
-        LOG_INF("%s: key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
-                key_offset, curr_key.key_num, draft.size());
+        LOG_DBG("%s: key_idx = %zu, key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
+                curr_key.key_idx, key_offset, curr_key.key_num, draft.size());
 
         map.last_draft_created   = false;
         map.last_draft_key_idx   = key_offset;
@@ -305,7 +487,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         }
     }
 
-    if (sum_occur > 0 && max_occur < 3 * sum_occur) {
+    if (sum_occur > 0 && max_occur < 2 * sum_occur) {
         // The most frequent value is not much more frequent than the other values.
         // We do not use the draft.
         return;
@@ -347,21 +529,3 @@ void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
             n_accepted, curr_value.n_accepted);
     curr_value.n_accepted = n_accepted;
 }
-
-// Helper functions.
-//
-
-// Print the values of a sublist of `llama_tokens & inp` to a string in the form [v0, v1, v2, ...].
-std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length) {
-    std::ostringstream oss;
-    oss << '[';
-    for (size_t i = 0; i < length; ++i) {
-        if (i > 0) {
-            oss << ", ";
-        }
-        oss << inp[start + i];
-    }
-    oss << ']';
-    return oss.str();
-}
-

common/ngram-map.h

@@ -9,8 +9,11 @@
 // 2. ngram_map: lookup of n-grams followed by m-grams in token history using a map.
 //    The map is a vector of key n-grams, and for each key n-gram there is a list of value m-grams.
 //
+// ref: https://github.com/ggml-org/llama.cpp/pull/18471
+//
 
 #include "llama.h"
+#include "common.h"
 
 #include <vector>
 
@@ -24,23 +27,9 @@ struct common_ngram_simple_config {
     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);
 
 
@@ -50,10 +39,13 @@ llama_tokens common_ngram_simple_draft(
 // maximum number of m-gram values stored for each key n-gram.
 #define COMMON_NGRAM_MAX_VALUES 4
 
+// number of entries in the (optional, size 0 to disable) map from ngram-hash to ngram-index.
+#define COMMON_NGRAM_HASH_MAP_SIZE 262144
+
 // 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)
+    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)
     int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
 };
 
@@ -73,23 +65,43 @@ struct common_ngram_map {
 
     bool key_only;       // true if only key n-grams are used, no values.
 
-    // first draft: vector only, no map.
     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.
+
     common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
                      uint16_t check_rate, uint16_t min_hits)
         : size_key(sz_key), size_value(sz_value), key_only(only_keys),
-          check_rate(check_rate), min_hits(min_hits) {}
+          check_rate(check_rate), 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
+    }
+
+    // In reasoning chats the previous reasoning block will be removed from context history.
+    // A rebuild of the ngram map is needed after that.
+
+    size_t   size_last_begin      = 0; // number of tokens at previous start of generation
 
     bool     last_draft_created   = false; // true if a draft was created at last call.
-    size_t   last_draft_key_idx   = 0; // index of last key used for draft generation.
+    size_t   last_draft_key_idx   = 0; // index of last key used for draft generation (0 = no draft)
     uint16_t last_draft_value_idx = 0; // index of last value used for draft generation.
 
     size_t   idx_last_check       = 0; // index of last check in context history
+
+    // optional map "hash to ngram-index" for faster lookup of n-grams. map is empty if unused.
+    //
+    // uint32_t instead of size_t (size of current histories is << UINT32_MAX)
+    std::vector<uint32_t> key_map;              // key_map[hash] = index of ngram in context window
+    uint32_t              key_map_last_idx = 0; // index of the last ngram added to key_map
 };
 
+// Initialize the n-gram map with the given token history.
+// map:                the ngram map to initialize.
+// tokens:             the token history to base the map on.
+void common_ngram_map_begin(
+    common_ngram_map & map,
+    const llama_tokens & tokens);
 
 // Searches for the n-gram in the history and checks whether a draft sequence should be generated.
 // map:                the ngram map to search in.

common/ngram-mod.cpp

@@ -0,0 +1,60 @@
+#include "ngram-mod.h"
+
+//
+// common_ngram_mod
+//
+
+common_ngram_mod::common_ngram_mod(uint16_t n, size_t size) : n(n), used(0) {
+    entries.resize(size);
+
+    reset();
+}
+
+size_t common_ngram_mod::idx(const entry_t * tokens) const {
+    size_t res = 0;
+
+    for (size_t i = 0; i < n; ++i) {
+        res = res*6364136223846793005ULL + tokens[i];
+    }
+
+    res = res % entries.size();
+
+    return res;
+}
+
+void common_ngram_mod::add(const entry_t * tokens) {
+    const size_t i = idx(tokens);
+
+    if (entries[i] == EMPTY) {
+        used++;
+    }
+
+    entries[i] = tokens[n];
+}
+
+common_ngram_mod::entry_t common_ngram_mod::get(const entry_t * tokens) const {
+    const size_t i = idx(tokens);
+
+    return entries[i];
+}
+
+void common_ngram_mod::reset() {
+    std::fill(entries.begin(), entries.end(), EMPTY);
+    used = 0;
+}
+
+size_t common_ngram_mod::get_n() const {
+    return n;
+}
+
+size_t common_ngram_mod::get_used() const {
+    return used;
+}
+
+size_t common_ngram_mod::size() const {
+    return entries.size();
+}
+
+size_t common_ngram_mod::size_bytes() const {
+    return entries.size() * sizeof(entries[0]);
+}

common/ngram-mod.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include <cstdint>
+#include <vector>
+#include <cstddef>
+
+//
+// common_ngram_mod
+// ref: https://github.com/ggml-org/llama.cpp/pull/19164
+//
+
+// basic n-gram hasher
+struct common_ngram_mod {
+    using entry_t = int32_t;
+
+    static constexpr entry_t EMPTY = -1;
+
+    common_ngram_mod(uint16_t n, size_t size);
+
+    size_t  idx(const entry_t * tokens) const;
+    void    add(const entry_t * tokens);
+    entry_t get(const entry_t * tokens) const; // return -1 if not found
+
+    void reset();
+
+    size_t get_n()    const;
+    size_t get_used() const;
+
+    size_t size()       const;
+    size_t size_bytes() const;
+
+private:
+    size_t n; // ngram size to hash
+
+    size_t used;
+
+    std::vector<entry_t> entries;
+};

common/speculative.cpp

@@ -6,6 +6,7 @@
 #include "log.h"
 #include "ngram-cache.h"
 #include "ngram-map.h"
+#include "ngram-mod.h"
 #include "sampling.h"
 
 #include <algorithm>
@@ -23,6 +24,7 @@ const std::vector<enum common_speculative_type> common_speculative_types = {
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MOD,
     COMMON_SPECULATIVE_TYPE_NGRAM_CACHE
 };
 
@@ -33,6 +35,7 @@ const std::map<std::string, enum common_speculative_type> common_speculative_typ
     {"ngram_simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
     {"ngram_map_k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
     {"ngram_map_k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
+    {"ngram_mod",     COMMON_SPECULATIVE_TYPE_NGRAM_MOD},
     {"ngram_cache",   COMMON_SPECULATIVE_TYPE_NGRAM_CACHE}
 };
 
@@ -110,6 +113,8 @@ 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.
@@ -119,7 +124,9 @@ struct common_speculative_state {
     // TODO: track performance of most recent calls
     const bool gen_perf = true; // whether to generate performance stats.
 
-    int64_t gen_duration_us = 0; // total time spent in this implementation in microseconds.
+    int64_t t_begin_us  = 0; // total time spent in refresh of this implementation in microseconds.
+    int64_t t_draft_us  = 0; // total time spent in generating drafts in this implementation in microseconds.
+    int64_t t_accept_us = 0; // total time spent in accumulation of this implementation in microseconds.
 
     common_speculative_state(enum common_speculative_type type) : type(type) {}
 
@@ -456,12 +463,14 @@ 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;
+
+    uint16_t check_id = 0; // used to control the frequency of generating drafts
 
     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);
@@ -472,7 +481,13 @@ 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);
+        ++check_id;
+        if (check_id < config.check_rate) {
+            return;
+        }
+        check_id = 0;
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
         GGML_UNUSED(params);
     }
 
@@ -492,7 +507,7 @@ struct common_speculative_state_ngram_map_k : public common_speculative_state {
         : common_speculative_state(type), map(std::move(map)) {}
 
     void begin(const llama_tokens & prompt) override {
-        GGML_UNUSED(prompt);
+        common_ngram_map_begin(map, prompt);
     }
 
     void draft(
@@ -509,6 +524,132 @@ struct common_speculative_state_ngram_map_k : public common_speculative_state {
     }
 };
 
+struct common_speculative_state_ngram_mod : public common_speculative_state {
+    common_ngram_mod & mod;
+
+    // the last position in the prompt that was added to the ngram container
+    size_t i_last = 0;
+
+    // length of the last drafted n‑gram (number of tokens returned by draft)
+    size_t n_draft_last = 0;
+
+    // consecutive accept rounds with low acceptance fraction (< 0.5)
+    int n_low = 0;
+
+    // enable trace logging if LLAMA_TRACE is set
+    const bool verbose;
+
+    common_speculative_state_ngram_mod(enum common_speculative_type type, common_ngram_mod & mod)
+        : common_speculative_state(type), mod(mod), verbose(std::getenv("LLAMA_TRACE") != nullptr) {
+        static_assert(sizeof(llama_token) == sizeof(common_ngram_mod::entry_t));
+    }
+
+    void begin(const llama_tokens & prompt) override {
+        i_last = 0;
+
+        n_draft_last = 0;
+
+        const size_t n = mod.get_n();
+
+        if (prompt.size() < n) {
+            return;
+        }
+
+        for (size_t i = 0; i < prompt.size() - n; ++i) {
+            mod.add(prompt.data() + i);
+        }
+
+        i_last = prompt.size() - n;
+
+        const double f = (double)mod.get_used() / (double)mod.size();
+        LOG_INF("%s: ngram_mod occupancy = %zu/%zu (%.2f)\n", __func__, mod.get_used(), mod.size(), f);
+
+        constexpr double f_thold = 0.25;
+        if (f > f_thold) {
+            LOG_WRN("%s: ngram_mod occupancy %.2f exceeds threshold (%.2f) - resetting\n", __func__, f, f_thold);
+
+            mod.reset();
+        }
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        GGML_UNUSED(params);
+
+        n_draft_last = 0;
+
+        const size_t cur_len = prompt_tgt.size();
+        if (cur_len < mod.get_n()) {
+            return;
+        }
+
+        const size_t n = mod.get_n();
+
+        // add new ngrams in chunks
+        if (i_last + 32 < cur_len) {
+            for (size_t i = i_last; i < cur_len - n; ++i) {
+                mod.add(prompt_tgt.data() + i);
+            }
+
+            i_last = cur_len - n;
+        }
+
+        result.resize(n + params.n_max);
+        for (size_t i = 0; i < n - 1; ++i) {
+            result[i] = prompt_tgt[cur_len - n + 1 + i];
+        }
+        result[n - 1] = id_last;
+
+        for (int i = 0; i < params.n_max; ++i) {
+            const llama_token token = mod.get(result.data() + i);
+            if (token == common_ngram_mod::EMPTY) {
+                if (i < params.n_min) {
+                    result.clear();
+                    return;
+                }
+
+                result.resize(n + i);
+                break;
+            }
+            result[n + i] = token;
+        }
+
+        // only return the m tokens that were drafted
+        for (size_t i = 0; n + i < result.size(); ++i) {
+            result[i] = result[n + i];
+        }
+        result.resize(result.size() - n);
+
+        // store length of drafted n‑gram for later acceptance analysis
+        n_draft_last = result.size();
+    }
+
+    void accept(uint16_t n_accepted) override {
+        if (verbose) {
+            LOG_INF("%s: accepted %d tokens from %zu drafted tokens\n", __func__, n_accepted, n_draft_last);
+        }
+
+        // compute acceptance fraction if we have a recorded draft length
+        if (n_draft_last > 0) {
+            const double f_acc = (double)n_accepted / (double)n_draft_last;
+            if (f_acc < 0.5) {
+                n_low++;
+                if (n_low >= 3) {
+                    LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, n_low);
+
+                    mod.reset();
+                    n_low = 0;
+                }
+            } else {
+                n_low = 0;
+            }
+        }
+    }
+};
+
 struct common_speculative_state_ngram_cache : public common_speculative_state {
     uint16_t n_draft;
     bool save_dynamic;
@@ -650,6 +791,7 @@ std::string common_speculative_type_to_str(enum common_speculative_type type) {
         case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram_simple";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram_map_k";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram_map_k4v";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MOD:     return "ngram_mod";
         case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:   return "ngram_cache";
         default:                                    return "unknown";
     }
@@ -663,11 +805,47 @@ 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(
-        const common_params_speculative & params,
-              llama_context             * ctx_tgt) {
+        common_params_speculative & params,
+        llama_context             * ctx_tgt) {
     llama_context * ctx_dft = nullptr;
     if (params.model_dft) {
         ctx_dft = llama_init_from_model(params.model_dft, params.cparams_dft);
@@ -687,6 +865,7 @@ common_speculative * common_speculative_init(
         bool has_ngram_simple  = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
         bool has_ngram_map_k   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
         bool has_ngram_map_k4v = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        bool has_ngram_mod     = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
 
         // In a more complex implementation we could use the same implementation but with different parameters.
         // This was initially used in PR-18471 but removed to simplify the code.
@@ -701,6 +880,22 @@ common_speculative * common_speculative_init(
             // This implementation can guess tokens with high acceptance rate but is more expensive.
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
         }
+        if (has_ngram_mod) {
+            // shared instance for all speculative decoding contexts
+            if (!params.ngram_mod) {
+                params.ngram_mod = std::make_shared<common_ngram_mod>(params.ngram_size_n, 4*1024*1024);
+
+                LOG_INF("%s: initialized ngram_mod with n=%d, size=%zu (%.3f MB)\n", __func__,
+                        params.ngram_size_n, params.ngram_mod->size(),
+                        (float)(params.ngram_mod->size_bytes())/1024/1024);
+
+                if (params.ngram_size_n < 16) {
+                    LOG_WRN("%s: ngram_mod n=%d is too small - poor quality is possible, see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, params.ngram_size_n);
+                }
+            }
+
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, params));
+        }
         if (has_ngram_cache) {
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
         }
@@ -738,14 +933,14 @@ common_speculative * common_speculative_init(
                 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
                 };
                 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;
@@ -758,6 +953,11 @@ common_speculative * common_speculative_init(
                 ));
                 break;
             }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
+                GGML_ASSERT(config.params.ngram_mod);
+                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod));
+                break;
+            }
             case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
                 auto state = create_state_ngram_cache(
                         params.lookup_cache_static, params.lookup_cache_dynamic, config);
@@ -795,6 +995,7 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
     }
 
     for (auto & impl : spec->impls) {
+        common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
         impl->begin(prompt);
     }
 }
@@ -810,20 +1011,14 @@ 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->gen_duration_us += t_now_us - t_start_us; // accumulate duration for this implementation
         }
 
         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(),
+                    common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
                     impl.get()->drafts_call_count, result.size());
 
             spec->curr_impl = impl.get(); // set current implementation for stats
@@ -846,12 +1041,15 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
 
     GGML_ASSERT(impl);
 
-    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->drafts_accepted_count++;
+            impl->drafts_accepted_tokens += n_accepted;
+        }
 
-    impl->accept(n_accepted);
+        impl->accept(n_accepted);
+    }
 }
 
 void common_speculative_print_stats(const common_speculative * spec) {
@@ -863,8 +1061,10 @@ void common_speculative_print_stats(const common_speculative * spec) {
         std::string str_perf;
         if (impl->gen_perf) {
             std::ostringstream oss;
-            oss << std::fixed << std::setprecision(3) << impl->gen_duration_us / 1000.0;
-            str_perf = ", dur = " + oss.str() + " ms";
+            oss << std::fixed << std::setprecision(3) << impl->t_begin_us / 1000.0 << ", ";
+            oss << std::fixed << std::setprecision(3) << impl->t_draft_us / 1000.0 << ", ";
+            oss << std::fixed << std::setprecision(3) << impl->t_accept_us / 1000.0;
+            str_perf = ", dur(b,g,a) = " + oss.str() + " ms";
         } else {
             str_perf = "";
         }

common/speculative.h

@@ -14,9 +14,13 @@ 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(
-        const common_params_speculative & params,
-              llama_context             * ctx_tgt);
+        common_params_speculative & params,
+        llama_context             * ctx_tgt);
 
 void common_speculative_free(common_speculative * spec);
 

convert_hf_to_gguf.py

@@ -586,6 +586,10 @@ class ModelBase:
                             gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
                             gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
                             gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
+                            # Kimi KDA conv weights should be F32
+                            gguf.MODEL_TENSOR.SSM_CONV1D_Q,
+                            gguf.MODEL_TENSOR.SSM_CONV1D_K,
+                            gguf.MODEL_TENSOR.SSM_CONV1D_V,
                         )
                     )
                     or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
@@ -903,10 +907,10 @@ class TextModel(ModelBase):
         if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
             logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
-        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+        if (n_experts := self.find_hparam(["num_local_experts", "num_experts"], optional=True)) is not None:
             self.gguf_writer.add_expert_count(n_experts)
             logger.info(f"gguf: expert count = {n_experts}")
-        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+        if (n_experts_used := self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
             logger.info(f"gguf: experts used count = {n_experts_used}")
         if (n_expert_groups := self.hparams.get("n_group")) is not None:
@@ -916,7 +920,7 @@ class TextModel(ModelBase):
             self.gguf_writer.add_expert_group_used_count(n_group_used)
             logger.info(f"gguf: expert groups used count = {n_group_used}")
 
-        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
+        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None:
             if score_func == "sigmoid":
                 self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
             elif score_func == "softmax":
@@ -4098,39 +4102,27 @@ class Qwen2MoeModel(TextModel):
         # process the experts separately
         name = name.replace("language_model.", "") # InternVL
 
-        # handle aggregated expert tensors
-        # GGUF stores dimensions reversed from PyTorch, so:
-        # PyTorch (A,B,C) -> GGUF writes [C,B,A] -> GGML reads ne={C,B,A}
-        # Input shapes from HF: (n_expert, n_ff_exp, n_embd) or (n_expert, n_embd, n_ff_exp)
-        # Expected GGML ne: {n_embd, n_ff_exp, n_expert} for gate/up, {n_ff_exp, n_embd, n_expert} for down
+        # handle pre-packed expert tensors (e.g. Qwen3.5 MoE, Qwen3Next)
+        # HF stores these using nn.Linear convention: [n_expert, out_features, in_features]
+        # This matches the individual expert stacking path below (which stacks
+        # per-expert [out, in] weights into [n_expert, out, in]), so no permute is needed.
         if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
             mapped = f"{name}.weight" if not name.endswith(".weight") else name
-            # Input: (n_expert=128, n_ff_exp=768, n_embd=2048)
-            # Want GGML ne: {n_ff_exp, n_embd, n_expert} = {768, 2048, 128}
-            # Need PyTorch: (128, 2048, 768) [reversed of GGML]
-            # So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
-            permuted = data_torch.permute(0, 2, 1).contiguous()
-            yield from super().modify_tensors(permuted, mapped, bid)
+            # HF: [n_expert, n_embd, n_ff] → GGML: {n_ff, n_embd, n_expert} ✓
+            yield from super().modify_tensors(data_torch, mapped, bid)
             return
 
         if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
-            if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
-                raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
-            split_dim = data_torch.shape[-1] // 2
-            gate = data_torch[..., :split_dim].contiguous()
-            up = data_torch[..., split_dim:].contiguous()
-            # Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
-            # Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
-            # Need PyTorch: (128, 768, 2048) [reversed of GGML]
-            # So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
-            base_name = name.removesuffix(".weight")
-            base = base_name.rsplit('.', 1)[0]
-            mapped_gate = f"{base}.gate_proj.weight"
-            mapped_up = f"{base}.up_proj.weight"
-            perm_gate = gate.permute(0, 2, 1).contiguous()
-            perm_up = up.permute(0, 2, 1).contiguous()
-            yield from super().modify_tensors(perm_gate, mapped_gate, bid)
-            yield from super().modify_tensors(perm_up, mapped_up, bid)
+            # HF: [n_expert, 2*n_ff, n_embd] → split on dim=1
+            n_ff = data_torch.shape[1] // 2
+            gate = data_torch[:, :n_ff, :].contiguous()
+            up = data_torch[:, n_ff:, :].contiguous()
+            # gate/up: [n_expert, n_ff, n_embd] → GGML: {n_embd, n_ff, n_expert} ✓
+            base_name = name.removesuffix(".weight").removesuffix(".gate_up_proj")
+            mapped_gate = f"{base_name}.gate_proj.weight"
+            mapped_up = f"{base_name}.up_proj.weight"
+            yield from super().modify_tensors(gate, mapped_gate, bid)
+            yield from super().modify_tensors(up, mapped_up, bid)
             return
 
         if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
@@ -4340,6 +4332,40 @@ class Qwen3NextModel(Qwen2MoeModel):
             yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Qwen3_5ForCausalLM", "Qwen3_5TextForCausalLM")
+class Qwen3_5Model(Qwen3NextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3_5
+
+    # Stores whichever of in_proj_a/in_proj_b is seen first, keyed by layer
+    _pending_ba: dict[int | None, tuple[str, Tensor]] = {}
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Handle split in_proj_b + in_proj_a → concatenated SSM_BETA_ALPHA
+        # safetensors sorts alphabetically so in_proj_a arrives before in_proj_b
+        if "in_proj_a.weight" in name or "in_proj_b.weight" in name:
+            which = "a" if "in_proj_a" in name else "b"
+            if bid not in self._pending_ba:
+                self._pending_ba[bid] = (which, data_torch)
+                return
+            prev_which, prev_tensor = self._pending_ba.pop(bid)
+            assert prev_which != which, f"duplicate in_proj_{which} for layer {bid}"
+            b_tensor = prev_tensor if prev_which == "b" else data_torch
+            a_tensor = prev_tensor if prev_which == "a" else data_torch
+            ba_combined = torch.cat([b_tensor, a_tensor], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.SSM_BETA_ALPHA, bid, ".weight"), ba_combined)
+            return
+        else:
+            # Qwen3Next uses .qkvz tensor, so we use the super to get the other functionalities
+            # (norm correction, A_log to A etc.) for free
+            # Qwen2Moe already does the gate_up conversion properly, just use that
+            yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3_5MoeForCausalLM", "Qwen3_5MoeTextForCausalLM")
+class Qwen3_5MoeModel(Qwen3_5Model):
+    model_arch = gguf.MODEL_ARCH.QWEN3_5_MOE
+
+
 @ModelBase.register("RND1")
 class RND1Model(Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.RND1
@@ -5013,6 +5039,221 @@ class CodeShellModel(TextModel):
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
 
+@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
+class KimiLinearModel(TextModel):
+    """Kimi-Linear model with hybrid MLA+KDA architecture"""
+    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_gpt2()
+            return
+        except Exception:
+            pass
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        if tokpre == "kimi-k2":
+            # Build merges list using the approach similar to HunYuanMoE
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer.model._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[QwenModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+                if len(merged) == 2:
+                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+            # Build token list
+            vocab_size = self.hparams["vocab_size"]
+            special_tokens = tokenizer.special_tokens
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
+            tokens: list[str] = []
+            toktypes: list[int] = []
+
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                else:
+                    token = reverse_vocab[i]
+                    tokens.append(token)
+                    if i in special_tokens.values():
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.NORMAL)
+
+            self.gguf_writer.add_tokenizer_model("gpt2")
+            self.gguf_writer.add_tokenizer_pre(tokpre)
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_types(toktypes)
+            self.gguf_writer.add_token_merges(merges)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+            special_vocab.add_to_gguf(self.gguf_writer)
+            # override eos id in config.json with tiktoken eos id
+            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
+        else:
+            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
+
+    def set_gguf_parameters(self):
+        # note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)
+        self.hparams["num_key_value_heads"] = 1
+
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+        # KDA & MLA params
+        # Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
+        linear_attn_config = self.hparams["linear_attn_config"]
+        # n_head == 0 for KDA layers, n_head > 0 for MLA layers
+        # full_attention_layers list will be used to distingush layer type
+        _num_kv_heads = list()
+        _full_attn_layers = linear_attn_config["full_attn_layers"]
+        for il in range(self.hparams["num_hidden_layers"]):
+            if il + 1 in _full_attn_layers:
+                _num_kv_heads.append(self.hparams["num_key_value_heads"])
+            else:
+                _num_kv_heads.append(0)
+        assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]
+        self.gguf_writer.add_head_count_kv(_num_kv_heads)
+
+        if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:
+            self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)
+        if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:
+            self.gguf_writer.add_kda_head_dim(kda_head_dim)
+
+        # MLA params - use add_* methods that handle arch substitution
+        # Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)
+        if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:
+            self.gguf_writer.add_q_lora_rank(q_lora_rank)
+        # To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA
+        kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)
+        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
+
+        # MLA head dimensions
+        # Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim
+        qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")
+        # Rotation - use qk_rope_head_dim for Kimi
+        qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)
+        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
+        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
+        v_head_dim = self.hparams.get("v_head_dim")
+
+        # Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
+        if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:
+            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
+        elif qk_nope_head_dim is not None:
+            n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
+            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
+
+        # n_embd_head_v_mla = v_head_dim
+        if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:
+            self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)
+        elif v_head_dim is not None:
+            self.gguf_writer.add_value_length_mla(v_head_dim)
+
+        # moe_intermediate_size (1024 for Kimi)
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        # num_shared_experts (1 for Kimi)
+        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
+        # first_k_dense_replace (1 for Kimi - first layer uses dense MLP)
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        # Routed scaling factor (expert_weights_scale = 2.446 for Kimi)
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")
+
+        # Handle KDA conv1d weights
+        # HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest
+        # llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest
+        # GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
+        # Memory layouts match: both have conv_step (d_conv) changing fastest
+        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
+            # HF shape: [d_inner, d_conv] e.g. [4096, 4]
+            # Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
+            if data_torch.ndim == 2:
+                d_inner, d_conv = data_torch.shape
+                # Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)
+                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
+            elif data_torch.ndim == 3:
+                # Already 3D [d_inner, 1, d_conv] from unsqueeze
+                d_inner, _, d_conv = data_torch.shape
+                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, 1, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
+
+        # Kimi specific bias
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # Handle A_log: iHF stores as [1, 1, num_heads, 1]
+        # llama.cpp expects ggml ne = [1, num_heads, 1, 1]
+        # GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]
+        if name.endswith(".A_log"):
+            data_torch = -torch.exp(data_torch)
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+            logger.info("Changed dt_bias to dt_proj.bias")
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                # w1: gate, w2: down, w3: up
+                for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),
+                                   ("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),
+                                   ("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:
+                    datas: list[Tensor] = []
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+                    data_torch = torch.stack(datas, dim=0)
+                    new_name = self.format_tensor_name(tname, bid)
+                    yield from super().modify_tensors(data_torch, new_name, bid)
+            return
+
+        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
+        if name.endswith("kv_b_proj.weight"):
+            name_kb = name.replace("kv_b_proj", "k_b_proj")
+            name_vb = name.replace("kv_b_proj", "v_b_proj")
+            n_head_kv = self.hparams["num_key_value_heads"]
+            v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)
+            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
+            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+            yield from super().modify_tensors(k_b, name_kb, bid)
+            yield from super().modify_tensors(v_b, name_vb, bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("InternLM2ForCausalLM")
 class InternLM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.INTERNLM2
@@ -7693,6 +7934,135 @@ class MimoV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Step3p5ForCausalLM")
+class Step35Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.STEP35
+
+    def set_gguf_parameters(self):
+        rope_theta = self.hparams.get("rope_theta")
+        if isinstance(rope_theta, list):
+            self.hparams["rope_theta"] = float(rope_theta[0])
+            self.hparams["local_rope_theta"] = float(rope_theta[1])
+            self.rope_parameters["rope_theta"] = self.hparams["rope_theta"]
+            self.rope_parameters["sliding_attention"] = {"rope_theta": self.hparams["local_rope_theta"]}
+
+        super().set_gguf_parameters()
+
+        layer_types = self.hparams.get("layer_types") or []
+        partial_rotary_factors = self.hparams.get("partial_rotary_factors") or []
+        attn_other = self.hparams.get("attention_other_setting") or {}
+
+        n_head_base = self.hparams["num_attention_heads"]
+        n_kv_base = self.hparams["num_attention_groups"]
+
+        n_head_swa = attn_other.get("num_attention_heads", n_head_base)
+        n_kv_swa = attn_other.get("num_attention_groups", n_kv_base)
+
+        layer_types = layer_types[: self.block_count]
+        partial_rotary_factors = partial_rotary_factors[: self.block_count]
+        assert [1.0 if lt == "sliding_attention" else 0.5 for lt in layer_types] == partial_rotary_factors
+        head_arr = [n_head_swa if lt == "sliding_attention" else n_head_base for lt in layer_types]
+        kv_arr = [n_kv_swa if lt == "sliding_attention" else n_kv_base for lt in layer_types]
+        swa_pat = [lt == "sliding_attention" for lt in layer_types]
+
+        self.gguf_writer.add_head_count(head_arr)
+        self.gguf_writer.add_head_count_kv(kv_arr)
+
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern(swa_pat)
+
+        self.gguf_writer.add_value_length(self.hparams["head_dim"])
+
+        # MoE params
+        self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["share_expert_dim"])
+
+        if (moe_router_scaling_factor := self.hparams.get("moe_router_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(moe_router_scaling_factor)
+        if (norm_expert_weight := self.hparams.get("norm_expert_weight")) is not None:
+            self.gguf_writer.add_expert_weights_norm(norm_expert_weight)
+
+        # leading dense blocks
+        leading_dense = 0
+        moe_layers_enum = self.hparams.get("moe_layers_enum")
+        if isinstance(moe_layers_enum, str) and moe_layers_enum.strip():
+            moe_layers = sorted(int(i) for i in moe_layers_enum.strip().split(","))
+            if moe_layers:
+                leading_dense = max(0, moe_layers[0])
+        self.gguf_writer.add_leading_dense_block_count(leading_dense)
+        self.gguf_writer.add_moe_every_n_layers(int(self.hparams.get("moe_every_n_layer", 1)))
+
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-5))
+
+        # Optional per-layer SwiGLU clamps.
+        if (limits := self.hparams.get("swiglu_limits")) is not None:
+            limits_f = [0.0 if v is None else float(v) for v in limits[: self.block_count]]
+            self.gguf_writer.add_swiglu_clamp_exp(limits_f)
+        if (limits_shared := self.hparams.get("swiglu_limits_shared")) is not None:
+            limits_shared_f = [0.0 if v is None else float(v) for v in limits_shared[: self.block_count]]
+            self.gguf_writer.add_swiglu_clamp_shexp(limits_shared_f)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # remove mtp layers
+        if (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None:
+            il = int(m.group(1))
+            n_main = int(self.hparams.get("num_hidden_layers", self.block_count))
+            if il >= n_main:
+                return
+        if name.endswith("norm.weight"):
+            data_torch += 1.0
+        # Map router bias (expert selection bias) to a GGUF bias tensor
+        if name.endswith(".moe.router_bias"):
+            name += ".bias"
+
+        if name.endswith((".self_attn.g_proj.weight", ".moe.gate.weight", ".moe.up_proj.weight", ".moe.gate_proj.weight", ".moe.down_proj.weight")):
+            data_torch = data_torch.squeeze().contiguous()
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # Step35 can optionally use Llama-3 style RoPE scaling (HF: rope_scaling.rope_type == "llama3").
+        # llama.cpp represents this via a single extra tensor: "rope_freqs.weight" (aka MODEL_TENSOR.ROPE_FREQS).
+        rope_params = self.rope_parameters.get("full_attention", self.rope_parameters)
+        rope_type = rope_params.get("rope_type") or ""
+        if rope_type.lower() != "llama3":
+            return
+
+        # Step35 configs can carry per-layer rope_theta as a list; for llama3 rope factors we use the base value.
+        rope_theta = self.hparams.get("rope_theta", 10000.0)
+        if isinstance(rope_theta, list):
+            rope_theta = rope_theta[0]
+        base = float(rope_theta)
+        if (dim := self.hparams.get("head_dim")) is None:
+            dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+        dim = int(dim)
+
+        freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+        factor = float(rope_params.get("factor", 8.0))
+        low_freq_factor = float(rope_params.get("low_freq_factor", 1.0))
+        high_freq_factor = float(rope_params.get("high_freq_factor", 4.0))
+        old_context_len = int(rope_params.get("original_max_position_embeddings", self.hparams.get("original_max_position_embeddings", 8192)))
+
+        low_freq_wavelen = old_context_len / low_freq_factor
+        high_freq_wavelen = old_context_len / high_freq_factor
+
+        rope_factors: list[float] = []
+        for freq in freqs:
+            wavelen = 2 * math.pi / float(freq)
+            if wavelen < high_freq_wavelen:
+                rope_factors.append(1.0)
+            elif wavelen > low_freq_wavelen:
+                rope_factors.append(factor)
+            else:
+                smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                rope_factors.append(1.0 / ((1.0 - smooth) / factor + smooth))
+
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
+
+
 @ModelBase.register("PanguEmbeddedForCausalLM")
 class PanguEmbeddedModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PANGU_EMBED
@@ -8806,6 +9176,7 @@ class GraniteMoeModel(GraniteModel):
             gate, up = data_torch.split(ffn_dim, dim=-2)
             yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
             yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid)
+            return
 
         has_experts = bool(self.hparams.get('num_local_experts'))
 

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.
 
@@ -111,15 +113,15 @@ 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  |
 
 *Notes:*
 
 - **Memory**
-  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-cli`.
+  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/llama-completion`.
   - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
 
 - **Execution Unit (EU)**
@@ -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
@@ -423,16 +327,12 @@ Choose one of following methods to run.
 - Use device 0:
 
 ```sh
-./examples/sycl/run-llama2.sh 0
-# OR
-./examples/sycl/run-llama3.sh 0
+./examples/sycl/test.sh -mg 0
 ```
 - Use multiple devices:
 
 ```sh
-./examples/sycl/run-llama2.sh
-# OR
-./examples/sycl/run-llama3.sh
+./examples/sycl/test.sh
 ```
 
 2. Command line
@@ -455,13 +355,13 @@ Examples:
 - Use device 0:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm none -mg 0
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm none -mg 0 --mmap
 ```
 
 - Use multiple devices:
 
 ```sh
-ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer
+ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 99 -sm layer --mmap
 ```
 
 *Notes:*
@@ -577,13 +477,13 @@ Or, use CMake presets to build:
 
 ```sh
 cmake --preset x64-windows-sycl-release
-cmake --build build-x64-windows-sycl-release -j --target llama-cli
+cmake --build build-x64-windows-sycl-release -j --target llama-completion
 
 cmake -DGGML_SYCL_F16=ON --preset x64-windows-sycl-release
-cmake --build build-x64-windows-sycl-release -j --target llama-cli
+cmake --build build-x64-windows-sycl-release -j --target llama-completion
 
 cmake --preset x64-windows-sycl-debug
-cmake --build build-x64-windows-sycl-debug -j --target llama-cli
+cmake --build build-x64-windows-sycl-debug -j --target llama-completion
 ```
 
 #### 3. Visual Studio
@@ -608,7 +508,7 @@ You can use Visual Studio to open the `llama.cpp` folder directly as a CMake pro
 - For a minimal experimental setup, you can build only the inference executable using:
 
     ```Powershell
-    cmake --build build --config Release -j --target llama-cli
+    cmake --build build --config Release -j --target llama-completion
     ```
 
 ##### - Generating a Visual Studio Solution
@@ -714,13 +614,7 @@ Choose one of following methods to run.
 1. Script
 
 ```
-examples\sycl\win-run-llama-2.bat
-```
-
-or
-
-```
-examples\sycl\win-run-llama-3.bat
+examples\sycl\win-test.bat
 ```
 
 2. Command line
@@ -744,13 +638,13 @@ Examples:
 - Use device 0:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm none -mg 0
+build\bin\llama-completion.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm none -mg 0 --mmap
 ```
 
 - Use multiple devices:
 
 ```
-build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer
+build\bin\llama-completion.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 99 -sm layer --mmap
 ```
 
 
@@ -776,15 +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
 
@@ -792,7 +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/snapdragon/CMakeUserPresets.json

@@ -1,10 +1,5 @@
 {
   "version": 5,
-  "cmakeMinimumRequired": {
-      "major": 3,
-      "minor": 28,
-      "patch": 0
-  },
   "configurePresets": [
     {
         "name": "arm64-android-snapdragon",

docs/backend/snapdragon/windows.md

@@ -128,7 +128,7 @@ However, additional settings are required for generating and signing HTP Ops lib
 > $env:HEXAGON_HTP_CERT="c:\Users\MyUsers\Certs\ggml-htp-v1.pfx"
 > $env:WINDOWS_SDK_BIN="C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0\arm64"
 
-> cmake --preset arm64-windows-snapdragon -B build-wos
+> cmake --preset arm64-windows-snapdragon-release -B build-wos
 ...
 > cmake --install build-wos --prefix pkg-snapdragon
 ```

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/multimodal/minicpmo2.6.md

@@ -9,7 +9,7 @@ Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch m
 ### Build llama.cpp
 Readme modification time: 20250206
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash

docs/multimodal/minicpmo4.0.md

@@ -8,11 +8,11 @@ Download [MiniCPM-o-4](https://huggingface.co/openbmb/MiniCPM-o-4) PyTorch model
 ### Build llama.cpp
 Readme modification time: 20250206
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 

docs/multimodal/minicpmv2.5.md

@@ -8,7 +8,7 @@ Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-
 ### Build llama.cpp
 Readme modification time: 20250206
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash

docs/multimodal/minicpmv2.6.md

@@ -8,7 +8,7 @@ Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch m
 ### Build llama.cpp
 Readme modification time: 20250206
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash

docs/multimodal/minicpmv4.0.md

@@ -8,11 +8,11 @@ Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model
 ### Build llama.cpp
 Readme modification time: 20250731
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 

docs/multimodal/minicpmv4.5.md

@@ -8,11 +8,11 @@ Download [MiniCPM-V-4_5](https://huggingface.co/openbmb/MiniCPM-V-4_5) PyTorch m
 ### Build llama.cpp
 Readme modification time: 20250826
 
-If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 

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

@@ -6,7 +6,7 @@ llama.cpp supports speculative decoding, a technique that can significantly acce
 
 ## Implementations
 
-The `llama-server` application supports several implementations of speculative decoding:
+The `llama-server` application supports several implementations of speculative decoding. An implementation with draft model can be mixed with an implementation without draft model.
 
 ### Draft Model (`draft`)
 
@@ -32,12 +32,21 @@ An example to use this approach can be the rewriting of source code by a LLM.
 
 This implementation looks for the last n-gram in history that matches the current n-gram and creates a draft using the m tokens following the matched n-gram. It is the simplest self-speculative approach with minimal overhead.
 
+```
+llama-server [...] --spec-type ngram-simple --draft-max 64
+```
+
 #### n-gram Map Key (`ngram-map-k`)
 
-This implementation looks for the current n-gram of size n (called the _key_) in the token history. If the key n-gram is followed by the same m tokens (called the _mgram_) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument `--spec-ngram-min-hits`) before generating drafts.
+This implementation looks for the current n-gram of size n (called the _key_) in the token history. If the key n-gram is followed by the same m tokens (called the _mgram_) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument `--spec-ngram-min-hits`, default is 1) before generating drafts.
 
 The number of accepted tokens is stored for each used n-gram.
 
+**Example:**
+```
+llama-server [...] --spec-type ngram-map-k --draft-max 64
+```
+
 #### n-gram Map Key-4-Values (`ngram-map-k4v`)
 
 This experimental implementation looks for the current n-gram of size n (called the _key_) in the token history. For each key, up to four _values_ (n-grams of size m, called _mgrams_) are tracked. An internal statistic counts the occurrences of each mgram after the key n-gram. If one mgram is significantly more frequent than the others, it is used as the draft.
@@ -45,17 +54,65 @@ This experimental implementation looks for the current n-gram of size n (called
 The number of accepted tokens is stored for each used n-gram.
 
 **Example:** Server options to be used if there are a lot of longer repetitions.
-```bash
-llama-server [...] --spec-type ngram-map-k4v --spec-ngram-size-n 8 --spec-ngram-size-m 8 --spec-ngram-min-hits 2
+```
+llama-server [...] --spec-type ngram-map-k4v --spec-ngram-size-n 8 --spec-ngram-size-m 8 --spec-ngram-min-hits 2 --draft-max 64
 ```
 
+### n-gram Mod (`ngram-mod`)
+
+Add basic ngram hasher for speculative decoding:
+
+- For each ngram, compute a hash using LCG
+- For each computed hash, store the next token
+- During speculation, iteratively compute the rolling hash of the last n tokens and pick the next token from the storage
+
+Some characteristics:
+
+- Lightweight (~16 MB)
+- Constant memory and complexity
+- Can generate variable draft lengths (i.e. m is not fixed)
+
+Currently, a single hash pool is shared across all server slots, so different requests can benefit from each other.
+
+**Sample usage:**
+
+```
+# notes:
+# - small `n` are not recommended
+# - MoEs require long drafts
+# - dense models: can reduce `--draft-min` and `--draft-max`
+
+llama-server ... --spec-type ngram-mod --spec-ngram-size-n 24 --draft-min 48 --draft-max 64
+```
+
+Applications:
+
+- Iterating over a block of text/code (e.g. in llama.vim)
+- Reasoning models (when they have to repeat their thinking in the final answer)
+- Summarization
+
+Example Video:
+
+- See #19164
+
+### Differences between ngram-simple, ngram-map and ngram-mod
+
+- ngram-simple looks for a previous matching n-gram and inserts the following m-gram.
+- ngram-map-k looks for a previous matching n-gram and inserts the following m-gram but uses an internal hash-map of n-grams in the current context window.
+- ngram-mod uses a hash pool which is shared across all server slots. The hash pool is a map from n-gram hash to the next token (not the next m-gram as in ngram-map).
 
 ## Command-Line Options
 
 If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.
 
 ```
---spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v]
+--draft, --draft-n, --draft-max N       number of tokens to draft for speculative decoding (default: 16)
+                                        (env: LLAMA_ARG_DRAFT_MAX)
+--draft-min, --draft-n-min N            minimum number of draft tokens to use for speculative decoding
+                                        (default: 0)
+                                        (env: LLAMA_ARG_DRAFT_MIN)
+[...]
+--spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
                                         type of speculative decoding to use when no draft model is provided
                                         (default: none)
 --spec-ngram-size-n N                   ngram size N for ngram-simple/ngram-map speculative decoding, length
@@ -78,6 +135,7 @@ Specifies a type of speculative decoding without draft model.
 | `ngram-simple` | Use simple n-gram pattern matching |
 | `ngram-map-k` | Use n-gram pattern matching with n-gram-keys |
 | `ngram-map-k4v` | Use n-gram pattern matching with n-gram-keys and up to four m-gram values (experimental) |
+| `ngram-mod` | Use basic ngram hasher for speculative decoding with shared pool |
 
 **Example:** Server-instance used to refactor source code.
 ```bash
@@ -112,9 +170,15 @@ statistics ngram_simple: #calls = 15, #gen drafts = 5, #acc drafts = 5, #gen tok
 statistics draft: #calls = 10, #gen drafts = 10, #acc drafts = 10, #gen tokens = 110, #acc tokens = 98
 ```
 
+```
+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
 - `#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)
 - `#acc tokens`: number of tokens accepted by the main model
+- `dur(b,g,a): durations of begin (new prompt), generation and accumulation (process acceptance).
 

examples/deprecation-warning/README.md

@@ -1,7 +1,7 @@
 # Migration notice for binary filenames
 
 > [!IMPORTANT]
-[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
+[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggml-org/llama.cpp/pull/7809)
 
 This migration was important, but it is a breaking change that may not always be immediately obvious to users.
 

examples/deprecation-warning/deprecation-warning.cpp

@@ -28,7 +28,7 @@ int main(int argc, char** argv) {
     fprintf(stdout, "\n");
     fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
     fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str());
-    fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
+    fprintf(stdout, " See https://github.com/ggml-org/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
     fprintf(stdout, "\n");
 
     return EXIT_FAILURE;

examples/json_schema_to_grammar.py

@@ -402,7 +402,7 @@ class SchemaConverter:
             Transforms a regular expression pattern into a GBNF rule.
 
             Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
-            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
+            Output: https://github.com/ggml-org/llama.cpp/blob/master/grammars/README.md
 
             Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
 

examples/lookahead/lookahead.cpp

@@ -50,6 +50,12 @@ int main(int argc, char ** argv) {
     const int N = 5;  // n-gram size
     const int G = 15; // max verification n-grams
 
+    // lookahead requires W + G + 1 sequences for parallel Jacobi decoding
+    params.n_parallel = W + G + 1;
+
+    // unified KV cache is required for coupled sequences in batch splitting
+    params.kv_unified = true;
+
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -115,7 +121,7 @@ int main(int argc, char ** argv) {
     // seq_id == 0           : the current input token
     // seq_id [1, W]         : tokens from the past N - 1 Jacobi iterations
     // seq_id [W + 1, W + G] : verification n-grams
-    llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
+    llama_batch batch = llama_batch_init(llama_n_ctx(ctx), 0, W + G + 1);
 
     // target model sampling context
     struct common_sampler * smpl = common_sampler_init(model, params.sampling);

examples/lookup/lookup.cpp

@@ -106,7 +106,7 @@ int main(int argc, char ** argv){
 
     std::vector<llama_token> draft;
 
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+    llama_batch batch_tgt = llama_batch_init(llama_n_ctx(ctx), 0, 1);
 
     const auto t_dec_start = ggml_time_us();
 

examples/model-conversion/Makefile

@@ -33,11 +33,14 @@ DEVICE ?= auto
 causal-convert-model-bf16: OUTTYPE=bf16
 causal-convert-model-bf16: causal-convert-model
 
+causal-convert-model-debug: DEBUG=--debug
+causal-convert-model-debug: causal-convert-model
+
 causal-convert-model:
 	$(call validate_model_path,causal-convert-model)
 	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
 	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
-	./scripts/causal/convert-model.sh
+	./scripts/causal/convert-model.sh $(DEBUG)
 
 causal-convert-mm-model-bf16: OUTTYPE=bf16
 causal-convert-mm-model-bf16: MM_OUTTYPE=f16

examples/model-conversion/scripts/causal/convert-model.sh

@@ -4,12 +4,17 @@ set -e
 
 # Parse command line arguments
 MMPROJ=""
+DEBUG=""
 while [[ $# -gt 0 ]]; do
     case $1 in
         --mmproj)
             MMPROJ="--mmproj"
             shift
             ;;
+        --debug)
+            DEBUG="1"
+            shift
+            ;;
         *)
             shift
             ;;
@@ -28,7 +33,12 @@ echo "Data  type: ${TYPE}"
 echo "Converted model path:: ${CONVERTED_MODEL}"
 echo "Metadata override: ${METADATA_OVERRIDE}"
 
-CMD_ARGS=("python" "../../convert_hf_to_gguf.py" "--verbose")
+if [[ -n "$DEBUG" ]]; then
+    CMD_ARGS=("python" "-m" "pdb")
+else
+    CMD_ARGS=("python")
+fi
+CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
 CMD_ARGS+=("--outtype" "${TYPE}")

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()

examples/sycl/run-llama2.sh

@@ -18,13 +18,14 @@ CONTEXT=4096
 #support malloc device memory more than 4GB.
 export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
 
+LOAD_MODE='--mmap'
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
 fi

examples/sycl/run-llama3.sh

@@ -1,31 +0,0 @@
-#!/usr/bin/env bash
-
-#  MIT license
-#  Copyright (C) 2025 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-# If you want more control, DPC++ Allows selecting a specific device through the
-# following environment variable
-export ONEAPI_DEVICE_SELECTOR="level_zero:0"
-source /opt/intel/oneapi/setvars.sh
-
-#export GGML_SYCL_DEBUG=1
-
-#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
-
-INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
-MODEL_FILE=models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf
-NGL=99 # Layers offloaded to the GPU. If the device runs out of memory, reduce this value according to the model you are using.
-CONTEXT=4096
-
-#support malloc device memory more than 4GB.
-export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-
-if [ $# -gt 0 ]; then
-    GGML_SYCL_DEVICE=$1
-    echo "Using $GGML_SYCL_DEVICE as the main GPU"
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none
-else
-    #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT}
-fi

examples/sycl/test.sh

@@ -0,0 +1,130 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+Help() {
+  cat << EOF
+Usage: $(basename "$0") [OPTIONS]
+
+This script processes files with specified options.
+
+Options:
+  -h, --help    Display this help message and exit.
+  -c, --context <value>    Set context length. Bigger need more memory.
+  -p, --promote <value>    Prompt to start generation with.
+  -m, --model   <value>    Full model file path.
+  -mg,--main-gpu <value>   Set main GPU ID (0 - n) for single GPU mode.
+  -sm,--split-mode <value> How to split the model across multiple GPUs, one of:
+                            - none: use one GPU only
+                            - layer (default): split layers and KV across GPUs
+                            - row: split rows across GPUs
+  -ngl,--n-gpu-layers <value>  Max. number of layers to store in VRAM (default: -1)
+  -lv,--log-verbosity <value>  Set the verbosity threshold. Messages with a higher verbosity will be
+                               ignored. Values:
+                                - 0: generic output
+                                - 1: error
+                                - 2: warning
+                                - 3: info
+                                - 4: debug
+
+
+EOF
+}
+
+BIN_FILE=./build/bin/llama-completion
+SEED=0
+GPUS_SETTING=""
+
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=models/llama-2-7b.Q4_0.gguf
+NGL=99
+CONTEXT=4096
+GGML_SYCL_DEVICE=-1
+SPLIT_MODE=layer
+LOG_VERBOSE=3
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        -c|--context)
+            CONTEXT=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -p|--promote)
+            # Option that is a simple flag (boolean)
+            INPUT_PROMPT="$2"
+            # Shift once to consume the option flag
+            shift
+            shift
+            ;;
+        -m|--model)
+            MODEL_FILE="$2"
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -mg|--main-gpu)
+            GGML_SYCL_DEVICE=$2
+            SPLIT_MODE=none
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -sm|--split-mode)
+            SPLIT_MODE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -ngl|--n-gpu-layers)
+            NGL=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -lv|--log-verbosity)
+            LOG_VERBOSE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -h|--help)
+            Help
+            exit 0
+            ;;
+        *)
+            # Handle unknown options or stop processing options
+            echo "Invalid option: $1"
+            # Optional: exit script or shift to treat remaining as positional args
+            exit 1
+            ;;
+    esac
+done
+
+
+
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+#support malloc device memory more than 4GB.
+export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+echo "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=${UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS}"
+
+if [ $GGML_SYCL_DEVICE -ne -1 ]; then
+    echo "Use $GGML_SYCL_DEVICE as main GPU"
+    #use signle GPU only
+    GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
+    export ONEAPI_DEVICE_SELECTOR="level_zero:${$GGML_SYCL_DEVICE}"
+    echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
+else
+   echo "Use all Intel GPUs, including iGPU & dGPU"
+ fi
+
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
+

examples/sycl/win-run-llama2.bat

@@ -7,5 +7,5 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 
 :: support malloc device memory more than 4GB.
 set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-
-.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0
+set LOAD_MODE="--mmap"
+.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0 %LOAD_MODE%

examples/sycl/win-test.bat

@@ -7,5 +7,5 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 
 :: support malloc device memory more than 4GB.
 set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-
-.\build\bin\llama-completion.exe -m models\Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf -no-cnv -p %INPUT2% -n 400 -s 0 -e -ngl 99
+set LOAD_MODE="--mmap"
+.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0 %LOAD_MODE%

ggml/CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 project("ggml" C CXX ASM)
 
 ### GGML Version

ggml/include/ggml-cann.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

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/include/ggml.h

@@ -6,7 +6,7 @@
 // This documentation is still a work in progress.
 // If you wish some specific topics to be covered, feel free to drop a comment:
 //
-//   https://github.com/ggerganov/whisper.cpp/issues/40
+//   https://github.com/ggml-org/whisper.cpp/issues/40
 //
 // ## Overview
 //

ggml/src/ggml-backend.cpp

@@ -258,6 +258,7 @@ void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor *
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
     if (backend->iface.set_tensor_async == NULL) {
+        ggml_backend_synchronize(backend);
         ggml_backend_tensor_set(tensor, data, offset, size);
     } else {
         backend->iface.set_tensor_async(backend, tensor, data, offset, size);
@@ -271,6 +272,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     if (backend->iface.get_tensor_async == NULL) {
+        ggml_backend_synchronize(backend);
         ggml_backend_tensor_get(tensor, data, offset, size);
     } else {
         backend->iface.get_tensor_async(backend, tensor, data, offset, size);

ggml/src/ggml-cann/acl_tensor.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

ggml/src/ggml-cann/acl_tensor.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

ggml/src/ggml-cann/aclnn_ops.h

@@ -1,5 +1,5 @@
 /**
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

ggml/src/ggml-cann/common.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

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

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023-2024 The ggml authors
+ * Copyright (c) 2023-2026 The ggml authors
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to

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/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;
+                        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)*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:
                     {
@@ -2929,11 +2939,12 @@ 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);
@@ -3673,6 +3684,11 @@ 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

@@ -8042,12 +8042,14 @@ 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 +8126,6 @@ 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 +8166,7 @@ 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 +8239,8 @@ 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 +8248,7 @@ 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 +8257,26 @@ 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);
+        }
     }
 }
 
@@ -8546,6 +8554,78 @@ 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 +8647,7 @@ 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 +8668,92 @@ 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 KV_TILE_SZ = ggml_fa_tile_config::KV;
+        static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+        const bool use_tiled = !use_ref &&
+                               (q->type == GGML_TYPE_F32 &&
+                                kv_is_f32_or_f16 &&
+                                k->type == v->type &&
+                                nek1 % KV_TILE_SZ == 0 &&
+                                neq1 >= Q_TILE_SZ);
+
+        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/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.

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)) {
@@ -2973,8 +2979,7 @@ 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;
     }
 
@@ -3910,14 +3915,14 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
         // Launch graph
         CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
 #else
+        GGML_UNUSED(graph_key);
         graph_evaluated_or_captured = true;
 #endif  // USE_CUDA_GRAPH
     }
 }
 
-static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
-
 #ifdef USE_CUDA_GRAPH
+static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
     ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
     if (graph->graph == nullptr) {
@@ -3930,12 +3935,8 @@ static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, co
     }
 
     return graph->is_enabled();
-#else
-    GGML_UNUSED(cuda_ctx);
-    GGML_UNUSED(graph_key);
-    return false;
-#endif // USE_CUDA_GRAPH
 }
+#endif // USE_CUDA_GRAPH
 
 static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
@@ -5053,16 +5054,6 @@ 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

@@ -3697,13 +3697,20 @@ 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/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-hexagon/CMakeLists.txt

@@ -1,8 +1,20 @@
 file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
 file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
 
-if (NOT IS_DIRECTORY "${HEXAGON_SDK_ROOT}" OR NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
-    message(FATAL_ERROR "Make sure HEXAGON_SDK_ROOT and HEXAGON_TOOLS_ROOT point to the correct Hexagon SDK installation.")
+if (NOT IS_DIRECTORY "${HEXAGON_SDK_ROOT}")
+    message(FATAL_ERROR "Make sure HEXAGON_SDK_ROOT point to the correct Hexagon SDK installation.")
+endif()
+
+if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+    message("Try to read HEXAGON_TOOLS_ROOT from hexagon_sdk.json")
+    file(READ "${HEXAGON_SDK_ROOT}/hexagon_sdk.json" HEXAGON_SDK_CONFIG_PATH)
+    string(JSON HEXAGON_TOOLS_PATH GET ${HEXAGON_SDK_CONFIG_PATH} "root" "tools" "info" 0 "path")
+    message("Found HEXAGON_TOOLS_PATH: ${HEXAGON_TOOLS_PATH}")
+    set(HEXAGON_TOOLS_ROOT "${HEXAGON_SDK_ROOT}/${HEXAGON_TOOLS_PATH}")
+    file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+    if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+        message(FATAL_ERROR "Make sure HEXAGON_TOOLS_ROOT point to the correct Hexagon SDK installation.")
+    endif()
 endif()
 
 message(STATUS "hexagon: using ${HEXAGON_SDK_ROOT} and ${HEXAGON_TOOLS_ROOT} for building libggml-htp skels")

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

@@ -17,6 +17,12 @@
 #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
@@ -33,23 +39,19 @@ static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict
     #pragma unroll(4)
     for (i = 0; i < nvec; i++) {
         // Load y (fp32) and convert into fp16
-        HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+0], zero);  // 32 elements
-        HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+1], zero);  // 32 elements
-        HVX_Vector y_hf  = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
+        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_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)));
+        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 y0_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+0], zero);  // 32 elements
-        HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+1], zero);  // 32 elements
-        HVX_Vector y_hf  = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
+        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
 
         // Load x (fp16)
         HVX_Vector x_hf  = vx[i];
@@ -62,13 +64,72 @@ static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)));
+        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(Q6_Vsf_equals_Vqf32(rsum), hvx_vec_splat_f32(s));
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(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));
+}
 
-    hvx_vec_store_u(r, 4, 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
@@ -91,7 +152,7 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        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) {
@@ -103,12 +164,62 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        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(Q6_Vsf_equals_Vqf32(rsum), hvx_vec_splat_f32(s));
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
-    hvx_vec_store_u(r, 4, 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));
+}
+
+static inline void hvx_dot_f16_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 vx0 = (const HVX_Vector * restrict) x0;  // fp16
+    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
+
+    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(4)
+    for (i = 0; i < nvec; i++) {
+        HVX_Vector y_hf  = vy[i];
+        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) {
+        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_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));
 }
 
 // MAD: y (F32) += x (F16) * s (float)
@@ -317,20 +428,22 @@ 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");
+            static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 <= 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
             HVX_Vector_x4 scores_x4;
             HVX_Vector v_max = hvx_vec_splat_f32(-INFINITY);
             for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
-                float __attribute__((aligned(VLEN))) scores_arr[FLASH_ATTN_BLOCK_SIZE];
-                for (int j = 0; j < VLEN_FP32; ++j) {
+                float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
+                for (int 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 (q->type == HTP_TYPE_F32) {
-                        hvx_dot_f32_f16_aa(&scores_arr[j], q_ptr_vtcm, k_ptr, DK, scale);
+                    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(&scores_arr[j], q_ptr_vtcm, k_ptr, DK, scale);
+                        hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
                     }
                 }
 
@@ -403,7 +516,7 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 float s_val;
                 const uint8_t * k_ptr = k_base + ic * size_k_row_padded;
 
-                if (q->type == HTP_TYPE_F32) {
+                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);

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

@@ -28,19 +28,16 @@ static void hvx_vec_dump_f16(char * pref, HVX_Vector v) {
 }
 
 static void hvx_vec_dump_f32_n(char * pref, HVX_Vector v, uint32_t n) {
-    union {
-        HVX_Vector v;
-        float      d[32];
-    } u = { .v = v };
+    HVX_VectorAlias u = { .v = v };
 
     const uint32_t n0 = n / 16;
     const uint32_t n1 = n % 16;
     int            i  = 0;
     for (; i < n0; i++) {
-        hex_dump_f32_line(pref, u.d + (16 * i), 16);
+        hex_dump_f32_line(pref, u.fp32 + (16 * i), 16);
     }
     if (n1) {
-        hex_dump_f32_line(pref, u.d + (16 * i), n1);
+        hex_dump_f32_line(pref, u.fp32 + (16 * i), n1);
     }
 }
 

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

@@ -44,6 +44,45 @@ static inline HVX_Vector hvx_vec_reduce_sum_qf32(HVX_Vector in) {
     return hvx_vec_reduce_sum_n_qf32(in, 32);
 }
 
+#if __HVX_ARCH__ > 75
+
+static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
+    HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
+    HVX_Vector  sum_sf  = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));
+
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 2));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 4));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 8));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 16));
+    return sum_sf;
+}
+
+static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n) {
+    unsigned int total = n * 4;  // total vec nbytes
+    unsigned int width = 4;      // fp32 nbytes
+
+    HVX_Vector sum = in, sum_t;
+    while (width < total) {
+        sum_t = Q6_V_vror_VR(sum, width);       // rotate right
+        sum   = Q6_Vsf_vadd_VsfVsf(sum, sum_t); // elementwise sum
+        width = width << 1;
+    }
+    return sum;
+}
+
+#else
+
+static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
+    HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
+    HVX_Vector  sum_qf  = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));
+
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 2));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 4));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 8));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 16));
+    return Q6_Vsf_equals_Vqf32(sum_qf);
+}
+
 static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n) {
     unsigned int total = n * 4;  // total vec nbytes
     unsigned int width = 4;      // fp32 nbytes
@@ -57,6 +96,8 @@ static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n)
     return sum;
 }
 
+#endif
+
 static inline HVX_Vector hvx_vec_reduce_sum_f32(HVX_Vector in) {
     return hvx_vec_reduce_sum_n_f32(in, 32);
 }

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

@@ -11,6 +11,7 @@
 
 #include "hex-dma.h"
 #include "hvx-utils.h"
+#include "hvx-dump.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -320,7 +321,7 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -344,7 +345,7 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -362,14 +363,14 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
         // Zero out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -402,7 +403,7 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);                                     // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);                           // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
     HVX_Vector r1_sum = Q6_V_vsplat_R(0);
 
@@ -432,8 +433,8 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -456,20 +457,18 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -493,7 +492,7 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -517,7 +516,7 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -535,14 +534,14 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
         // Zero out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -605,8 +604,8 @@ static void vec_dot_q8x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -629,20 +628,18 @@ static void vec_dot_q8x4x2_q8x4x2_rx2(const int n,
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
@@ -669,7 +666,7 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -708,7 +705,7 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers
@@ -741,14 +738,14 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
         // Zero-out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -781,13 +778,13 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);                                     // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);                           // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
     HVX_Vector r1_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
     // Compute combined scale (fp32).
-    // Apply scale to acc and accumulate into the row sum (qf32).
+    // Apply scale to acc and accumulate into the row sum (f32).
 
     const uint32_t nb   = n / qk;  // num full blocks
     int32_t        nloe = n % qk;  // num leftover elemements (must be signed)
@@ -829,8 +826,8 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers
@@ -867,24 +864,22 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(r0_d, vy_d));
         HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(r1_d, vy_d));
 
-        // Zero-out unused scales
+        // Zero-out unused values
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_f16_f16_aa(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -913,7 +908,7 @@ static void vec_dot_f16_f16_aa(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 
@@ -957,11 +952,8 @@ static void vec_dot_f16_f16_aa_rx2(const int n,
         rsum1 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum1, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)));
     }
 
-    rsum0 = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum0));
-    rsum1 = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum1));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(rsum1, rsum0, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(Q6_Vsf_equals_Vqf32(rsum0), Q6_Vsf_equals_Vqf32(rsum1));
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_f16_f16_uu(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -990,7 +982,7 @@ static void vec_dot_f16_f16_uu(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 
@@ -1042,7 +1034,8 @@ static void vec_dot_f16_f32_uu(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    // Convert into fp32 and reduce
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 

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

@@ -154,8 +154,8 @@ static void hvx_fast_softmax_f32(const uint8_t * restrict src,
         v_pad[i] = v3;
     }
 
-    v       = hvx_vec_reduce_sum_qf32(sum_vec);
-    sum_vec = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(v));
+    v       = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_vec));
+    sum_vec = hvx_vec_repl4(v);
 
     HVX_VectorPred pos_sum   = Q6_Q_vcmp_gt_VwVw(sum_vec, zero_v);
     HVX_Vector     v4        = hvx_vec_inverse_f32(sum_vec);

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

@@ -57,8 +57,8 @@ static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
         sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
     }
 
-    HVX_Vector reduced_sum = hvx_vec_reduce_sum_qf32(sum_v);
-    sum_v                  = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(reduced_sum));
+    HVX_Vector reduced_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+    sum_v                  = hvx_vec_repl4(reduced_sum);
 
     HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
     HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);

ggml/src/ggml-metal/CMakeLists.txt

@@ -71,7 +71,7 @@ else()
         #       disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
         #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)

ggml/src/ggml-metal/ggml-metal-context.h

@@ -15,14 +15,22 @@ typedef struct ggml_metal * ggml_metal_t;
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev);
 void ggml_metal_free(ggml_metal_t ctx);
 
+const char * ggml_metal_get_name(ggml_metal_t ctx);
+
 void ggml_metal_synchronize(ggml_metal_t ctx);
 
 void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
 enum ggml_status ggml_metal_graph_compute (ggml_metal_t ctx, struct ggml_cgraph * gf);
 void             ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf);
 
+void ggml_metal_event_record(ggml_metal_t ctx, ggml_metal_event_t ev);
+void ggml_metal_event_wait  (ggml_metal_t ctx, ggml_metal_event_t ev);
+
+ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx);
+
 void ggml_metal_set_n_cb            (ggml_metal_t ctx, int n_cb);
 void ggml_metal_set_abort_callback  (ggml_metal_t ctx, ggml_abort_callback abort_callback, void * user_data);
 bool ggml_metal_supports_family     (ggml_metal_t ctx, int family);

ggml/src/ggml-metal/ggml-metal-context.m

@@ -24,9 +24,13 @@ struct ggml_metal_command_buffer {
 };
 
 struct ggml_metal {
+    char name[128];
+
     ggml_metal_device_t  dev;
     ggml_metal_library_t lib;
 
+    ggml_metal_event_t ev_cpy; // for async copies
+
     dispatch_queue_t d_queue;
 
     // additional, inference-time compiled pipelines
@@ -117,7 +121,11 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
         }
     }
 
-    //const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+    res->ev_cpy = ggml_metal_device_event_init(dev);
+
+    const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+
+    snprintf(res->name, sizeof(res->name), "%s", props_dev->name);
 
     res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
@@ -206,9 +214,15 @@ void ggml_metal_free(ggml_metal_t ctx) {
 
     dispatch_release(ctx->d_queue);
 
+    ggml_metal_device_event_free(ctx->dev, ctx->ev_cpy);
+
     free(ctx);
 }
 
+const char * ggml_metal_get_name(ggml_metal_t ctx) {
+    return ctx->name;
+}
+
 void ggml_metal_synchronize(ggml_metal_t ctx) {
     // wait for any backend operations to finish
     if (ctx->cmd_buf_last) {
@@ -273,8 +287,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
         // wrap the source data into a Metal buffer
         id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
         id<MTLBuffer> buf_src = [device newBufferWithBytes:data
-                                                         length:size
-                                                        options:MTLResourceStorageModeShared];
+                                                    length:size
+                                                   options:MTLResourceStorageModeShared];
 
         GGML_ASSERT(buf_src);
 
@@ -316,9 +330,9 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
     @autoreleasepool {
         id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
         id<MTLBuffer> buf_dst = [device newBufferWithBytesNoCopy:data
-                                                               length:size
-                                                              options:MTLResourceStorageModeShared
-                                                          deallocator:nil];
+                                                          length:size
+                                                         options:MTLResourceStorageModeShared
+                                                     deallocator:nil];
 
         GGML_ASSERT(buf_dst);
 
@@ -356,9 +370,52 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
     }
 }
 
+bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    @autoreleasepool {
+        struct ggml_metal_buffer_id bid_src = ggml_metal_get_buffer_id(src);
+        struct ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(dst);
+
+        if (bid_src.metal == nil || bid_dst.metal == nil) {
+            return false;
+        }
+
+        // queue the copy operation into the Metal context
+        // this will be queued at the end, after any currently ongoing GPU operations
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx_src->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+        id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+        [encoder copyFromBuffer:bid_src.metal
+                   sourceOffset:bid_src.offs
+                       toBuffer:bid_dst.metal
+              destinationOffset:bid_dst.offs
+                           size:ggml_nbytes(src)];
+
+        [encoder endEncoding];
+
+        ggml_metal_event_t ev_cpy = ggml_metal_get_ev_cpy(ctx_src);
+        ggml_metal_event_encode_signal(ev_cpy, cmd_buf);
+
+        [cmd_buf commit];
+
+        // do not wait here for completion
+        //[cmd_buf waitUntilCompleted];
+
+        // instead, remember a reference to the command buffer and wait for it later if needed
+        [ctx_src->cmd_bufs_ext addObject:cmd_buf];
+        ctx_src->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+
+        ggml_metal_event_wait(ctx_dst, ev_cpy);
+
+        return true;
+    }
+}
+
 enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
     // number of nodes encoded by the main thread (empirically determined)
-    const int n_main = 64;
+    const int n_main = MAX(64, 0.1*gf->n_nodes);
 
     // number of threads in addition to the main thread
     const int n_cb = ctx->n_cb;
@@ -530,6 +587,42 @@ void ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf) {
     //printf("%s: graph optimize took %.3f ms\n", __func__, (ggml_time_us() - t_start) / 1000.0);
 }
 
+void ggml_metal_event_record(ggml_metal_t ctx, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+
+        ggml_metal_event_encode_signal(ev, cmd_buf);
+
+        [cmd_buf commit];
+
+        [ctx->cmd_bufs_ext addObject:cmd_buf];
+        ctx->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+    }
+}
+
+void ggml_metal_event_wait(ggml_metal_t ctx, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+
+        ggml_metal_event_encode_wait(ev, cmd_buf);
+
+        [cmd_buf commit];
+
+        [ctx->cmd_bufs_ext addObject:cmd_buf];
+        ctx->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+    }
+}
+
+ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx) {
+    return ctx->ev_cpy;
+}
+
 void ggml_metal_set_n_cb(ggml_metal_t ctx, int n_cb) {
     if (ctx->n_cb != n_cb) {
         ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_COMMAND_BUFFERS);

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -17,10 +17,12 @@ struct ggml_metal_device_deleter {
 
 typedef std::unique_ptr<ggml_metal_device, ggml_metal_device_deleter> ggml_metal_device_ptr;
 
-ggml_metal_device_t ggml_metal_device_get(void) {
-    static ggml_metal_device_ptr ctx { ggml_metal_device_init() };
+ggml_metal_device_t ggml_metal_device_get(int device) {
+    static std::vector<ggml_metal_device_ptr> devs;
 
-    return ctx.get();
+    devs.emplace_back(ggml_metal_device_init(device));
+
+    return devs.back().get();
 }
 
 struct ggml_metal_pipelines {
@@ -174,6 +176,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    const int n = op->src[0]->ne[0];
+
+    snprintf(base, 256, "kernel_diag_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_n=%d", base, n);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    res.nsg  = 1;
+    res.smem = 0;
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_metal_library_t lib, ggml_type tsrc) {
     char base[256];
     char name[256];
@@ -532,6 +554,36 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    const int nsg = 8;
+    const int n   = op->src[1]->ne[1];
+    const int k   = op->src[1]->ne[0];
+
+    snprintf(base, 256, "kernel_solve_tri_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_nsg=%d_n=%d_k=%d", base, nsg, n, k);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, nsg, FC_SOLVE_TRI + 0);
+        ggml_metal_cv_set_int16(cv, n,   FC_SOLVE_TRI + 1);
+        ggml_metal_cv_set_int16(cv, k,   FC_SOLVE_TRI + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.nsg  = nsg;
+    res.smem = GGML_PAD(GGML_PAD(n, 32)*nsg*sizeof(float), 16);
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
     char base[256];
     char name[256];
@@ -1340,34 +1392,78 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
     GGML_UNUSED(op);
 }
 
-ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(
-        ggml_metal_library_t lib,
-        ggml_op op,
-        int32_t n_fuse,
-        bool row) {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_library_t lib, const ggml_tensor * op, int32_t n_fuse) {
     char base[256];
     char name[256];
 
-    const char * op_str = "undefined";
-    switch (op) {
-        case GGML_OP_ADD:   op_str = "add";   break;
-        case GGML_OP_SUB:   op_str = "sub";   break;
-        case GGML_OP_MUL:   op_str = "mul";   break;
-        case GGML_OP_DIV:   op_str = "div";   break;
+    int op_num = -1;
+
+    switch (op->op) {
+        case GGML_OP_ADD: op_num = 0; break;
+        case GGML_OP_SUB: op_num = 1; break;
+        case GGML_OP_MUL: op_num = 2; break;
+        case GGML_OP_DIV: op_num = 3; break;
         default: GGML_ABORT("fatal error");
     };
 
-    if (row) {
-        snprintf(base, 256, "kernel_%s_row_c4_fuse_%d", op_str, n_fuse);
-    } else {
-        snprintf(base, 256, "kernel_%s_fuse_%d", op_str, n_fuse);
-    }
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t1_str = ggml_type_name(op->src[1]->type);
+    const char * t_str  = ggml_type_name(op->type);
 
-    snprintf(name, 256, "%s", base);
+    const bool is_c4 = (op->src[0]->ne[0] % 4 == 0) && (op->src[1]->ne[0] % 4 == 0);
+
+    const bool is_rb = ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && (ggml_nrows(op->src[1]) == 1) && ggml_nelements(op) < 65536;
+
+    snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s%s", t0_str, t1_str, t_str, is_c4 ? "_4" : "");
+    snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d", base, op_num, n_fuse, is_rb);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
+        ggml_metal_cv_set_int16(cv, n_fuse, FC_BIN + 1);
+        ggml_metal_cv_set_bool (cv, is_rb,  FC_BIN + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.c4  = is_c4;
+    res.cnt = is_rb;
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one(ggml_metal_library_t lib, ggml_op op) {
+    char base[256];
+    char name[256];
+
+    int op_num = -1;
+
+    switch (op) {
+        case GGML_OP_ADD: op_num = 0; break;
+        case GGML_OP_SUB: op_num = 1; break;
+        case GGML_OP_MUL: op_num = 2; break;
+        case GGML_OP_DIV: op_num = 3; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s", "f32", "f32", "f32");
+    snprintf(name, 256, "%s_op=%d_nf=%d", base, op_num, 1);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
+        ggml_metal_cv_set_int16(cv, 1,      FC_BIN + 1);
+        ggml_metal_cv_set_bool (cv, false,  FC_BIN + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
     }
 
     return res;

ggml/src/ggml-metal/ggml-metal-device.h

@@ -53,6 +53,9 @@ struct ggml_metal_pipeline_with_params {
     int nr1;
 
     size_t smem;
+
+    bool c4;
+    bool cnt;
 };
 
 int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline);
@@ -108,6 +111,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_1d
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -121,6 +125,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -132,7 +137,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
-struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse );
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
@@ -205,7 +211,9 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
 //
 
 struct ggml_metal_device_props {
+    int device;
     char name[128];
+    char desc[128];
 
     size_t max_buffer_size;
     size_t max_working_set_size;
@@ -224,11 +232,15 @@ struct ggml_metal_device_props {
     int op_offload_min_batch_size;
 };
 
-ggml_metal_device_t ggml_metal_device_init(void);
+typedef struct ggml_metal_event * ggml_metal_event_t;
+
+void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
+void ggml_metal_event_encode_wait  (ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
+
+ggml_metal_device_t ggml_metal_device_init(int device);
 void ggml_metal_device_free(ggml_metal_device_t dev);
 
-// return a singleton that is automatically destroyed when the program exits
-ggml_metal_device_t ggml_metal_device_get(void);
+ggml_metal_device_t ggml_metal_device_get(int device);
 
 void * ggml_metal_device_get_obj  (ggml_metal_device_t dev); // id<MTLDevice>
 void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQueue>
@@ -240,6 +252,10 @@ void ggml_metal_device_rsets_rm (ggml_metal_device_t dev, ggml_metal_rset_t rset
 
 void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev);
 
+ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev);
+void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev);
+void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev);
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total);
 bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_tensor * op);
 

ggml/src/ggml-metal/ggml-metal-device.m

@@ -24,9 +24,6 @@
 static const NSInteger MTLGPUFamilyMetal3_GGML = 5001;
 static const NSInteger MTLGPUFamilyMetal4_GGML = 5002;
 
-// virtual address for GPU memory allocations
-static atomic_uintptr_t g_addr_device = 0x000000400ULL;
-
 #if !GGML_METAL_EMBED_LIBRARY
 // Here to assist with NSBundle Path Hack
 @interface GGMLMetalClass : NSObject
@@ -349,10 +346,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta
 
     struct ggml_metal_pipeline_with_params res = {
         /*.pipeline =*/ nil,
+        /*.nsg      =*/ 0,
         /*.nr0      =*/ 0,
         /*.nr1      =*/ 0,
-        /*.nsg      =*/ 0,
         /*.smem     =*/ 0,
+        /*.c4       =*/ false,
+        /*.cnt      =*/ false,
     };
 
     res.pipeline = ggml_metal_pipelines_get(lib->pipelines, name);
@@ -365,10 +364,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta
 struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv) {
     struct ggml_metal_pipeline_with_params res = {
         /*.pipeline =*/ nil,
+        /*.nsg      =*/ 0,
         /*.nr0      =*/ 0,
         /*.nr1      =*/ 0,
-        /*.nsg      =*/ 0,
         /*.smem     =*/ 0,
+        /*.c4       =*/ false,
+        /*.cnt      =*/ false,
     };
 
     [lib->lock lock];
@@ -523,6 +524,9 @@ struct ggml_metal_device {
     ggml_metal_library_t library;
 
     struct ggml_metal_device_props props;
+
+    // virtual address for GPU memory allocations
+    atomic_uintptr_t addr_virt;
 };
 
 //
@@ -618,7 +622,7 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets) {
     free(rsets);
 }
 
-ggml_metal_device_t ggml_metal_device_init(void) {
+ggml_metal_device_t ggml_metal_device_init(int device) {
     ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));
 
     assert(dev != NULL);
@@ -632,6 +636,9 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                 GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
             }
 
+            dev->addr_virt = 0x000000400ULL;
+
+            dev->props.device = device;
             dev->props.has_simdgroup_reduction  = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
             dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
@@ -792,7 +799,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                 dev->props.max_working_set_size   = dev->mtl_device.maxBufferLength;
             }
 
-            strncpy(dev->props.name, [[dev->mtl_device name] UTF8String], sizeof(dev->props.name) - 1);
+            snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device);
+            snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", [[dev->mtl_device name] UTF8String]);
 
             dev->library = ggml_metal_library_init(dev);
             if (!dev->library) {
@@ -922,6 +930,59 @@ void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev) {
     atomic_store_explicit(&dev->rsets->d_loop, 2*dev->rsets->keep_alive_s, memory_order_relaxed);
 }
 
+struct ggml_metal_event {
+    void * obj; // id<MTLEvent>
+
+    atomic_int value;
+};
+
+void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
+    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+
+    id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
+
+    [cmd_buf encodeSignalEvent:event value:atomic_fetch_add_explicit(&ev->value, 1, memory_order_relaxed) + 1];
+}
+
+void ggml_metal_event_encode_wait(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
+    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+
+    id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
+
+    [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
+}
+
+ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
+    id<MTLEvent> event = [dev->mtl_device newEvent];
+
+    ggml_metal_event_t ev = calloc(1, sizeof(struct ggml_metal_event));
+
+    ev->obj = (__bridge void *)event;
+    ev->value = 0;
+
+    return ev;
+}
+
+void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
+    id<MTLEvent> event = ev->obj;
+    [event release];
+
+    free(ev);
+
+    GGML_UNUSED(dev);
+}
+
+void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLEvent> event = ev->obj;
+
+        id<MTLCommandBuffer> cmd_buf = [dev->mtl_queue commandBuffer];
+        [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
+        [cmd_buf commit];
+        [cmd_buf waitUntilCompleted];
+    }
+}
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {
     if (@available(macOS 10.12, iOS 16.0, *)) {
         *total = dev->mtl_device.recommendedMaxWorkingSetSize;
@@ -997,7 +1058,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_ADD_ID:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ACC:
         case GGML_OP_REPEAT:
         case GGML_OP_SCALE:
@@ -1096,6 +1157,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
             return true;
+        case GGML_OP_SOLVE_TRI:
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return has_simdgroup_reduction;
@@ -1177,6 +1239,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                         return false;
                 };
             }
+        case GGML_OP_DIAG:
+            return true;
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
             return has_simdgroup_reduction;
@@ -1344,8 +1408,8 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
         res->all_data = ggml_metal_host_malloc(size_aligned);
         res->is_shared = true;
     } else {
-        // use virtual address from g_addr_device counter
-        res->all_data = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);
+        // use virtual address
+        res->all_data = (void *) atomic_fetch_add_explicit(&dev->addr_virt, size_aligned, memory_order_relaxed);
         res->is_shared = false;
     }
     res->all_size = size_aligned;

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -78,13 +78,15 @@
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
 #define FC_SSM_CONV                    900
-#define FC_COUNT_EQUAL                 1000
+#define FC_SOLVE_TRI                   1000
+#define FC_COUNT_EQUAL                 1100
+#define FC_BIN                         1200
 
 // op-specific constants
-#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NQPSG 8
 #define OP_FLASH_ATTN_EXT_NCPSG 64
 
-#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
 // kernel argument structs
@@ -733,6 +735,33 @@ typedef struct {
     uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
     int32_t  ne00t;
     int32_t  ne00;
@@ -764,6 +793,25 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_set_rows;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_diag;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -341,6 +341,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_rwkv(ctx, idx);
             } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                n_fuse = ggml_metal_op_solve_tri(ctx, idx);
+            } break;
         case GGML_OP_MUL_MAT:
             {
                 n_fuse = ggml_metal_op_mul_mat(ctx, idx);
@@ -357,6 +361,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_set_rows(ctx, idx);
             } break;
+        case GGML_OP_DIAG:
+            {
+                n_fuse = ggml_metal_op_diag(ctx, idx);
+            } break;
         case GGML_OP_L2_NORM:
             {
                 n_fuse = ggml_metal_op_l2_norm(ctx, idx);
@@ -699,7 +707,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
         /*.o1   =*/ { 0 },
     };
 
-    auto pipeline = ggml_metal_library_get_pipeline_bin(lib, GGML_OP_ADD, 1, false);
+    auto pipeline = ggml_metal_library_get_pipeline_bin_one(lib, GGML_OP_ADD);
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -1255,6 +1263,48 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS(int32_t,  ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS(int32_t,  ne, op, ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
+
+    ggml_metal_kargs_diag args = {
+        /*.ne00 =*/ne00,
+        /*.ne01 =*/ne01,
+        /*.ne02 =*/ne02,
+        /*.ne03 =*/ne03,
+        /*.nb00 =*/nb00,
+        /*.nb01 =*/nb01,
+        /*.nb02 =*/nb02,
+        /*.nb03 =*/nb03,
+        /*.ne0  =*/ne0,
+        /*.ne1  =*/ne1,
+        /*.ne2  =*/ne2,
+        /*.ne3  =*/ne3,
+        /*.nb0  =*/nb0,
+        /*.nb1  =*/nb1,
+        /*.nb2  =*/nb2,
+        /*.nb3  =*/nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_diag(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, 32, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -1557,6 +1607,63 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_kargs_solve_tri args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, (ne10 + nsg - 1)/nsg, ne02, ne03, 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -2295,7 +2402,7 @@ size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
     //    return res;
     //}
 
-    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPSG : OP_FLASH_ATTN_EXT_NQPSG;
     const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
 
     const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
@@ -2411,7 +2518,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
 
         GGML_ASSERT(nqptg <= 32);
@@ -2578,9 +2685,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int nkpsg = 1*ncpsg;
+        const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 1  == 0);
@@ -2632,6 +2739,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_op_concurrency_reset(ctx);
         }
 
+        // note: for simplicity assume the K is larger or equal than V
+        GGML_ASSERT(ne10 >= ne20);
+
         // ne00 + 2*ncpsg*(nsg)
         // for each query, we load it as f16 in shared memory (ne00)
         // and store the soft_max values and the mask
@@ -2639,28 +2749,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // ne20*(nsg)
         // each simdgroup has a full f32 head vector in shared mem to accumulate results
         //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*GGML_PAD(ne20, 128)*(nsg))*(sizeof(float)/2), 16))
-
-        int64_t nsgmax = 2;
-        while (true) {
-            const size_t smem = FATTN_SMEM(nsgmax);
-            // avoid using more than half of the threadgroup memory - can cause slow downs especially for large head sizes
-            if (smem > props_dev->max_theadgroup_memory_size/2) {
-                break;
-            }
-            nsgmax *= 2;
-        }
-        nsgmax /= 2;
-
-        // simdgroups per threadgroup (a.k.a. warps)
-        //const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
-        const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) 1024/32)));
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
-        while (nsg <= nsgt) {
-            nsg *= 2;
-        }
-        nsg /= 2;
 
         // workgroups
         // each workgroup handles nsg*nkpsg cache values
@@ -2673,7 +2764,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         } else {
             nwg = 32;
             nsg = 1;
-            while (2*nwg*nsg*nkpsg < ne11 && nsg < 4) {
+            while (2*nwg*nsg*ncpsg < ne11 && nsg < 4) {
                 nsg *= 2;
             }
         }
@@ -2739,7 +2830,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
         } else {
             // sanity checks
             assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
@@ -2752,7 +2843,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
 
             // sync the 2 kernels
             ggml_metal_op_concurrency_reset(ctx);
@@ -2804,8 +2895,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
-    bool bcast_row = false;
-
     ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
     ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
     ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
@@ -2899,18 +2988,7 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
 
     struct ggml_metal_pipeline_with_params pipeline;
 
-    if (ggml_nelements(op->src[1]) == ne10 && ggml_is_contiguous(op->src[1]) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(op->src[0]));
-
-        // src1 is a row
-        GGML_ASSERT(ne11 == 1);
-
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, true);
-
-        bcast_row = true;
-    } else {
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, false);
-    }
+    pipeline = ggml_metal_library_get_pipeline_bin(lib, op, n_fuse);
 
     if (n_fuse > 1) {
         bid_dst = ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
@@ -2924,20 +3002,28 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
         }
     }
 
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne10 = ne10/4;
+        args.ne0  = ne0/4;
+    }
+
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
     ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
     ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
     ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
 
-    if (bcast_row) {
-        const int64_t n = ggml_nelements(op)/4;
+    if (pipeline.cnt) {
+        const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
 
         ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
     } else {
-        int nth = 32;
+        const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
-        while (16*nth < ne0 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        int nth = 1;
+
+        while (2*nth < args.ne0 && nth < nth_max) {
             nth *= 2;
         }
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -56,10 +56,12 @@ int ggml_metal_op_sum_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rwkv              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal.cpp

@@ -7,11 +7,15 @@
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
 
-// globals
+#include <mutex>
+#include <string>
 
-// initialized in ggml_backend_metal_reg
-static ggml_backend_reg    g_ggml_metal_reg;
-static ggml_backend_device g_ggml_metal_device;
+#define GGML_METAL_NAME "MTL"
+#define GGML_METAL_MAX_DEVICES 16
+
+// number of Metal devices
+// note: can be overriden with GGML_METAL_DEVICES env to simulate virtual devices
+static int g_devices = 1;
 
 ////////////////////////////////////////////////////////////////////////////////
 // backend interface
@@ -165,10 +169,28 @@ static ggml_backend_buffer_i ggml_backend_metal_buffer_private_i = {
     /* .reset           = */ NULL,
 };
 
+static bool ggml_backend_buffer_is_metal(ggml_backend_buffer_t buffer) {
+    return buffer->iface.free_buffer == ggml_backend_metal_buffer_shared_free_buffer ||
+           buffer->iface.free_buffer == ggml_backend_metal_buffer_private_free_buffer;
+}
+
 //
 // buffer types
 //
 
+struct ggml_backend_metal_buffer_type {
+    int device;
+    std::string name;
+};
+
+struct ggml_backend_metal_buffer_type_deleter {
+    void operator()(ggml_backend_metal_buffer_type * ctx) const {
+        delete ctx;
+    }
+};
+
+typedef std::unique_ptr<ggml_backend_metal_buffer_type, ggml_backend_metal_buffer_type_deleter> ggml_backend_metal_buffer_type_ptr;
+
 // common method for allocating shread or private Metal buffers
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size, bool shared) {
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)buft->device->context;
@@ -218,9 +240,9 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
 // default (shared) buffer type
 
 static const char * ggml_backend_metal_buffer_type_shared_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;
 
-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }
 
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_shared_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -249,29 +271,54 @@ static bool ggml_backend_metal_buffer_type_shared_is_host(ggml_backend_buffer_ty
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_shared(void) {
-    static ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_shared_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_shared_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_shared_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_shared_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_shared_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_shared_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_shared(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
 
-    return &ggml_backend_buffer_type_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx =
+                new ggml_backend_metal_buffer_type {
+                    /* .device = */ i,
+                    /* .name   = */ GGML_METAL_NAME + std::to_string(i),
+                };
+            ctxs.emplace_back(raw_ctx);
+
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_shared_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_shared_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_shared_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_shared_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_shared_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_shared_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }
 
 // default (private) buffer type
 
 static const char * ggml_backend_metal_buffer_type_private_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal_Private";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;
 
-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }
 
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_private_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -300,29 +347,53 @@ static bool ggml_backend_metal_buffer_type_private_is_host(ggml_backend_buffer_t
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_private(void) {
-    static ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_private_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_private_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_private_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_private_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_private_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_private_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_private(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
 
-    return &ggml_backend_buffer_type_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx = new ggml_backend_metal_buffer_type{
+                /* .device = */ i,
+                /* .name   = */ GGML_METAL_NAME + std::to_string(i) + "_Private"
+            };
+            ctxs.emplace_back(raw_ctx);
+
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_private_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_private_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_private_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_private_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_private_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_private_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }
 
 // mapped buffer type
 
 static const char * ggml_backend_metal_buffer_type_mapped_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal_Mapped";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;
 
-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }
 
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_mapped_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -352,31 +423,55 @@ static bool ggml_backend_metal_buffer_type_mapped_is_host(ggml_backend_buffer_ty
     GGML_UNUSED(buft);
 }
 
-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_mapped(void) {
-    // note: not obvious, but this buffer type still needs to implement .alloc_buffer:
-    //       https://github.com/ggml-org/llama.cpp/pull/15832#discussion_r2333177099
-    static ggml_backend_buffer_type ggml_backend_buffer_type_mapped_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_mapped_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_mapped_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_mapped_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_mapped_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_mapped_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_mapped_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_mapped(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
 
-    return &ggml_backend_buffer_type_mapped_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx = new ggml_backend_metal_buffer_type{
+                /* .device = */ i,
+                /* .name   = */ GGML_METAL_NAME + std::to_string(i) + "_Mapped"
+            };
+            ctxs.emplace_back(raw_ctx);
+
+            // note: not obvious, but this buffer type still needs to implement .alloc_buffer:
+            //       https://github.com/ggml-org/llama.cpp/pull/15832#discussion_r2333177099
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_mapped_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_mapped_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_mapped_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_mapped_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_mapped_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_mapped_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }
 
 // backend
 
 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
-    return "Metal";
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;
 
-    GGML_UNUSED(backend);
+    return ggml_metal_get_name(ctx);
 }
 
 static void ggml_backend_metal_free(ggml_backend_t backend) {
@@ -409,12 +504,24 @@ static void ggml_backend_metal_get_tensor_async(ggml_backend_t backend, const gg
 }
 
 static bool ggml_backend_metal_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    return false;
+    if (!ggml_backend_is_metal(backend_src) || !ggml_backend_is_metal(backend_dst)) {
+        return false;
+    }
 
-    GGML_UNUSED(backend_src);
-    GGML_UNUSED(backend_dst);
-    GGML_UNUSED(src);
-    GGML_UNUSED(dst);
+    if (!ggml_backend_buffer_is_metal(src->buffer) || !ggml_backend_buffer_is_metal(dst->buffer)) {
+        return false;
+    }
+
+    ggml_metal_t ctx_src = (ggml_metal_t)backend_src->context;
+    ggml_metal_t ctx_dst = (ggml_metal_t)backend_dst->context;
+
+    //ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
+    //ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    //ggml_metal_buffer_t buf_ctx_src = (ggml_metal_buffer_t)buf_src->context;
+    //ggml_metal_buffer_t buf_ctx_dst = (ggml_metal_buffer_t)buf_dst->context;
+
+    return ggml_metal_cpy_tensor_async(ctx_src, ctx_dst, src, dst);
 }
 
 static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
@@ -423,6 +530,20 @@ static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend,
     return ggml_metal_graph_compute(ctx, cgraph);
 }
 
+static void ggml_backend_metal_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_event_record(ctx, ev);
+}
+
+static void ggml_backend_metal_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_event_wait(ctx, ev);
+}
+
 static void ggml_backend_metal_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_metal_t ctx = (ggml_metal_t)backend->context;
 
@@ -435,7 +556,6 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
     ggml_metal_t ctx = (ggml_metal_t)backend->context;
 
     ggml_metal_set_n_cb(ctx, n_cb);
-
 }
 
 static ggml_backend_i ggml_backend_metal_i = {
@@ -450,12 +570,8 @@ static ggml_backend_i ggml_backend_metal_i = {
     /* .graph_plan_update       = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
-
-    // the events API is needed only for multi-GPU setups, so likely no need to implement it for Metal
-    // in any case, these docs seem relevant if we ever decide to implement it:
-    // https://developer.apple.com/documentation/metal/mtlcommandbuffer#Synchronizing-Passes-with-Events
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
+    /* .event_record            = */ ggml_backend_metal_event_record,
+    /* .event_wait              = */ ggml_backend_metal_event_wait,
     /* .graph_optimize          = */ ggml_backend_metal_graph_optimize,
 };
 
@@ -519,15 +635,17 @@ void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
 // backend device
 
 static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) {
-    return "Metal";
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
 
-    GGML_UNUSED(dev);
+    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);
+
+    return props_dev->name;
 }
 
 static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) {
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
 
-    return ggml_metal_device_get_props(ctx_dev)->name;
+    return ggml_metal_device_get_props(ctx_dev)->desc;
 }
 
 static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
@@ -550,14 +668,14 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
     ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
     props->caps = {
-        /* .async                 = */ true,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ true,
-        /* .events                = */ false,
+        /* .async                = */ true,
+        /* .host_buffer          = */ false,
+        /* .buffer_from_host_ptr = */ true,
+        /* .events               = */ true,
     };
 }
 
-static ggml_backend_t ggml_backend_metal_device_init(ggml_backend_dev_t dev, const char * params) {
+static ggml_backend_t ggml_backend_metal_device_init_backend(ggml_backend_dev_t dev, const char * params) {
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
 
     ggml_metal_t ctx = ggml_metal_init(ctx_dev);
@@ -587,7 +705,7 @@ static ggml_backend_buffer_type_t ggml_backend_metal_device_get_buffer_type(ggml
 
     const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);
 
-    return props_dev->use_shared_buffers ? ggml_backend_metal_buffer_type_shared() : ggml_backend_metal_buffer_type_private();
+    return props_dev->use_shared_buffers ? ggml_backend_metal_buffer_type_shared(props_dev->device) : ggml_backend_metal_buffer_type_private(props_dev->device);
 }
 
 static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
@@ -595,7 +713,9 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backen
 
     ggml_metal_buffer_t res = ggml_metal_buffer_map(ctx_dev, ptr, size, max_tensor_size);
 
-    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type_mapped(), ggml_backend_metal_buffer_shared_i, res, size);
+    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type_mapped(props_dev->device), ggml_backend_metal_buffer_shared_i, res, size);
 }
 
 static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
@@ -606,9 +726,10 @@ static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const
 
 static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
     return
+        buft->device == dev && (
         buft->iface.get_name == ggml_backend_metal_buffer_type_shared_get_name ||
         buft->iface.get_name == ggml_backend_metal_buffer_type_private_get_name ||
-        buft->iface.get_name == ggml_backend_metal_buffer_type_mapped_get_name;
+        buft->iface.get_name == ggml_backend_metal_buffer_type_mapped_get_name);
 
     GGML_UNUSED(dev);
 }
@@ -632,45 +753,97 @@ static bool ggml_backend_metal_device_offload_op(ggml_backend_dev_t dev, const g
             get_op_batch_size(op) >= ggml_metal_device_get_props(ctx_dev)->op_offload_min_batch_size;
 }
 
+static ggml_backend_event_t ggml_backend_metal_device_event_new(ggml_backend_dev_t dev) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t event = ggml_metal_device_event_init(ctx_dev);
+    GGML_ASSERT(event);
+
+    ggml_backend_event_t ev = new ggml_backend_event {
+        /* .device  = */ dev,
+        /* .context = */ event,
+    };
+
+    return ev;
+}
+
+static void ggml_backend_metal_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_device_event_free(ctx_dev, ev);
+
+    delete event;
+}
+
+static void ggml_backend_metal_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t evt = (ggml_metal_event_t)event->context;
+
+    ggml_metal_device_event_synchronize(ctx_dev, evt);
+}
+
 static ggml_backend_device_i ggml_backend_metal_device_i = {
     /* .get_name             = */ ggml_backend_metal_device_get_name,
     /* .get_description      = */ ggml_backend_metal_device_get_description,
     /* .get_memory           = */ ggml_backend_metal_device_get_memory,
     /* .get_type             = */ ggml_backend_metal_device_get_type,
     /* .get_props            = */ ggml_backend_metal_device_get_props,
-    /* .init_backend         = */ ggml_backend_metal_device_init,
+    /* .init_backend         = */ ggml_backend_metal_device_init_backend,
     /* .get_buffer_type      = */ ggml_backend_metal_device_get_buffer_type,
     /* .get_host_buffer_type = */ NULL,
     /* .buffer_from_host_ptr = */ ggml_backend_metal_device_buffer_mapped,
     /* .supports_op          = */ ggml_backend_metal_device_supports_op,
     /* .supports_buft        = */ ggml_backend_metal_device_supports_buft,
     /* .offload_op           = */ ggml_backend_metal_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
+    /* .event_new            = */ ggml_backend_metal_device_event_new,
+    /* .event_free           = */ ggml_backend_metal_device_event_free,
+    /* .event_synchronize    = */ ggml_backend_metal_device_event_synchronize,
 };
 
 // backend registry
 
+struct ggml_backend_metal_reg {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+typedef struct ggml_backend_metal_reg * ggml_backend_metal_reg_t;
+
+static ggml_backend_metal_reg_t ggml_backend_metal_reg_init(void) {
+    ggml_backend_metal_reg_t ctx = new struct ggml_backend_metal_reg;
+
+    return ctx;
+}
+
+static void ggml_backend_metal_reg_free(ggml_backend_metal_reg_t ctx) {
+    delete ctx;
+}
+
+struct ggml_backend_metal_reg_deleter {
+    void operator()(ggml_backend_metal_reg_t ctx) {
+        ggml_backend_metal_reg_free(ctx);
+    }
+};
+
+typedef std::unique_ptr<struct ggml_backend_metal_reg, ggml_backend_metal_reg_deleter> ggml_backend_metal_reg_ptr;
+
 static const char * ggml_backend_metal_reg_get_name(ggml_backend_reg_t reg) {
-    return "Metal";
+    return GGML_METAL_NAME;
 
     GGML_UNUSED(reg);
 }
 
 static size_t ggml_backend_metal_reg_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
+    ggml_backend_metal_reg_t ctx = (ggml_backend_metal_reg_t)reg->context;
+    return ctx->devices.size();
 }
 
 static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    return &g_ggml_metal_device;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(index);
+    ggml_backend_metal_reg_t ctx = (ggml_backend_metal_reg_t)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
 }
 
 static ggml_backend_feature g_ggml_backend_metal_features[] = {
@@ -698,27 +871,67 @@ static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const
 
 static ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_name         = */ ggml_backend_metal_reg_get_name,
-    /* .device_count     = */ ggml_backend_metal_reg_device_count,
-    /* .device_get       = */ ggml_backend_metal_reg_device_get,
+    /* .get_device_count = */ ggml_backend_metal_reg_device_count,
+    /* .get_device       = */ ggml_backend_metal_reg_device_get,
     /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
-ggml_backend_reg_t ggml_backend_metal_reg(void) {
-    {
-        g_ggml_metal_reg = {
-            /* .api_version = */ GGML_BACKEND_API_VERSION,
-            /* .iface       = */ ggml_backend_metal_reg_i,
-            /* .context     = */ NULL,
-        };
+static ggml_backend_dev_t ggml_backend_metal_device_init(ggml_backend_reg_t reg, int device) {
+    return new ggml_backend_device {
+        /* .iface   = */ ggml_backend_metal_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ ggml_metal_device_get(device),
+    };
+}
 
-        g_ggml_metal_device = {
-            /* .iface   = */ ggml_backend_metal_device_i,
-            /* .reg     = */ &g_ggml_metal_reg,
-            /* .context = */ ggml_metal_device_get(),
-        };
+static void ggml_backend_metal_device_free(ggml_backend_dev_t dev) {
+    delete dev;
+}
+
+struct ggml_backend_device_deleter {
+    void operator()(ggml_backend_dev_t ctx) {
+        ggml_backend_metal_device_free(ctx);
+    }
+};
+
+typedef std::unique_ptr<ggml_backend_device, ggml_backend_device_deleter> ggml_backend_device_ptr;
+
+ggml_backend_reg_t ggml_backend_metal_reg(void) {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        const char * env = getenv("GGML_METAL_DEVICES");
+        if (env) {
+            g_devices = atoi(env);
+        }
+
+        static std::vector<ggml_backend_device_ptr> devs;
+
+        if (!initialized) {
+            static ggml_backend_metal_reg_ptr reg_ctx(ggml_backend_metal_reg_init());
+
+            for (int i = 0; i < g_devices; ++i) {
+                auto * dev = ggml_backend_metal_device_init(&reg, i);
+                devs.emplace_back(dev);
+
+                reg_ctx->devices.push_back(dev);
+            }
+
+            reg = {
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_metal_reg_i,
+                /* .context     = */ reg_ctx.get(),
+            };
+        }
+
+        initialized = true;
     }
 
-    return &g_ggml_metal_reg;
+    return &reg;
 }
 
 GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg)

ggml/src/ggml-metal/ggml-metal.metal

@@ -895,11 +895,13 @@ enum ggml_sort_order {
     GGML_SORT_ORDER_DESC,
 };
 
-// general-purpose kernel for addition, subtraction, multiplication and division of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across all dims
-// cons: not very efficient
-template <int F>
-kernel void kernel_add_fuse_impl(
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
         constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
@@ -907,138 +909,152 @@ kernel void kernel_add_fuse_impl(
         uint3   tgpig[[threadgroup_position_in_grid]],
         ushort3 tpitg[[thread_position_in_threadgroup]],
         ushort3   ntg[[threads_per_threadgroup]]) {
-    const int i03 = tgpig.z;
-    const int i02 = tgpig.y;
-    const int i01 = tgpig.x;
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
 
-    const int i13 = i03%args.ne13;
-    const int i12 = i02%args.ne12;
-    const int i11 = i01%args.ne11;
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.x;
+        const uint i1 = i0%args.ne10;
 
-    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
-    device       float * dst_ptr  = (device       float *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
 
-    device const float * src1_ptr[F];
-    for (short j = 0; j < F; ++j) {
-        src1_ptr[j] = (device const float *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
-    }
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
 
-    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-        const int i10 = i0%args.ne10;
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
 
-        float res = src0_ptr[i0];
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
 
-#pragma unroll
-        for (short j = 0; j < F; ++j) {
-            res += src1_ptr[j][i10];
-        }
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
 
-        dst_ptr[i0] = res;
-    }
-}
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
 
-typedef decltype(kernel_add_fuse_impl<2>) kernel_add_fuse_t;
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
 
-template [[host_name("kernel_add_fuse_1")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<1>;
-template [[host_name("kernel_add_fuse_2")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<2>;
-template [[host_name("kernel_add_fuse_3")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<3>;
-template [[host_name("kernel_add_fuse_4")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<4>;
-template [[host_name("kernel_add_fuse_5")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<5>;
-template [[host_name("kernel_add_fuse_6")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<6>;
-template [[host_name("kernel_add_fuse_7")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<7>;
-template [[host_name("kernel_add_fuse_8")]] kernel kernel_add_fuse_t kernel_add_fuse_impl<8>;
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
 
-kernel void kernel_sub_fuse_1(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    const int i03 = tgpig.z;
-    const int i02 = tgpig.y;
-    const int i01 = tgpig.x;
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
 
-    const int i13 = i03%args.ne13;
-    const int i12 = i02%args.ne12;
-    const int i11 = i01%args.ne11;
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
 
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
-
-    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-        const int i10 = i0%args.ne10;
-        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) - *((device float *)(src1_ptr + i10*args.nb10));
-    }
-}
-
-kernel void kernel_mul_fuse_1(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    const int i03 = tgpig.z;
-    const int i02 = tgpig.y;
-    const int i01 = tgpig.x;
-
-    const int i13 = i03%args.ne13;
-    const int i12 = i02%args.ne12;
-    const int i11 = i01%args.ne11;
-
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
-
-    if (args.ne10 == 1) {
-        const float x = *((device float *)(src1_ptr));
-        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * x;
+            dst_row[i0] = res;
         }
     } else {
-        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-            const int i10 = i0%args.ne10;
-            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * *((device float *)(src1_ptr + i10*args.nb10));
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = i0%args.ne10;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = i0%args.ne10;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
         }
     }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
 }
 
-kernel void kernel_div_fuse_1(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    const int i03 = tgpig.z;
-    const int i02 = tgpig.y;
-    const int i01 = tgpig.x;
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
 
-    const int i13 = i03%args.ne13;
-    const int i12 = i02%args.ne12;
-    const int i11 = i01%args.ne11;
-
-    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
-    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + args.o1[0];
-    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
-
-    if (args.ne10 == 1) {
-        const float x = 1.0f / *((device float *)(src1_ptr));
-        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * x;
-        }
-    } else {
-        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
-            const int i10 = i0%args.ne10;
-            *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) / *((device float *)(src1_ptr + i10*args.nb10));
-        }
-    }
-}
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
 
 kernel void kernel_add_id(
         constant ggml_metal_kargs_add_id & args,
@@ -1057,7 +1073,7 @@ kernel void kernel_add_id(
     const size_t nb1 = args.ne0 * sizeof(float);
     const size_t nb2 = args.ne1 * nb1;
 
-    device       float * dst_row  = (device       float *)((device char *)dst + i1*nb1 + i2*nb2);
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
     device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
     device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
 
@@ -1098,141 +1114,6 @@ template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat
 template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
 template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
 
-// assumption: src1 is a row
-// broadcast src1 into src0
-template <short F>
-kernel void kernel_add_row_c4_fuse_impl(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint tpig[[thread_position_in_grid]]) {
-    const uint nb = args.ne00/4;
-    const uint i  = tpig % nb;
-
-    device const float4 * src0_row = (device const float4 *) (src0);
-    device       float4 *  dst_row = (device       float4 *) (dst);
-
-    float4 res = src0_row[tpig];
-
-#pragma unroll(F)
-    for (short j = 0; j < F; ++j) {
-        res += ((device const float4 *) (src1 + args.o1[j]))[i];
-    }
-
-    dst_row[tpig] = res;
-}
-
-typedef decltype(kernel_add_row_c4_fuse_impl<1>) kernel_add_row_c4_fuse_t;
-
-template [[host_name("kernel_add_row_c4_fuse_1")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<1>;
-template [[host_name("kernel_add_row_c4_fuse_2")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<2>;
-template [[host_name("kernel_add_row_c4_fuse_3")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<3>;
-template [[host_name("kernel_add_row_c4_fuse_4")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<4>;
-template [[host_name("kernel_add_row_c4_fuse_5")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<5>;
-template [[host_name("kernel_add_row_c4_fuse_6")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<6>;
-template [[host_name("kernel_add_row_c4_fuse_7")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<7>;
-template [[host_name("kernel_add_row_c4_fuse_8")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<8>;
-
-template <short F>
-kernel void kernel_sub_row_c4_fuse_impl(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint tpig[[thread_position_in_grid]]) {
-
-    const uint nb = args.ne00/4;
-    const uint i  = tpig % nb;
-
-    device const float4 * src0_row = (device const float4 *) (src0);
-    device       float4 *  dst_row = (device       float4 *) (dst);
-
-    device const float4 * src1_row[F];
-    for (short j = 0; j < F; ++j) {
-        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
-    }
-
-    float4 res = src0_row[tpig];
-
-#pragma unroll(F)
-    for (short j = 0; j < F; ++j) {
-        res -= src1_row[j][i];
-    }
-
-    dst_row[tpig] = res;
-}
-
-typedef decltype(kernel_sub_row_c4_fuse_impl<1>) kernel_sub_row_c4_fuse_t;
-
-template [[host_name("kernel_sub_row_c4_fuse_1")]] kernel kernel_sub_row_c4_fuse_t kernel_sub_row_c4_fuse_impl<1>;
-
-template <short F>
-kernel void kernel_mul_row_c4_fuse_impl(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint tpig[[thread_position_in_grid]]) {
-
-    const uint nb = args.ne00/4;
-    const uint i  = tpig % nb;
-
-    device const float4 * src0_row = (device const float4 *) (src0);
-    device       float4 *  dst_row = (device       float4 *) (dst);
-
-    device const float4 * src1_row[F];
-    for (short j = 0; j < F; ++j) {
-        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
-    }
-
-    float4 res = src0_row[tpig];
-
-#pragma unroll(F)
-    for (short j = 0; j < F; ++j) {
-        res *= src1_row[j][i];
-    }
-
-    dst_row[tpig] = res;
-}
-
-typedef decltype(kernel_mul_row_c4_fuse_impl<1>) kernel_mul_row_c4_fuse_t;
-
-template [[host_name("kernel_mul_row_c4_fuse_1")]] kernel kernel_mul_row_c4_fuse_t kernel_mul_row_c4_fuse_impl<1>;
-
-template <short F>
-kernel void kernel_div_row_c4_fuse_impl(
-        constant ggml_metal_kargs_bin & args,
-        device const char * src0,
-        device const char * src1,
-        device       char * dst,
-        uint tpig[[thread_position_in_grid]]) {
-
-    const uint nb = args.ne00/4;
-    const uint i  = tpig % nb;
-
-    device const float4 * src0_row = (device const float4 *) (src0);
-    device       float4 *  dst_row = (device       float4 *) (dst);
-
-    device const float4 * src1_row[F];
-    for (short j = 0; j < F; ++j) {
-        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
-    }
-
-    float4 res = src0_row[tpig];
-
-#pragma unroll(F)
-    for (short j = 0; j < F; ++j) {
-        res /= src1_row[j][i];
-    }
-
-    dst_row[tpig] = res;
-}
-
-typedef decltype(kernel_div_row_c4_fuse_impl<1>) kernel_div_row_c4_fuse_t;
-
-template [[host_name("kernel_div_row_c4_fuse_1")]] kernel kernel_div_row_c4_fuse_t kernel_div_row_c4_fuse_impl<1>;
-
 kernel void kernel_scale_f32(
         constant ggml_metal_kargs_scale & args,
         device const float * src0,
@@ -2737,6 +2618,83 @@ kernel void kernel_rwkv_wkv7_f32(
     }
 }
 
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t ne02 = args.ne02;
+    const int32_t ne03 = args.ne03;
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}
+
 kernel void kernel_argmax_f32(
         constant ggml_metal_kargs_argmax & args,
         device   const char * src0,
@@ -5208,6 +5166,7 @@ constant int32_t FC_flash_attn_ext_blk_ncpsg [[function_constant(FC_FLASH_ATTN_E
 // scan the blocks of the mask that are not masked
 // 0 -     masked (i.e. full of -INF, skip)
 // 1 - not masked (i.e. at least one element of the mask is not -INF)
+// 2 - all zero
 kernel void kernel_flash_attn_ext_blk(
         constant ggml_metal_kargs_flash_attn_ext_blk & args,
         device const char * mask,
@@ -5229,27 +5188,29 @@ kernel void kernel_flash_attn_ext_blk(
 
     device const half * mask_src = (device const half *) (mask + (i1*Q)*args.nb31 + i2*args.nb32 + i3*args.nb33) + i0*C + tiisg;
 
-    // fast route
-    if (res == 0) {
-        if (simd_max(*mask_src) > -MAXHALF/2) {
-            res = 1;
-        }
-    }
-
     // detailed check of the elements of the block
     if ((C > NW || Q > 1) && res == 0) {
-        half m = -MAXHALF;
+        half mmin =  MAXHALF;
+        half mmax = -MAXHALF;
 
         FOR_UNROLL (short j = 0; j < Q; ++j) {
             FOR_UNROLL (short ii = 0; ii < C/NW; ++ii) {
-                m = max(m, mask_src[ii*NW]);
+                mmin = min(mmin, mask_src[ii*NW]);
+                mmax = max(mmax, mask_src[ii*NW]);
             }
 
             mask_src += args.nb31/2;
         }
 
-        if (simd_max(m) > -MAXHALF/2) {
-            res = 1;
+        mmin = simd_min(mmin);
+        mmax = simd_max(mmax);
+
+        if (mmax > -MAXHALF) {
+            if (mmin == 0.0 && mmax == 0.0) {
+                res = 2;
+            } else {
+                res = 1;
+            }
         }
     }
 
@@ -5491,9 +5452,13 @@ void kernel_flash_attn_ext_impl(
                 ic = 0;
             }
 
+            char blk_cur = 1;
+
             // read the mask into shared mem
             if (FC_flash_attn_ext_has_mask) {
-                if (blk[ic0] == 0) {
+                blk_cur = blk[ic0];
+
+                if (blk_cur == 0) {
                     FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
                         pm2[jj] += NW;
                     }
@@ -5501,16 +5466,22 @@ void kernel_flash_attn_ext_impl(
                     continue;
                 }
 
-                FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
-                    const short j = jj*NSG + sgitg;
+                if (blk_cur == 1) {
+                    FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
+                        const short j = jj*NSG + sgitg;
 
-                    if (FC_flash_attn_ext_bc_mask) {
-                        sm2[j*SH + tiisg] = (iq1 + j) < args.ne31 ? pm2[jj][tiisg] : half2(-MAXHALF, -MAXHALF);
-                    } else {
-                        sm2[j*SH + tiisg] = pm2[jj][tiisg];
+                        if (FC_flash_attn_ext_bc_mask) {
+                            sm2[j*SH + tiisg] = (iq1 + j) < args.ne31 ? pm2[jj][tiisg] : half2(-MAXHALF, -MAXHALF);
+                        } else {
+                            sm2[j*SH + tiisg] = pm2[jj][tiisg];
+                        }
+
+                        pm2[jj] += NW;
+                    }
+                } else if (blk_cur == 2) {
+                    FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
+                        pm2[jj] += NW;
                     }
-
-                    pm2[jj] += NW;
                 }
 
 #if 0
@@ -5675,10 +5646,12 @@ void kernel_flash_attn_ext_impl(
                 }
 
                 // mqk = mqk + slope*mask
-                if (FC_flash_attn_ext_has_bias) {
-                    s2 += s2_t(sm2[j*SH + tiisg])*slope;
-                } else {
-                    s2 += s2_t(sm2[j*SH + tiisg]);
+                if (blk_cur != 2) {
+                    if (FC_flash_attn_ext_has_bias) {
+                        s2 += s2_t(sm2[j*SH + tiisg])*slope;
+                    } else {
+                        s2 += s2_t(sm2[j*SH + tiisg]);
+                    }
                 }
 
                 M[jj] = simd_max(max(M[jj], max(s2[0], s2[1])));
@@ -5931,7 +5904,7 @@ template<
     void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
     short DK,         // K head size
     short DV,         // V head size
-    short Q  = OP_FLASH_ATTN_EXT_NQPTG, // queries per threadgroup
+    short Q  = OP_FLASH_ATTN_EXT_NQPSG, // queries per threadgroup
     short C  = OP_FLASH_ATTN_EXT_NCPSG> // cache items per threadgroup
 kernel void kernel_flash_attn_ext(
         constant ggml_metal_kargs_flash_attn_ext & args,
@@ -6141,11 +6114,10 @@ template<
     void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
     short DK,       // K head size
     short DV,       // V head size
-    short NE,       // head elements per thread
-    short Q,        // queries per threadgroup
-    short C,        // cache items per threadgroup
-    short NSG>      // number of simd groups
-void kernel_flash_attn_ext_vec_impl(
+    short NE = 4,   // head elements per thread
+    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPSG,  // queries per threadgroup
+    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
+kernel void kernel_flash_attn_ext_vec(
         constant ggml_metal_kargs_flash_attn_ext_vec & args,
         device const char * q,
         device const char * k,
@@ -6162,6 +6134,7 @@ void kernel_flash_attn_ext_vec_impl(
     static_assert(DV % 32 == 0, "DV must be divisible by 32");
 
 #define NWG  (FC_flash_attn_ext_vec_nwg)
+#define NSG  (FC_flash_attn_ext_vec_nsg)
 
 #define NS10 (FC_flash_attn_ext_vec_ns10)
 #define NS20 (FC_flash_attn_ext_vec_ns20)
@@ -6190,12 +6163,12 @@ void kernel_flash_attn_ext_vec_impl(
 
     const short T = PK + NSG*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*PK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*PK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*PK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*PK); // same as above but in s4_t
-    threadgroup half  * sm  = (threadgroup half  *) (shmem_f16 +   sgitg*SH + 2*C + Q*PK); // scratch buffer for mask
-    threadgroup o4_t  * so4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*PV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                      0*PK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                      0*PK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + NSG*PK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + NSG*PK); // same as above but in s4_t
+    threadgroup half  * sm  = (threadgroup half  *) (shmem_f16 +   sgitg*SH + 2*C + NSG*PK); // scratch buffer for mask
+    threadgroup o4_t  * so4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*PV       + NSG*PK + NSG*SH); // scratch buffer for the results
 
     // store the result for all queries in shared memory (the O matrix from the paper)
     so4 += tiisg;
@@ -6213,11 +6186,13 @@ void kernel_flash_attn_ext_vec_impl(
     // load heads from Q to shared memory
     device const float4 * q4 = (device const float4 *) ((device const char *) q);
 
-    for (short i = tiisg; i < PK4; i += NW) {
-        if (iq1 < args.ne01 && i < DK4) {
-            sq4[i] = (q4_t) q4[i];
-        } else {
-            sq4[i] = (q4_t) 0.0f;
+    if (iq1 < args.ne01) {
+        for (short i = tiisg; i < PK4; i += NW) {
+            if (i < DK4) {
+                sq4[i] = (q4_t) q4[i];
+            } else {
+                sq4[i] = (q4_t) 0.0f;
+            }
         }
     }
 
@@ -6295,7 +6270,7 @@ void kernel_flash_attn_ext_vec_impl(
             }
 
             // skip -INF blocks
-            if (simd_max(sm[tiisg]) == -INFINITY) {
+            if (simd_max(sm[tiisg]) <= -MAXHALF) {
                 continue;
             }
 
@@ -6569,57 +6544,11 @@ void kernel_flash_attn_ext_vec_impl(
     }
 
 #undef NWG
+#undef NSG
 #undef NS10
 #undef NS20
 }
 
-template<
-    typename q4_t,  // query types in shared memory
-    typename k4_t,  // key types in shared memory
-    typename v4_t,  // value types in shared memory
-    typename qk_t,  // Q*K types
-    typename s_t,   // soft-max types
-    typename s4_t,
-    typename o4_t,  // attention accumulation types
-    typename kd4_t, // key type in device memory
-    short nl_k,
-    void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &),
-    typename vd4_t, // value type in device memory
-    short nl_v,
-    void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
-    short DK,       // K head size
-    short DV,       // V head size
-    short NE = 4,   // head elements per thread
-    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPTG,  // queries per threadgroup
-    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
-kernel void kernel_flash_attn_ext_vec(
-        constant ggml_metal_kargs_flash_attn_ext_vec & args,
-        device const char * q,
-        device const char * k,
-        device const char * v,
-        device const char * mask,
-        device const char * sinks,
-        device const char * pad,
-        device       char * dst,
-        threadgroup  half * shmem_f16 [[threadgroup(0)]],
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort  tiisg[[thread_index_in_simdgroup]],
-        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
-#define FWD_TMPL q4_t, k4_t, v4_t, qk_t, s_t, s4_t, o4_t, kd4_t, nl_k, deq_k_t4, vd4_t, nl_v, deq_v_t4, DK, DV, NE, Q, C
-#define FWD_ARGS args, q, k, v, mask, sinks, pad, dst, shmem_f16, tgpig, tiisg, sgitg
-    switch (FC_flash_attn_ext_vec_nsg) {
-      // note: disabled cases to reduce library load time
-        case 1:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  1>(FWD_ARGS); break;
-        case 2:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  2>(FWD_ARGS); break;
-        case 4:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  4>(FWD_ARGS); break;
-      //case 8:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  8>(FWD_ARGS); break;
-      //case 16: kernel_flash_attn_ext_vec_impl<FWD_TMPL, 16>(FWD_ARGS); break;
-      //case 32: kernel_flash_attn_ext_vec_impl<FWD_TMPL, 32>(FWD_ARGS); break;
-    }
-#undef FWD_TMPL
-#undef FWD_ARGS
-}
-
 // note: I think the s_t can be half instead of float, because the Q*K scaling is done before storing to shared mem
 //       in the other (non-vec) kernel, we need s_t to also be float because we scale during the soft_max
 //
@@ -8782,6 +8711,26 @@ kernel void kernel_set_rows_f(
     }
 }
 
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
 constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
 constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -101,6 +101,8 @@ set(GGML_OPENCL_KERNELS
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
     mul_mm_q8_0_f32_l4_lm
+    mul_mm_q8_0_f32_8x4
+    gemv_noshuffle_general_q8_0_f32
     mul
     norm
     relu

ggml/src/ggml-opencl/kernels/concat.cl

@@ -1,109 +1,51 @@
-kernel void kernel_concat_f32_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
-    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
-    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
-    int dim
+kernel void kernel_concat_f32(
+    global  const char * src0,
+    ulong                offset0,
+    global  const char * src1,
+    ulong                offset1,
+    global        char * dst,
+    ulong                offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb00,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim
 ) {
-    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
-    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
-    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
-    int i0 = get_global_id(0); // Index along dst's 0th dimension
-    int i1 = get_global_id(1); // Index along dst's 1st dimension
-    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
 
-    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
-        return;
-    }
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
-    ulong src_idx;
+    global const float * x;
 
-    if (dim == 0) {
-        if (i0 < d_ne00) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
-            dst[dst_idx] = src1[src_idx];
-        }
-    } else if (dim == 1) {
-        if (i1 < d_ne01) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
-        }
-    } else if (dim == 2) {
-        if (i2 < d_ne02) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-
-            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
-        }
-    }
-}
-
-kernel void kernel_concat_f32_non_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-
-    long ne00, long ne01, long ne02, long ne03,
-    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
-
-    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
-
-    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
-    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
-    int dim
-) {
-    global const char * src0_base = p_src0 + off_src0;
-    global const char * src1_base = p_src1 + off_src1;
-    global char * dst_base        = p_dst + off_dst;
-
-    long current_i1 = get_global_id(0); // Index for dst_dim_1
-    long current_i2 = get_global_id(1); // Index for dst_dim_2
-    long current_i3 = get_global_id(2); // Index for dst_dim_3
-
-    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
-        return;
-    }
-
-    global const float * x_val_ptr;
-    global float * y_val_ptr;
-
-    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
-        bool use_src0;
-        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
-
-        if (dim == 0) {
-            use_src0 = (current_i0 < ne00);
-            if (!use_src0) { s_i0 = current_i0 - ne00; }
-        } else if (dim == 1) {
-            use_src0 = (current_i1 < ne01);
-            if (!use_src0) { s_i1 = current_i1 - ne01; }
-        } else if (dim == 2) {
-            use_src0 = (current_i2 < ne02);
-            if (!use_src0) { s_i2 = current_i2 - ne02; }
-        } else { // dim == 3
-            use_src0 = (current_i3 < ne03);
-            if (!use_src0) { s_i3 = current_i3 - ne03; }
-        }
-
-        if (use_src0) {
-            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
         } else {
-            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
         }
 
-        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
-        *y_val_ptr = *x_val_ptr;
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
     }
 }