common : Fix corrupted memory error on json grammar initialization (#16038)

Initalizing RESERVED_NAME in is_reserved_name() is not thread
safe and leads to corrupted memory when used from multiple threads
as can be seen in the asan trace below. This fixes the initialization
to make it thread-safe.

    #0 0x000100abd018 in std::__1::pair<std::__1::__hash_iterator<std::__1::__hash_node<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, void*>*>, bool> std::__1::__hash_table<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::hash<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::equal_to<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::allocator<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>>::__emplace_unique_key_args<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&>(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&) __hash_table:1565
    #1 0x000100ab0320 in SchemaConverter::visit(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&) json-schema-to-grammar.cpp:802
    #2 0x000100aafc48 in std::__1::__function::__func<build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&)::$_2, std::__1::allocator<build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&)::$_2>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> (std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>::operator()(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&) function.h:319
    #3 0x000100a2c938 in std::__1::__function::__func<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0::operator()(common_grammar_builder const&) const::'lambda'(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&), std::__1::allocator<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0::operator()(common_grammar_builder const&) const::'lambda'(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>, void (nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>::operator()(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&) function.h:319
    #4 0x000100a139f8 in foreach_function(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&, std::__1::function<void (nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)> const&) chat.cpp:762
    #5 0x000100a2a7f4 in std::__1::__function::__func<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0, std::__1::allocator<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0>, void (common_grammar_builder const&)>::operator()(common_grammar_builder const&) function.h:319
    #6 0x000100aa98f4 in build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&) json-schema-to-grammar.cpp:982
    #7 0x0001009c9314 in common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool) chat.cpp:1110
    #8 0x0001009b8afc in common_chat_templates_apply_jinja(common_chat_templates const*, common_chat_templates_inputs const&) chat.cpp:1992
    #9 0x0001009b533c in common_chat_templates_apply(common_chat_templates const*, common_chat_templates_inputs const&) chat.cpp:2074
    #10 0x000100810120 in llamacpp_apply_chat_template+0x724 (predict_oai-98384e17fb94e863:arm64+0x100090120)
    ...

==45482==Register values:
 x[0] = 0x00006020004147f8   x[1] = 0x00006080000013c8   x[2] = 0x0000000000000000   x[3] = 0x0000604006289738
 x[4] = 0x0000000000000002   x[5] = 0x0000000000000001   x[6] = 0x04034000004b4000   x[7] = 0x0000000000000001
 x[8] = 0xbebebebebebebebe   x[9] = 0x17d7d7d7d7d7d7d7  x[10] = 0x00000c04000828ff  x[11] = 0x0000000000000001
x[12] = 0x000000002018d383  x[13] = 0x0000000000000000  x[14] = 0xfa0000000000fafa  x[15] = 0x000010700001ffff
x[16] = 0x000000019dc012c0  x[17] = 0x00000001021284f8  x[18] = 0x0000000000000000  x[19] = 0x00000001700acdc0
x[20] = 0x0000000000000002  x[21] = 0x000000002018d384  x[22] = 0x16dd16fd2e731151  x[23] = 0x0000007000020000
x[24] = 0x0000000100c69c08  x[25] = 0x0000000100c69c20  x[26] = 0x00006080000013c7  x[27] = 0x0000000100c69c00
x[28] = 0x00000001700acd60     fp = 0x00000001700aceb0     lr = 0x0000000100abce30     sp = 0x00000001700acd60
AddressSanitizer can not provide additional info.
SUMMARY: AddressSanitizer: SEGV __hash_table:1565 in std::__1::pair<std::__1::__hash_iterator<std::__1::__hash_node<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, void*>*>, bool> std::__1::__hash_table<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::hash<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::equal_to<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::allocator<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>>::__emplace_unique_key_args<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&>(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&)
Thread T5 created by T0 here:
    #0 0x0001020b99d4 in pthread_create+0x5c (libclang_rt.asan_osx_dynamic.dylib:arm64e+0x359d4)
    #1 0x000100873910 in std::sys::pal::unix::thread::Thread::new::h77254fdd87a28e05+0x118 (predict_oai-98384e17fb94e863:arm64+0x1000f3910)
    #2 0x0001007c7a1c in test::run_test::haeb3c2bcd5ed6cf6+0x76c (predict_oai-98384e17fb94e863:arm64+0x100047a1c)
    #3 0x0001007aedb0 in test::console::run_tests_console::he9d142d704f3a986+0x149c (predict_oai-98384e17fb94e863:arm64+0x10002edb0)
    #4 0x0001007c5758 in test::test_main::hf86a5e20735245b9+0x118 (predict_oai-98384e17fb94e863:arm64+0x100045758)
    #5 0x0001007c5da0 in test::test_main_static::h61ee9c8fd30abca0+0x54 (predict_oai-98384e17fb94e863:arm64+0x100045da0)
    ...

==45482==ABORTING

.clang-format

@@ -22,6 +22,13 @@ AllowShortIfStatementsOnASingleLine: Never
 AllowShortLambdasOnASingleLine: Inline
 AllowShortLoopsOnASingleLine: false
 AlwaysBreakBeforeMultilineStrings: true
+# Treat CUDA keywords/attributes as "attribute macros" and avoid breaking lines inside them
+AttributeMacros:
+  - __host__
+  - __device__
+  - __global__
+  - __forceinline__
+  - __launch_bounds__
 BinPackArguments: true
 BinPackParameters: false # OnePerLine
 BitFieldColonSpacing: Both

.devops/rocm.Dockerfile

@@ -4,7 +4,7 @@ ARG UBUNTU_VERSION=24.04
 ARG ROCM_VERSION=6.4
 ARG AMDGPU_VERSION=6.4
 
-# Target the CUDA build image
+# Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
 ### Build image
@@ -15,16 +15,13 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # This is mostly tied to rocBLAS supported archs.
 # gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
 # gfx906 is deprecated
-#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
+#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
 
-ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
-#ARG ROCM_DOCKER_ARCH=gfx1100
+ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
+#ARG ROCM_DOCKER_ARCH='gfx1151'
 
-# Set nvcc architectured
+# Set ROCm architectures
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
-# Enable ROCm
-# ENV CC=/opt/rocm/llvm/bin/clang
-# ENV CXX=/opt/rocm/llvm/bin/clang++
 
 RUN apt-get update \
     && apt-get install -y \
@@ -39,8 +36,16 @@ WORKDIR /app
 
 COPY . .
 
+RUN git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
+    cmake -S . -B build \
+        -DGGML_HIP=ON \
+        -DGGML_HIP_ROCWMMA_FATTN=ON \
+        -DCMAKE_HIP_FLAGS="-I$(pwd)/rocwmma/library/include/" \
+        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
+        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
+        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.github/workflows/build.yml

@@ -56,7 +56,7 @@ env:
 
 jobs:
   macOS-latest-cmake-arm64:
-    runs-on: macos-14
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -88,6 +88,7 @@ jobs:
             -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
 
       - name: Test
         id: cmake_test
@@ -126,7 +127,8 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -136,7 +138,7 @@ jobs:
           ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-arm64-webgpu:
-    runs-on: macos-14
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -709,6 +711,7 @@ jobs:
 
   macOS-latest-swift:
     runs-on: macos-latest
+    needs: ios-xcode-build
 
     strategy:
       matrix:
@@ -725,6 +728,12 @@ jobs:
           key: macOS-latest-swift
           evict-old-files: 1d
 
+      - name: Download xcframework artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+
       - name: Dependencies
         id: depends
         continue-on-error: true
@@ -746,11 +755,6 @@ jobs:
             -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          ./build-xcframework.sh
-
   windows-msys2:
     runs-on: windows-2025
 
@@ -1050,9 +1054,13 @@ jobs:
         run:  examples/sycl/win-build-sycl.bat
 
   windows-latest-cmake-hip:
-    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-2022
 
+    env:
+      # The ROCm version must correspond to the version used in the HIP SDK.
+      ROCM_VERSION: "6.4.2"
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
+
     steps:
       - name: Clone
         id: checkout
@@ -1061,23 +1069,46 @@ jobs:
       - name: Clone rocWMMA repository
         id: clone_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+          git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
 
-      - name: Install
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
         id: verify
         run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1141,8 +1172,17 @@ jobs:
         run: |
           ./build-xcframework.sh
 
+      - name: Upload xcframework artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+          retention-days: 1
+
       - name: Build Xcode project
-        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+        run: |
+          xcodebuild -downloadPlatform iOS
+          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
 
   android-build:
     runs-on: ubuntu-latest

.github/workflows/close-issue.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - uses: actions/stale@v5
         with:
-          exempt-issue-labels: "refactoring,help wanted,good first issue,research,bug,roadmap"
+          exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
           days-before-issue-stale: 30
           days-before-issue-close: 14
           stale-issue-label: "stale"

.github/workflows/release.yml

@@ -108,7 +108,8 @@ jobs:
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Determine tag name
@@ -528,11 +529,14 @@ jobs:
   windows-hip:
     runs-on: windows-2022
 
+    env:
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
+
     strategy:
       matrix:
         include:
           - name: "radeon"
-            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
@@ -542,29 +546,52 @@ jobs:
       - name: Clone rocWMMA repository
         id: clone_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+          git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
+          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
           evict-old-files: 1d
 
-      - name: Install
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
         id: verify
         run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: Build
         id: cmake_build
@@ -585,9 +612,12 @@ jobs:
             -DLLAMA_CURL=OFF
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
+          md "build\bin\hipblaslt\library"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
 
       - name: Pack artifacts
         id: pack_artifacts

CMakeLists.txt

@@ -58,6 +58,12 @@ if (MSVC)
     add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
 endif()
 
+if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
+    set(LLAMA_TOOLS_INSTALL_DEFAULT OFF)
+else()
+    set(LLAMA_TOOLS_INSTALL_DEFAULT ${LLAMA_STANDALONE})
+endif()
+
 #
 # option list
 #
@@ -82,6 +88,7 @@ option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 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})
 
 # 3rd party libs
 option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)

CONTRIBUTING.md

@@ -16,6 +16,9 @@
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
 - Consider adding yourself to [CODEOWNERS](CODEOWNERS)
+- Let authors, who are also collaborators, merge their own PRs
+- When merging a PR by a contributor, make sure you have a good understanding of the changes
+- Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 
 # Coding guidelines
 

ci/run.sh

@@ -270,7 +270,9 @@ function gg_run_ctest_with_model_debug {
     local model; model=$(gg_get_model)
     cd build-ci-debug
     set -e
+
     (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
     set +e
     cd ..
 }
@@ -281,7 +283,15 @@ function gg_run_ctest_with_model_release {
     local model; model=$(gg_get_model)
     cd build-ci-release
     set -e
+
     (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    # test memory leaks
+    #if [[ ! -z ${GG_BUILD_METAL} ]]; then
+    #    # TODO: this hangs for some reason ...
+    #    (time leaks -quiet -atExit -- ./bin/test-thread-safety -m $model --parallel 2 -t 2 -p "hello") 2>&1 | tee -a $OUT/${ci}-leaks.log
+    #fi
+
     set +e
     cd ..
 }
@@ -860,10 +870,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-if [ -z ${GG_BUILD_SYCL} ]; then
-    # SYCL build breaks with debug build flags
-    test $ret -eq 0 && gg_run ctest_debug
-fi
+test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -871,9 +878,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run test_scripts_debug
-        fi
+        test $ret -eq 0 && gg_run test_scripts_debug
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -884,9 +889,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
             test $ret -eq 0 && gg_run pythia_2_8b
             #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run ctest_with_model_debug
-        fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi

common/arg.cpp

@@ -745,6 +745,124 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 
 #endif // LLAMA_USE_CURL
 
+//
+// Docker registry functions
+//
+
+static std::string common_docker_get_token(const std::string & repo) {
+    std::string url = "https://auth.docker.io/token?service=registry.docker.io&scope=repository:" + repo + ":pull";
+
+    common_remote_params params;
+    auto                 res = common_remote_get_content(url, params);
+
+    if (res.first != 200) {
+        throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
+    }
+
+    std::string            response_str(res.second.begin(), res.second.end());
+    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
+
+    if (!response.contains("token")) {
+        throw std::runtime_error("Docker registry token response missing 'token' field");
+    }
+
+    return response["token"].get<std::string>();
+}
+
+static std::string common_docker_resolve_model(const std::string & docker) {
+    // Parse ai/smollm2:135M-Q4_K_M
+    size_t      colon_pos = docker.find(':');
+    std::string repo, tag;
+    if (colon_pos != std::string::npos) {
+        repo = docker.substr(0, colon_pos);
+        tag  = docker.substr(colon_pos + 1);
+    } else {
+        repo = docker;
+        tag  = "latest";
+    }
+
+    // ai/ is the default
+    size_t      slash_pos = docker.find('/');
+    if (slash_pos == std::string::npos) {
+        repo.insert(0, "ai/");
+    }
+
+    LOG_INF("%s: Downloading Docker Model: %s:%s\n", __func__, repo.c_str(), tag.c_str());
+    try {
+        // --- helper: digest validation ---
+        auto validate_oci_digest = [](const std::string & digest) -> std::string {
+            // Expected: algo:hex ; start with sha256 (64 hex chars)
+            // You can extend this map if supporting other algorithms in future.
+            static const std::regex re("^sha256:([a-fA-F0-9]{64})$");
+            std::smatch m;
+            if (!std::regex_match(digest, m, re)) {
+                throw std::runtime_error("Invalid OCI digest format received in manifest: " + digest);
+            }
+            // normalize hex to lowercase
+            std::string normalized = digest;
+            std::transform(normalized.begin()+7, normalized.end(), normalized.begin()+7, [](unsigned char c){
+                return std::tolower(c);
+            });
+            return normalized;
+        };
+
+        std::string token = common_docker_get_token(repo);  // Get authentication token
+
+        // Get manifest
+        const std::string    url_prefix = "https://registry-1.docker.io/v2/" + repo;
+        std::string          manifest_url = url_prefix + "/manifests/" + tag;
+        common_remote_params manifest_params;
+        manifest_params.headers.push_back("Authorization: Bearer " + token);
+        manifest_params.headers.push_back(
+            "Accept: application/vnd.docker.distribution.manifest.v2+json,application/vnd.oci.image.manifest.v1+json");
+        auto manifest_res = common_remote_get_content(manifest_url, manifest_params);
+        if (manifest_res.first != 200) {
+            throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
+        }
+
+        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
+        std::string            gguf_digest;  // Find the GGUF layer
+        if (manifest.contains("layers")) {
+            for (const auto & layer : manifest["layers"]) {
+                if (layer.contains("mediaType")) {
+                    std::string media_type = layer["mediaType"].get<std::string>();
+                    if (media_type == "application/vnd.docker.ai.gguf.v3" ||
+                        media_type.find("gguf") != std::string::npos) {
+                        gguf_digest = layer["digest"].get<std::string>();
+                        break;
+                    }
+                }
+            }
+        }
+
+        if (gguf_digest.empty()) {
+            throw std::runtime_error("No GGUF layer found in Docker manifest");
+        }
+
+        // Validate & normalize digest
+        gguf_digest = validate_oci_digest(gguf_digest);
+        LOG_DBG("%s: Using validated digest: %s\n", __func__, gguf_digest.c_str());
+
+        // Prepare local filename
+        std::string model_filename = repo;
+        std::replace(model_filename.begin(), model_filename.end(), '/', '_');
+        model_filename += "_" + tag + ".gguf";
+        std::string local_path = fs_get_cache_file(model_filename);
+
+        const std::string blob_url = url_prefix + "/blobs/" + gguf_digest;
+        if (!common_download_file_single(blob_url, local_path, token, false)) {
+            throw std::runtime_error("Failed to download Docker Model");
+        }
+
+        LOG_INF("%s: Downloaded Docker Model to: %s\n", __func__, local_path.c_str());
+        return local_path;
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: Docker Model download failed: %s\n", __func__, e.what());
+        throw;
+    }
+}
+
 //
 // utils
 //
@@ -795,7 +913,9 @@ static handle_model_result common_params_handle_model(
     handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
-        if (!model.hf_repo.empty()) {
+        if (!model.docker_repo.empty()) {  // Handle Docker URLs by resolving them to local paths
+            model.path = common_docker_resolve_model(model.docker_repo);
+        } else if (!model.hf_repo.empty()) {
             // short-hand to avoid specifying --hf-file -> default it to --model
             if (model.hf_file.empty()) {
                 if (model.path.empty()) {
@@ -1184,7 +1304,7 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     } else {
         for (const auto & device : dev_names) {
             auto * dev = ggml_backend_dev_by_name(device.c_str());
-            if (!dev || ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+            if (!dev || ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
                 throw std::invalid_argument(string_format("invalid device: %s", device.c_str()));
             }
             devices.push_back(dev);
@@ -1194,7 +1314,7 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
-static void add_rpc_devices(std::string servers) {
+static void add_rpc_devices(const std::string & servers) {
     auto rpc_servers = string_split<std::string>(servers, ',');
     if (rpc_servers.empty()) {
         throw std::invalid_argument("no RPC servers specified");
@@ -1263,6 +1383,18 @@ static std::string list_builtin_chat_templates() {
     return msg.str();
 }
 
+static bool is_truthy(const std::string & value) {
+    return value == "on" || value == "enabled" || value == "1";
+}
+
+static bool is_falsey(const std::string & value) {
+    return value == "off" || value == "disabled" || value == "0";
+}
+
+static bool is_autoy(const std::string & value) {
+    return value == "auto" || value == "-1";
+}
+
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     // load dynamic backends
     ggml_backend_load_all();
@@ -1544,21 +1676,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_chunks = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL}));
-    add_opt(common_arg(
-        {"-fa", "--flash-attn"}, "FA",
-        string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')", llama_flash_attn_type_name(params.flash_attn_type)),
-        [](common_params & params, const std::string & value) {
-            if (value == "on" || value == "enabled" || value == "1") {
-                params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
-            } else if (value == "off" || value == "disabled" || value == "0") {
-                params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
-            } else if (value == "auto" || value == "-1") {
-                params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
-            } else {
-                throw std::runtime_error(string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
-            }
-        }
-    ).set_env("LLAMA_ARG_FLASH_ATTN"));
+    add_opt(common_arg({ "-fa", "--flash-attn" }, "[on|off|auto]",
+                       string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')",
+                                     llama_flash_attn_type_name(params.flash_attn_type)),
+                       [](common_params & params, const std::string & value) {
+                           if (is_truthy(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
+                           } else if (is_falsey(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
+                           } else if (is_autoy(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
+                           } else {
+                               throw std::runtime_error(
+                                   string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
+                           }
+                       }).set_env("LLAMA_ARG_FLASH_ATTN"));
     add_opt(common_arg(
         {"-p", "--prompt"}, "PROMPT",
         "prompt to start generation with; for system message, use -sys",
@@ -1572,7 +1704,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.system_prompt = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_DIFFUSION}));
     add_opt(common_arg(
         {"--no-perf"},
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@@ -2384,24 +2516,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--list-devices"},
         "print list of available devices and exit",
         [](common_params &) {
-            std::vector<ggml_backend_dev_t> rpc_devices;
-            std::vector<ggml_backend_dev_t> all_devices;
+            std::vector<ggml_backend_dev_t> devices;
             for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
                 auto * dev = ggml_backend_dev_get(i);
-                if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
-                    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
-                    if (ggml_backend_reg_name(reg) == std::string("RPC")) {
-                        rpc_devices.push_back(dev);
-                    } else {
-                        all_devices.push_back(dev);
-                    }
+                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+                    devices.push_back(dev);
                 }
             }
-            // insert RPC devices in front
-            all_devices.insert(all_devices.begin(), rpc_devices.begin(), rpc_devices.end());
             printf("Available devices:\n");
-            for (size_t i = 0; i < all_devices.size(); ++i) {
-                auto * dev = all_devices[i];
+            for (auto * dev : devices) {
                 size_t free, total;
                 ggml_backend_dev_memory(dev, &free, &total);
                 printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
@@ -2425,7 +2548,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--cpu-moe", "-cmoe"},
         "keep all Mixture of Experts (MoE) weights in the CPU",
         [](common_params & params) {
-            params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
         }
     ).set_env("LLAMA_ARG_CPU_MOE"));
     add_opt(common_arg(
@@ -2438,7 +2561,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             for (int i = 0; i < value; ++i) {
                 // keep strings alive and avoid leaking memory by storing them in a static vector
                 static std::list<std::string> buft_overrides;
-                buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                buft_overrides.push_back(llm_ffn_exps_block_regex(i));
                 params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
             }
         }
@@ -2447,7 +2570,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--cpu-moe-draft", "-cmoed"},
         "keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
         [](common_params & params) {
-            params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+            params.speculative.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
     add_opt(common_arg(
@@ -2459,7 +2582,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             for (int i = 0; i < value; ++i) {
                 static std::list<std::string> buft_overrides_draft;
-                buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
                 params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
             }
         }
@@ -2624,6 +2747,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.url = value;
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
+    add_opt(common_arg(
+        { "-dr", "--docker-repo" }, "[<repo>/]<model>[:quant]",
+        "Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.\n"
+        "example: gemma3\n"
+        "(default: unused)",
+        [](common_params & params, const std::string & value) {
+            params.model.docker_repo = value;
+        }
+    ).set_env("LLAMA_ARG_DOCKER_REPO"));
     add_opt(common_arg(
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
@@ -3134,13 +3266,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ));
-    add_opt(common_arg(
-        {"--log-colors"},
-        "Enable colored logging",
-        [](common_params &) {
-            common_log_set_colors(common_log_main(), true);
-        }
-    ).set_env("LLAMA_LOG_COLORS"));
+    add_opt(common_arg({ "--log-colors" }, "[on|off|auto]",
+                       "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
+                       "'auto' enables colors when output is to a terminal",
+                       [](common_params &, const std::string & value) {
+                           if (is_truthy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
+                           } else if (is_falsey(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
+                           } else if (is_autoy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
+                           } else {
+                               throw std::invalid_argument(
+                                   string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
+                           }
+                       }).set_env("LLAMA_LOG_COLORS"));
     add_opt(common_arg(
         {"-v", "--verbose", "--log-verbose"},
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",

common/chat.cpp

@@ -163,6 +163,19 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
     throw std::runtime_error("Invalid tool_choice: " + tool_choice);
 }
 
+bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates) {
+    common_chat_templates_inputs dummy_inputs;
+    common_chat_msg msg;
+    msg.role = "user";
+    msg.content = "test";
+    dummy_inputs.messages = {msg};
+    dummy_inputs.enable_thinking = false;
+    const auto rendered_no_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
+    dummy_inputs.enable_thinking = true;
+    const auto rendered_with_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
+    return rendered_no_thinking.prompt != rendered_with_thinking.prompt;
+}
+
 template <>
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {
     std::vector<common_chat_msg> msgs;
@@ -618,6 +631,7 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
+        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1: return "DeepSeek V3.1";
         case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
         case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
         case COMMON_CHAT_FORMAT_GRANITE: return "Granite";
@@ -685,11 +699,13 @@ static void parse_json_tool_calls(
         size_t from = std::string::npos;
         auto first = true;
         while (true) {
+            auto start_pos = builder.pos();
             auto res = function_regex_start_only && first
                 ? builder.try_consume_regex(*function_regex_start_only)
                 : function_regex
                     ? builder.try_find_regex(*function_regex, from)
                     : std::nullopt;
+
             if (res) {
                 std::string name;
                 if (get_function_name) {
@@ -724,6 +740,8 @@ static void parse_json_tool_calls(
                     return;
                 }
                 throw common_chat_msg_partial_exception("incomplete tool call");
+            } else {
+                builder.move_to(start_pos);
             }
             break;
         }
@@ -1375,6 +1393,71 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
     }
     return data;
 }
+
+static common_chat_params common_chat_params_init_deepseek_v3_1(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // Pass thinking context for DeepSeek V3.1 template
+    json additional_context = {
+        {"thinking", inputs.enable_thinking},
+    };
+
+    auto prompt = apply(tmpl, inputs,
+                       /* messages_override= */ inputs.messages,
+                       /* tools_override= */ std::nullopt,
+                       additional_context);
+    data.prompt = prompt;
+    data.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    if (string_ends_with(data.prompt, "<think>")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null();
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+                tool_rules.push_back(builder.add_rule(name + "-call",
+                    "( \"<｜tool▁call▁begin｜>\" )? \"" + name + "<｜tool▁sep｜>"
+                    "\" " + builder.add_schema(name + "-args", parameters) + " "
+                    "\"<｜tool▁call▁end｜>\""));
+            });
+            // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag,
+            // so we accept common variants (then it's all constrained)
+            builder.add_rule("root",
+                std::string(data.thinking_forced_open ? "( \"</think>\" space )? " : "") +
+                "( \"<｜tool▁calls▁begin｜>\" | \"<｜tool_calls_begin｜>\" | \"<｜tool calls begin｜>\" | \"<｜tool\\\\_calls\\\\_begin｜>\" | \"<｜tool▁calls｜>\" ) "
+                "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " "
+                "\"<｜tool▁calls▁end｜>\""
+                " space");
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                // If thinking_forced_open, then we capture the </think> tag in the grammar,
+                // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
+                std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)" : "(?:<think>[\\s\\S]*?</think>\\s*)?") +
+                    "(<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)[\\s\\S]*"
+            });
+            data.preserved_tokens = {
+                "<think>",
+                "</think>",
+                "<｜tool▁calls▁begin｜>",
+                "<｜tool▁call▁begin｜>",
+                "<｜tool▁sep｜>",
+                "<｜tool▁call▁end｜>",
+                "<｜tool▁calls▁end｜>",
+            };
+        });
+    }
+    return data;
+}
+
 static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
     builder.try_parse_reasoning("<think>", "</think>");
     if (!builder.syntax().parse_tool_calls) {
@@ -1396,6 +1479,66 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
         tool_calls_end);
 }
 
+static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) {
+    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)");
+
+    static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>");
+    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
+    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
+
+    if (!builder.syntax().parse_tool_calls) {
+        LOG_DBG("%s: not parse_tool_calls\n", __func__);
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    LOG_DBG("%s: parse_tool_calls\n", __func__);
+
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ tool_calls_begin,
+        /* function_regex_start_only= */ std::nullopt,
+        function_regex,
+        close_regex,
+        tool_calls_end);
+}
+
+static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
+    // DeepSeek V3.1 outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
+    // First try to parse using the standard reasoning parsing method
+    LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
+
+    auto start_pos = builder.pos();
+    auto found_end_think = builder.try_find_literal("</think>");
+    builder.move_to(start_pos);
+
+    if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
+        LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else if (builder.try_parse_reasoning("<think>", "</think>")) {
+        // If reasoning was parsed successfully, the remaining content is regular content
+        LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
+        // </think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else {
+        if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
+          LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
+          common_chat_parse_deepseek_v3_1_content(builder);
+          return;
+        }
+        // If no reasoning tags found, check if we should treat everything as reasoning
+        if (builder.syntax().thinking_forced_open) {
+            // If thinking is forced open but no tags found, treat everything as reasoning
+            LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
+            builder.add_reasoning_content(builder.consume_rest());
+        } else {
+            LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
+            // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜>
+            common_chat_parse_deepseek_v3_1_content(builder);
+        }
+    }
+}
+
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     auto prompt = apply(tmpl, inputs);
@@ -2352,6 +2495,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         }
     }
 
+    // DeepSeek V3.1: detect based on specific patterns in the template
+    if (src.find("message['prefix'] is defined and message['prefix'] and thinking") != std::string::npos &&
+        params.json_schema.is_null()) {
+        return common_chat_params_init_deepseek_v3_1(tmpl, params);
+    }
+
     // DeepSeek R1: use handler in all cases except json schema (thinking / tools).
     if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_deepseek_r1(tmpl, params);
@@ -2524,6 +2673,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
             common_chat_parse_deepseek_r1(builder);
             break;
+        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1:
+            common_chat_parse_deepseek_v3_1(builder);
+            break;
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
             common_chat_parse_functionary_v3_2(builder);
             break;

common/chat.h

@@ -107,6 +107,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
+    COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
     COMMON_CHAT_FORMAT_GRANITE,
@@ -199,6 +200,8 @@ common_chat_msg           common_chat_parse(const std::string & input, bool is_p
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
+bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
+
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 // T can be std::string containing JSON or nlohmann::ordered_json
 template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);

common/common.h

@@ -193,10 +193,11 @@ struct common_params_sampling {
 };
 
 struct common_params_model {
-    std::string path    = ""; // model local path                                           // NOLINT
-    std::string url     = ""; // model url to download                                      // NOLINT
-    std::string hf_repo = ""; // HF repo                                                    // NOLINT
-    std::string hf_file = ""; // HF file                                                    // NOLINT
+    std::string path        = ""; // model local path                                       // NOLINT
+    std::string url         = ""; // model url to download                                  // NOLINT
+    std::string hf_repo     = ""; // HF repo                                                // NOLINT
+    std::string hf_file     = ""; // HF file                                                // NOLINT
+    std::string docker_repo = ""; // Docker repo                                            // NOLINT
 };
 
 struct common_params_speculative {
@@ -287,9 +288,9 @@ struct common_params {
     float   rope_freq_base        =  0.0f; // RoPE base frequency
     float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
     float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
-    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    float   yarn_attn_factor      = -1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = -1.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = -1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
 
     // offload params
@@ -452,7 +453,7 @@ struct common_params {
 
     std::string slot_save_path;
 
-    float slot_prompt_similarity = 0.5f;
+    float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
     bool is_pp_shared = false;
@@ -733,6 +734,20 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
 
+//
+// MoE utils
+//
+
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_exps";
+
+static std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
+}
+
+static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+    return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
+}
+
 //
 // training utils
 //

common/json-schema-to-grammar.cpp

@@ -257,12 +257,13 @@ std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
 };
 
 static bool is_reserved_name(const std::string & name) {
-    static std::unordered_set<std::string> RESERVED_NAMES;
-    if (RESERVED_NAMES.empty()) {
-        RESERVED_NAMES.insert("root");
-        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
-    }
+    static const std::unordered_set<std::string> RESERVED_NAMES = [] {
+        std::unordered_set<std::string> s;
+        s.insert("root");
+        for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
+        for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
+        return s;
+    }();
     return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }
 
@@ -843,9 +844,10 @@ public:
                 _build_object_rule(
                     properties, required, name,
                     schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+        } else if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
             std::unordered_set<std::string> required;
             std::vector<std::pair<std::string, json>> properties;
+            std::map<std::string, size_t> enum_values;
             std::string hybrid_name = name;
             std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
                 if (comp_schema.contains("$ref")) {
@@ -857,6 +859,14 @@ public:
                             required.insert(prop.key());
                         }
                     }
+                } else if (comp_schema.contains("enum")) {
+                    for (const auto & v : comp_schema["enum"]) {
+                        const auto rule = _generate_constant_rule(v);
+                        if (enum_values.find(rule) == enum_values.end()) {
+                            enum_values[rule] = 0;
+                        }
+                        enum_values[rule] += 1;
+                    }
                 } else {
                   // todo warning
                 }
@@ -870,6 +880,17 @@ public:
                     add_component(t, true);
                 }
             }
+            if (!enum_values.empty()) {
+                std::vector<std::string> enum_intersection;
+                for (const auto & p : enum_values) {
+                    if (p.second == schema["allOf"].size()) {
+                        enum_intersection.push_back(p.first);
+                    }
+                }
+                if (!enum_intersection.empty()) {
+                    return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ") space");
+                }
+            }
             return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
         } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
             json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];

common/log.cpp

@@ -4,17 +4,52 @@
 #include <condition_variable>
 #include <cstdarg>
 #include <cstdio>
+#include <cstdlib>
+#include <cstring>
 #include <mutex>
 #include <sstream>
 #include <thread>
 #include <vector>
 
+#if defined(_WIN32)
+#    include <io.h>
+#    include <windows.h>
+#    define isatty _isatty
+#    define fileno _fileno
+#else
+#    include <unistd.h>
+#endif // defined(_WIN32)
+
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
 void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
+// Auto-detect if colors should be enabled based on terminal and environment
+static bool common_log_should_use_colors_auto() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
+
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -353,6 +388,11 @@ struct common_log * common_log_init() {
 
 struct common_log * common_log_main() {
     static struct common_log log;
+    static std::once_flag    init_flag;
+    std::call_once(init_flag, [&]() {
+        // Set default to auto-detect colors
+        log.set_colors(common_log_should_use_colors_auto());
+    });
 
     return &log;
 }
@@ -380,8 +420,19 @@ void common_log_set_file(struct common_log * log, const char * file) {
     log->set_file(file);
 }
 
-void common_log_set_colors(struct common_log * log, bool colors) {
-    log->set_colors(colors);
+void common_log_set_colors(struct common_log * log, log_colors colors) {
+    if (colors == LOG_COLORS_AUTO) {
+        log->set_colors(common_log_should_use_colors_auto());
+        return;
+    }
+
+    if (colors == LOG_COLORS_DISABLED) {
+        log->set_colors(false);
+        return;
+    }
+
+    GGML_ASSERT(colors == LOG_COLORS_ENABLED);
+    log->set_colors(true);
 }
 
 void common_log_set_prefix(struct common_log * log, bool prefix) {

common/log.h

@@ -24,6 +24,12 @@
 #define LOG_DEFAULT_DEBUG 1
 #define LOG_DEFAULT_LLAMA 0
 
+enum log_colors {
+    LOG_COLORS_AUTO     = -1,
+    LOG_COLORS_DISABLED = 0,
+    LOG_COLORS_ENABLED  = 1,
+};
+
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
 extern int common_log_verbosity_thold;
@@ -65,10 +71,10 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
 //
 
-void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
-void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
-void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
-void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
+void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
+void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
+void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold

convert_hf_to_gguf.py

@@ -735,6 +735,9 @@ class TextModel(ModelBase):
         if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c":
             # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B
             res = "qwen2"
+        if chkhsh == "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273":
+            # ref: https://huggingface.co/alvarobartt/grok-2-tokenizer
+            res = "grok-2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -885,6 +888,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
+        if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
+            # ref: https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base
+            res = "llada-moe"
 
         if res is None:
             logger.warning("\n")
@@ -2682,12 +2688,20 @@ class BitnetModel(TextModel):
         yield (new_name, data_torch)
 
 
-@ModelBase.register("GrokForCausalLM")
+@ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")
 class GrokModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GROK
 
     def set_vocab(self):
-        self._set_vocab_sentencepiece()
+        if (self.dir_model / 'tokenizer.model').is_file():
+            self._set_vocab_sentencepiece()
+            return
+
+        if not (self.dir_model / 'tokenizer.json').is_file() or not (self.dir_model / 'chat_template.jinja').is_file():
+            logger.error('Error: Missing vocab and chat template, download files from https://huggingface.co/alvarobartt/grok-2-tokenizer')
+            sys.exit(1)
+
+        self._set_vocab_gpt2()
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -2695,11 +2709,46 @@ class GrokModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-    _experts: list[dict[str, Tensor]] | None = None
+        self.gguf_writer.add_attn_logit_softcapping(self.hparams.get("attn_logit_softcapping", 30.0))
+        self.gguf_writer.add_router_logit_softcapping(self.hparams.get("router_logit_softcapping", 30.0))
+        if (final_logit_softcap := self.hparams.get("final_logit_softcapping")):
+            self.gguf_writer.add_final_logit_softcapping(final_logit_softcap)
+
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+
+        # Treat "original" as "yarn", seems to have been a mistake
+        if self.hparams.get("rope_type") in ("yarn", "original"):
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(self.hparams["scaling_factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_yarn_ext_factor(self.hparams["extrapolation_factor"])
+            self.gguf_writer.add_rope_scaling_yarn_attn_factor(self.hparams["attn_factor"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_fast(self.hparams["beta_fast"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_slow(self.hparams["beta_slow"])
+
+        if temp_len := self.hparams.get("attn_temperature_len"):
+            self.gguf_writer.add_attn_temperature_length(temp_len)
+
+        self.gguf_writer.add_attn_output_scale(self.hparams.get("attn_output_multiplier", rope_dim**-0.5))
+        self.gguf_writer.add_embedding_scale(self.hparams["embedding_multiplier_scale"])
+        self.gguf_writer.add_logit_scale(self.hparams["output_multiplier_scale"])
+
+    _experts: list[dict[str, list[Tensor]]] | None = None
+    _cur_expert = ""
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        tensors: list[tuple[str, Tensor]] = []
+        is_expert = ".moe." in name or ".block_sparse_moe.experts." in name
+
+        if not is_expert:
+            tensors.append((self.map_tensor_name(name), data_torch))
+
         # process the experts separately
-        if name.find(".moe.") != -1:
+        if is_expert or self._cur_expert:
             n_experts = self.hparams["num_local_experts"]
 
             assert bid is not None
@@ -2707,32 +2756,41 @@ class GrokModel(TextModel):
             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:
-                tensors: list[tuple[str, Tensor]] = []
-
-                # merge the experts into a single 3d tensor
-                for wid in ["linear", "linear_1", "linear_v"]:
-                    datas: list[Tensor] = []
-
-                    for xid in range(n_experts):
-                        ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
-                        datas.append(self._experts[bid][ename])
-                        del self._experts[bid][ename]
-
-                    data_torch = torch.stack(datas, dim=0)
-
-                    merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
-
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
-            else:
+            # concatenate split tensors
+            if name in self._experts[bid]:
+                self._cur_expert = name
+                self._experts[bid][name].append(data_torch)
                 return []
+            elif is_expert:
+                self._cur_expert = name
+                self._experts[bid][name] = [data_torch]
+                return []
+            else:
+                self._cur_expert = ""
 
-        return [(self.map_tensor_name(name), data_torch)]
+            for bid in range(self.block_count):
+                if len(self._experts[bid]) >= n_experts * 3:
+                    # merge the experts into a single 3d tensor
+                    for wid in [("linear", "w1", 0), ("linear_1", "w2", 1), ("linear_v", "w3", 0)]:
+                        datas: list[Tensor] = []
+
+                        for xid in range(n_experts):
+                            ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid[0]}.weight"
+                            if ename not in self._experts[bid]:
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid[1]}.weight"
+                            tensor_list = self._experts[bid][ename]
+                            datas.append(torch.cat(tensor_list, dim=wid[2]) if len(tensor_list) > 1 else tensor_list[0])
+                            del self._experts[bid][ename]
+
+                        data_torch = torch.stack(datas, dim=0)
+
+                        merged_name = f"transformer.decoder_layer.{bid}.moe.{wid[0]}.weight"
+
+                        new_name = self.map_tensor_name(merged_name)
+
+                        yield (new_name, data_torch)
+
+        yield from tensors
 
 
 @ModelBase.register("DbrxForCausalLM")
@@ -5128,6 +5186,20 @@ class EmbeddingGemma(Gemma3Model):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+
+        # Override the sliding window size as it gets adjusted by the Gemma3TextConfig
+        # constructor. We want to use the value from the original model's config.json.
+        # ref: https://github.com/huggingface/transformers/pull/40700
+        with open(self.dir_model / "config.json", "r", encoding="utf-8") as f:
+            config = json.load(f)
+            orig_sliding_window = config.get("sliding_window")
+            if orig_sliding_window is None:
+                raise ValueError("sliding_window not found in model config - this is required for the model")
+
+            logger.info(f"Using original sliding_window from config: {orig_sliding_window} "
+                        f"instead of {self.hparams['sliding_window']}")
+            self.gguf_writer.add_sliding_window(orig_sliding_window)
+
         self._try_set_pooling_type()
 
 
@@ -5937,9 +6009,34 @@ class SeedOssModel(TextModel):
 
 
 @ModelBase.register("Olmo2ForCausalLM")
+@ModelBase.register("Olmo3ForCausalLM")
 class Olmo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.OLMO2
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_attn_factors(rope_scaling["attention_factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+        if "sliding_window" in self.hparams:
+            self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+
+            sliding_window_pattern = []
+            if "layer_types" in self.hparams:
+                sliding_window_pattern = [t == "sliding_attention" for t in self.hparams["layer_types"]]
+            else:
+                # Olmo2 does not use sliding window attention.
+                # Olmo3 defaults to using sliding window for all layers except every 4th.
+                for i in range(self.hparams["num_hidden_layers"]):
+                    sliding_window_pattern.append((i + 1) % 4 != 0)
+
+            self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+
 
 @ModelBase.register("OlmoeForCausalLM")
 class OlmoeModel(TextModel):
@@ -6687,6 +6784,8 @@ class T5Model(TextModel):
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
         self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        if (dec_n_layer := self.hparams.get("num_decoder_layers")) is not None:
+            self.gguf_writer.add_decoder_block_count(dec_n_layer)
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
         self.gguf_writer.add_key_length(self.hparams["d_kv"])
         self.gguf_writer.add_value_length(self.hparams["d_kv"])
@@ -8168,6 +8267,76 @@ class HunYuanMoEModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("LLaDAMoEModel", "LLaDAMoEModelLM")
+class LLaDAMoEModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LLADA_MOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+
+        if (expert_intermediate_size := self.hparams.get("expert_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
+
+        # number of experts used per token (top-k)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+
+        self.gguf_writer.add_mask_token_id(156895)
+        self.gguf_writer.add_causal_attention(False)
+        self.gguf_writer.add_diffusion_shift_logits(False)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    # Copied from: Qwen2MoeModel
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            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:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # Copied from: Qwen2MoeModel
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("HunYuanDenseV1ForCausalLM")
 class HunYuanModel(TextModel):
     model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE

convert_hf_to_gguf_update.py

@@ -139,6 +139,7 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
+    {"name": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -158,6 +159,7 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
     {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
     {"name": "qwen2",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "chkhsh": "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c"},
+    {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
 ]
 
 

convert_lora_to_gguf.py

@@ -12,7 +12,7 @@ import json
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
-from transformers import AutoConfig
+from transformers import AutoConfig, AutoTokenizer
 
 import torch
 
@@ -26,6 +26,8 @@ import gguf
 # reuse model definitions from convert_hf_to_gguf.py
 from convert_hf_to_gguf import LazyTorchTensor, ModelBase
 
+from gguf.constants import GGUFValueType
+
 logger = logging.getLogger("lora-to-gguf")
 
 
@@ -369,7 +371,31 @@ if __name__ == '__main__':
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
+                logger.debug("GGUF KV: %s = %d", gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
+                alora_invocation_tokens = lparams.get("alora_invocation_tokens")
+                invocation_string = lparams.get("invocation_string")
+                if invocation_string and not alora_invocation_tokens:
+                    logger.debug("Tokenizing invocation_string -> alora_invocation_tokens")
+                    base_model_path_or_id = hparams.get("_name_or_path")
+                    try:
+                        tokenizer = AutoTokenizer.from_pretrained(base_model_path_or_id)
+                    except ValueError:
+                        logger.error("Unable to load tokenizer from %s", base_model_path_or_id)
+                        raise
+                    # NOTE: There's an off-by-one with the older aLoRAs where
+                    # the invocation string includes the "<|start_of_turn|>"
+                    # token, but the adapters themselves were trained to
+                    # activate _after_ that first token, so we drop it here.
+                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
+                if alora_invocation_tokens:
+                    logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
+                    self.gguf_writer.add_key_value(
+                        gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS,
+                        alora_invocation_tokens,
+                        GGUFValueType.ARRAY,
+                        GGUFValueType.UINT32,
+                    )
 
             def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                 # Never add extra tensors (e.g. rope_freqs) for LoRA adapters

docs/backend/CANN.md

@@ -314,3 +314,11 @@ Converting the matmul weight format from ND to NZ to improve performance. Enable
 ### GGML_CANN_ACL_GRAPH
 
 Operators are executed using ACL graph execution, rather than in op-by-op (eager) mode. Enabled by default.
+
+### GGML_CANN_GRAPH_CACHE_CAPACITY
+
+Maximum number of compiled CANN graphs kept in the LRU cache, default is 12. When the number of cached graphs exceeds this capacity, the least recently used graph will be evicted.
+
+### GGML_CANN_PREFILL_USE_GRAPH
+
+Enable ACL graph execution during the prefill stage, default is false. This option is only effective when FA is enabled.

docs/build-s390x.md

@@ -42,18 +42,6 @@ cmake --build build --config Release -j $(nproc)
     cmake --build build --config Release -j $(nproc)
     ```
 
--   By default, NNPA is disabled by default. To enable it:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_NNPA=ON
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
 -   For debug builds:
 
     ```bash
@@ -164,15 +152,11 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 Only available in IBM z15/LinuxONE 3 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
 
-### 2. NNPA Vector Intrinsics Acceleration
-
-Only available in IBM z16/LinuxONE 4 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
-
-### 3. zDNN Accelerator (WIP)
+### 2. zDNN Accelerator (WIP)
 
 Only available in IBM z17/LinuxONE 5 or later system with the `-DGGML_ZDNN=ON` compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs will default back to CPU routines.
 
-### 4. Spyre Accelerator
+### 3. Spyre Accelerator
 
 _Only available with IBM z17 / LinuxONE 5 or later system. No support currently available._
 
@@ -230,10 +214,6 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
     CXXFLAGS="-include cstdint" pip3 install -r requirements.txt
     ```
 
-5. `-DGGML_NNPA=ON` generates gibberish output
-
-    Answer: We are aware of this as detailed in [this issue](https://github.com/ggml-org/llama.cpp/issues/14877). Please either try reducing the number of threads, or disable the compile option using `-DGGML_NNPA=OFF`.
-
 ## Getting Help on IBM Z & LinuxONE
 
 1. **Bugs, Feature Requests**
@@ -258,38 +238,38 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 ## Appendix B: SIMD Support Matrix
 
-|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
-| ---------- | ----------- | ---- | ---- | ----- |
-| FP32       | ✅          | ✅   | ✅   | ❓    |
-| FP16       | ✅          | ✅   | ❓   | ❓    |
-| BF16       | 🚫          | 🚫   | ❓   | ❓    |
-| Q4_0       | ✅          | ✅   | ❓   | ❓    |
-| Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| MXFP4      | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_0       | ✅          | ✅   | ❓   | ❓    |
-| Q5_1       | ✅          | ✅   | ❓   | ❓    |
-| Q8_0       | ✅          | ✅   | ❓   | ❓    |
-| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
-| Q3_K       | ✅          | ✅   | ❓   | ❓    |
-| Q4_K       | ✅          | ✅   | ❓   | ❓    |
-| Q5_K       | ✅          | ✅   | ❓   | ❓    |
-| Q6_K       | ✅          | ✅   | ❓   | ❓    |
-| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
-| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
-| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
-| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
-| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+|            | VX/VXE/VXE2 | zDNN | Spyre |
+|------------|-------------|------|-------|
+| FP32       | ✅           | ✅    | ❓     |
+| FP16       | ✅           | ✅    | ❓     |
+| BF16       | 🚫           | ✅    | ❓     |
+| Q4_0       | ✅           | ❓    | ❓     |
+| Q4_1       | ✅           | ❓    | ❓     |
+| MXFP4      | 🚫           | ❓    | ❓     |
+| Q5_0       | ✅           | ❓    | ❓     |
+| Q5_1       | ✅           | ❓    | ❓     |
+| Q8_0       | ✅           | ❓    | ❓     |
+| Q2_K       | 🚫           | ❓    | ❓     |
+| Q3_K       | ✅           | ❓    | ❓     |
+| Q4_K       | ✅           | ❓    | ❓     |
+| Q5_K       | ✅           | ❓    | ❓     |
+| Q6_K       | ✅           | ❓    | ❓     |
+| TQ1_0      | 🚫           | ❓    | ❓     |
+| TQ2_0      | 🚫           | ❓    | ❓     |
+| IQ2_XXS    | 🚫           | ❓    | ❓     |
+| IQ2_XS     | 🚫           | ❓    | ❓     |
+| IQ2_S      | 🚫           | ❓    | ❓     |
+| IQ3_XXS    | 🚫           | ❓    | ❓     |
+| IQ3_S      | 🚫           | ❓    | ❓     |
+| IQ1_S      | 🚫           | ❓    | ❓     |
+| IQ1_M      | 🚫           | ❓    | ❓     |
+| IQ4_NL     | ✅           | ❓    | ❓     |
+| IQ4_XS     | ✅           | ❓    | ❓     |
+| FP32->FP16 | 🚫           | ❓    | ❓     |
+| FP16->FP32 | 🚫           | ❓    | ❓     |
 
 -   ✅ - acceleration available
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Aug 22, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.

docs/ops.md

@@ -18,6 +18,7 @@ Legend:
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -26,6 +27,7 @@ Legend:
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
@@ -49,9 +51,11 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
@@ -61,7 +65,9 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -98,6 +104,7 @@ Legend:
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -333,17 +333,17 @@ static void print_params(struct my_llama_hparams * params) {
 }
 
 static void print_tensor_info(const struct ggml_context * ctx) {
-    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+    for (auto * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
         LOG_INF("%s: Allocating ", __func__);
         int64_t total = 1;
         int i = 0;
         for (; i < ggml_n_dims(t); ++i) {
-            if (i > 0) LOG("x ");
-            LOG("[%" PRId64 "] ", t->ne[i]);
+            if (i > 0) { LOG_INF("x "); }
+            LOG_INF("[%" PRId64 "] ", t->ne[i]);
             total *= t->ne[i];
         }
-        if (i > 1) LOG("= [%" PRId64 "] ", total);
-        LOG("float space for %s\n", ggml_get_name(t));
+        if (i > 1) { LOG_INF("= [%" PRId64 "] ", total); }
+        LOG_INF("float space for %s\n", ggml_get_name(t));
     }
 }
 

examples/diffusion/diffusion-cli.cpp

@@ -510,19 +510,27 @@ static void diffusion_generate(llama_context *          ctx,
     n_generated = params.max_length;
 }
 
-static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
+static std::string format_input_text(const std::string & prompt, const std::string & system_prompt, bool use_chat_template, llama_model * model) {
     if (!use_chat_template) {
         return prompt;
     }
 
     auto chat_templates = common_chat_templates_init(model, "");
-
     common_chat_templates_inputs inputs;
-    common_chat_msg              user_msg;
-    user_msg.role                = "user";
-    user_msg.content             = prompt;
-    inputs.add_generation_prompt = true;
+    common_chat_msg system_msg;
+
+    if (!system_prompt.empty()) {
+        system_msg.role = "system";
+        system_msg.content = system_prompt;
+        inputs.messages.push_back(system_msg);
+    }
+
+    common_chat_msg user_msg;
+    user_msg.role = "user";
+    user_msg.content = prompt;
+
     inputs.messages.push_back(user_msg);
+    inputs.add_generation_prompt = true;
 
     auto result = common_chat_templates_apply(chat_templates.get(), inputs);
 
@@ -579,7 +587,8 @@ int main(int argc, char ** argv) {
     llama_set_n_threads(ctx, params.cpuparams.n_threads, params.cpuparams_batch.n_threads);
 
     const llama_vocab * vocab            = llama_model_get_vocab(model);
-    std::string         formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
+
+    std::string         formatted_prompt = format_input_text(params.prompt, params.system_prompt, params.enable_chat_template, model);
 
     std::vector<llama_token> input_tokens = common_tokenize(vocab,
                                                             formatted_prompt,
@@ -596,6 +605,7 @@ int main(int argc, char ** argv) {
     }
 
     llama_token mask_token_id = llama_vocab_mask(vocab);
+
     GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
 
     bool visual_mode = params.diffusion.visual_mode;

examples/eval-callback/eval-callback.cpp

@@ -28,6 +28,15 @@ static std::string ggml_ne_string(const ggml_tensor * t) {
     return str;
 }
 
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
 static float ggml_get_float_value(uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
     float v;
@@ -43,6 +52,8 @@ static float ggml_get_float_value(uint8_t * data, ggml_type type, const size_t *
         v = (float) *(int16_t *) &data[i];
     } else if (type == GGML_TYPE_I8) {
         v = (float) *(int8_t *) &data[i];
+    } else if (type == GGML_TYPE_BF16) {
+        v = ggml_compute_bf16_to_fp32(*(ggml_bf16_t *) &data[i]);
     } else {
         GGML_ABORT("fatal error");
     }

examples/json_schema_to_grammar.py

@@ -586,9 +586,10 @@ class SchemaConverter:
             properties = list(schema.get('properties', {}).items())
             return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
 
-        elif schema_type in (None, 'object') and 'allOf' in schema:
+        elif schema_type in (None, 'object', 'string') and 'allOf' in schema:
             required = set()
             properties = []
+            enum_sets = []
             hybrid_name = name
             def add_component(comp_schema, is_required):
                 if (ref := comp_schema.get('$ref')) is not None:
@@ -600,6 +601,9 @@ class SchemaConverter:
                         if is_required:
                             required.add(prop_name)
 
+                if 'enum' in comp_schema:
+                    enum_sets.append(set(comp_schema['enum']))
+
             for t in schema['allOf']:
                 if 'anyOf' in t:
                     for tt in t['anyOf']:
@@ -607,6 +611,15 @@ class SchemaConverter:
                 else:
                     add_component(t, is_required=True)
 
+            if enum_sets:
+                enum_intersection = enum_sets[0]
+                for s in enum_sets[1:]:
+                    enum_intersection &= s
+
+                if enum_intersection:
+                    rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in sorted(enum_intersection))) + ') space'
+                    return self._add_rule(rule_name, rule)
+
             return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
 
         elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):

examples/model-conversion/requirements.txt

@@ -1,5 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch~=2.6.0
-torchvision~=0.21.0
-transformers~=4.55.0
-huggingface-hub~=0.34.0
+torch
+torchvision
+transformers
+huggingface-hub
+accelerate

examples/model-conversion/scripts/causal/run-org-model.py

@@ -9,15 +9,134 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
 import torch
 import numpy as np
 
-unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+### If you want to dump RoPE activations, apply this monkey patch to the model
+### class from Transformers that you are running (replace apertus.modeling_apertus
+### with the proper package and class for your model
+### === START ROPE DEBUG ===
+# from transformers.models.apertus.modeling_apertus import apply_rotary_pos_emb
 
-parser = argparse.ArgumentParser(description='Process model with specified path')
-parser.add_argument('--model-path', '-m', help='Path to the model')
+# orig_rope = apply_rotary_pos_emb
+# torch.set_printoptions(threshold=float('inf'))
+# torch.set_printoptions(precision=6, sci_mode=False)
+
+# def debug_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+#     # log inputs
+#     summarize(q, "RoPE.q_in")
+#     summarize(k, "RoPE.k_in")
+
+#     # call original
+#     q_out, k_out = orig_rope(q, k, cos, sin, position_ids, unsqueeze_dim)
+
+#     # log outputs
+#     summarize(q_out, "RoPE.q_out")
+#     summarize(k_out, "RoPE.k_out")
+
+#     return q_out, k_out
+
+# # Patch it
+# import transformers.models.apertus.modeling_apertus as apertus_mod  # noqa: E402
+# apertus_mod.apply_rotary_pos_emb = debug_rope
+### == END ROPE DEBUG ===
+
+
+def summarize(tensor: torch.Tensor, name: str, max_seq: int = 3, max_vals: int = 3):
+    """
+    Print a tensor in llama.cpp debug style.
+
+    Supports:
+    - 2D tensors (seq, hidden)
+    - 3D tensors (batch, seq, hidden)
+    - 4D tensors (batch, seq, heads, dim_per_head) via flattening heads × dim_per_head
+
+    Shows first and last max_vals of each vector per sequence position.
+    """
+    t = tensor.detach().to(torch.float32).cpu()
+
+    # Determine dimensions
+    if t.ndim == 3:
+        _, s, _ = t.shape
+    elif t.ndim == 2:
+        _, s = 1, t.shape[0]
+        t = t.unsqueeze(0)
+    elif t.ndim == 4:
+        _, s, _, _ = t.shape
+    else:
+        print(f"Skipping tensor due to unsupported dimensions: {t.ndim}")
+        return
+
+    ten_shape = t.shape
+
+    print(f"ggml_debug: {name} = (f32)  ... = {{{ten_shape}}}")
+    print("                                     [")
+    print("                                      [")
+
+    # Determine indices for first and last sequences
+    first_indices = list(range(min(s, max_seq)))
+    last_indices = list(range(max(0, s - max_seq), s))
+
+    # Check if there's an overlap between first and last indices or if we're at the edge case of s = 2 * max_seq
+    has_overlap = bool(set(first_indices) & set(last_indices)) or (max_seq * 2 == s)
+
+    # Combine indices
+    if has_overlap:
+        # If there's overlap, just use the combined unique indices
+        indices = sorted(list(set(first_indices + last_indices)))
+        separator_index = None
+    else:
+        # If no overlap, we'll add a separator between first and last sequences
+        indices = first_indices + last_indices
+        separator_index = len(first_indices)
+
+    for i, si in enumerate(indices):
+        # Add separator if needed
+        if separator_index is not None and i == separator_index:
+            print("                                       ...")
+
+        # Extract appropriate slice
+        vec = t[0, si]
+        if vec.ndim == 2:  # 4D case: flatten heads × dim_per_head
+            flat = vec.flatten().tolist()
+        else:  # 2D or 3D case
+            flat = vec.tolist()
+
+        # First and last slices
+        first = flat[:max_vals]
+        last = flat[-max_vals:] if len(flat) >= max_vals else flat
+        first_str = ", ".join(f"{v:12.4f}" for v in first)
+        last_str = ", ".join(f"{v:12.4f}" for v in last)
+
+        print(f"                                       [{first_str}, ..., {last_str}]")
+
+    print("                                      ],")
+    print("                                     ]")
+    print(f"                                     sum = {t.sum().item():.6f}\n")
+
+
+def debug_hook(name):
+    def fn(_m, input, output):
+        if isinstance(input, torch.Tensor):
+            summarize(input, name + "_in")
+        elif isinstance(input, (tuple, list)) and isinstance(input[0], torch.Tensor):
+            summarize(input[0], name + "_in")
+        if isinstance(output, torch.Tensor):
+            summarize(output, name + "_out")
+        elif isinstance(output, (tuple, list)) and isinstance(output[0], torch.Tensor):
+            summarize(output[0], name + "_out")
+
+    return fn
+
+
+unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
+
+parser = argparse.ArgumentParser(description="Process model with specified path")
+parser.add_argument("--model-path", "-m", help="Path to the model")
 args = parser.parse_args()
 
-model_path = os.environ.get('MODEL_PATH', args.model_path)
+model_path = os.environ.get("MODEL_PATH", args.model_path)
 if model_path is None:
-    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+    parser.error(
+        "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
+    )
 
 config = AutoConfig.from_pretrained(model_path)
 
@@ -34,18 +153,30 @@ config = AutoConfig.from_pretrained(model_path)
 
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
-    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    unreleased_module_path = (
+        f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    )
     class_name = f"{unreleased_model_name}ForCausalLM"
     print(f"Importing unreleased model module: {unreleased_module_path}")
 
     try:
-        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
-        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+        model_class = getattr(
+            importlib.import_module(unreleased_module_path), class_name
+        )
+        model = model_class.from_pretrained(
+            model_path
+        )  # Note: from_pretrained, not fromPretrained
     except (ImportError, AttributeError) as e:
         print(f"Failed to import or load model: {e}")
         exit(1)
 else:
-    model = AutoModelForCausalLM.from_pretrained(model_path)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, device_map="auto", offload_folder="offload"
+    )
+
+for name, module in model.named_modules():
+    if len(list(module.children())) == 0:  # only leaf modules
+        module.register_forward_hook(debug_hook(name))
 
 model_name = os.path.basename(model_path)
 # Printing the Model class to allow for easier debugging. This can be useful

examples/simple/simple.cpp

@@ -145,6 +145,20 @@ int main(int argc, char ** argv) {
 
     llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
 
+    if (llama_model_has_encoder(model)) {
+        if (llama_encode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            decoder_start_token_id = llama_vocab_bos(vocab);
+        }
+
+        batch = llama_batch_get_one(&decoder_start_token_id, 1);
+    }
+
     // main loop
 
     const auto t_main_start = ggml_time_us();

ggml/CMakeLists.txt

@@ -134,7 +134,6 @@ option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
 option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
-option(GGML_NNPA             "ggml: enable nnpa"             OFF)  # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")

ggml/include/ggml-backend.h

@@ -132,6 +132,8 @@ extern "C" {
         GGML_BACKEND_DEVICE_TYPE_CPU,
         // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
+        // integrated GPU device using host memory
+        GGML_BACKEND_DEVICE_TYPE_IGPU,
         // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
         GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
@@ -150,11 +152,21 @@ extern "C" {
 
     // all the device properties
     struct ggml_backend_dev_props {
+        // device name
         const char * name;
+        // device description
         const char * description;
+        // device free memory in bytes
         size_t memory_free;
+        // device total memory in bytes
         size_t memory_total;
+        // device type
         enum ggml_backend_dev_type type;
+        // device id
+        //   for PCI devices, this should be the PCI bus id formatted as "domain:bus:device.function" (e.g. "0000:01:00.0")
+        //   if the id is unknown, this should be NULL
+        const char * device_id;
+        // device capabilities
         struct ggml_backend_dev_caps caps;
     };
 

ggml/include/ggml-cpu.h

@@ -101,7 +101,6 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
-    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -135,6 +134,7 @@ extern "C" {
     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_BACKEND_API void ggml_cpu_fp32_to_i32 (const float *,     int32_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

ggml/include/ggml-metal.h

@@ -43,14 +43,8 @@ GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_DEPRECATED(
-        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
-
 GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

ggml/include/ggml-zdnn.h

@@ -7,8 +7,6 @@
 extern "C" {
 #endif
 
-GGML_BACKEND_API ggml_backend_t ggml_backend_zdnn_init(void);
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
 
 #ifdef __cplusplus

ggml/include/ggml.h

@@ -1404,6 +1404,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // note: casting from f32 to i32 will discard the fractional part
     GGML_API struct ggml_tensor * ggml_cast(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1528,7 +1529,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // supports 3D: a->ne[2] == b->ne[1]
+    // supports 4D a:
+    // a     [n_embd, ne1, ne2, ne3]
+    // b I32 [n_rows, ne2, ne3, 1]
+    //
+    // return [n_embd, n_rows, ne2, ne3]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,  // data

ggml/src/ggml-backend-impl.h

@@ -8,7 +8,7 @@
 extern "C" {
 #endif
 
-    #define GGML_BACKEND_API_VERSION 1
+    #define GGML_BACKEND_API_VERSION 2
 
     //
     // Backend buffer type
@@ -114,6 +114,9 @@ extern "C" {
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
         // wait for an event on on a different stream
         void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
+
+        // (optional) sort/optimize the nodes in the graph
+        void                      (*optimize_graph)    (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     };
 
     struct ggml_backend {

ggml/src/ggml-backend-reg.cpp

@@ -400,9 +400,8 @@ ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const
 
 ggml_backend_t ggml_backend_init_best(void) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
-    if (!dev) {
-        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    }
+    dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     if (!dev) {
         return nullptr;
     }

ggml/src/ggml-backend.cpp

@@ -463,6 +463,13 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
     backend->iface.event_wait(backend, event);
 }
 
+static void ggml_backend_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend);
+    if (backend->iface.optimize_graph != NULL) {
+        backend->iface.optimize_graph(backend, cgraph);
+    }
+}
+
 // Backend device
 
 const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
@@ -1298,6 +1305,10 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
 
+        // Optimize this split of the graph. This needs to happen before we make graph_copy,
+        // so they are in sync.
+        ggml_backend_optimize_graph(sched->backends[split->backend_id], &split->graph);
+
         // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
             assert(graph_copy->size > (graph_copy->n_nodes + 1));

ggml/src/ggml-blas/ggml-blas.cpp

@@ -270,6 +270,7 @@ static struct ggml_backend_i blas_backend_i = {
     /* .graph_compute           = */ ggml_backend_blas_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .optimize_graph          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_blas_guid(void) {

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2268,8 +2268,6 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
  *                           stream, and persistent buffers for rope init/cache.
  * @param dst                The destination ggml_tensor whose computation
  *                           depends on the RoPE values (usually Qcur/Kcur).
- * @param sin_tensor_buffer  Pre-allocated buffer for storing repeated sin values.
- * @param cos_tensor_buffer  Pre-allocated buffer for storing repeated cos values.
  * @param theta_scale        Scalar exponent base for computing theta scale values.
  * @param freq_scale         Frequency scaling factor, applied to theta scale.
  * @param attn_factor        Attention scaling factor, applied to sin/cos.
@@ -2277,17 +2275,23 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
  *                           (dim expansion vs repeat_interleave).
  */
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
-                             void* sin_tensor_buffer, void* cos_tensor_buffer,
                              float* corr_dims, float ext_factor,
                              float theta_scale, float freq_scale,
                              float attn_factor, bool is_neox) {
-    // int sin/cos cache, cache has different repeat method depond on
-    // @param.is_neox
-
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
+    if(src2 == nullptr && ctx.rope_cache.cached
+        && ctx.rope_cache.ext_factor == ext_factor
+        && ctx.rope_cache.theta_scale == theta_scale
+        && ctx.rope_cache.freq_scale == freq_scale
+        && ctx.rope_cache.attn_factor == attn_factor
+        && ctx.rope_cache.is_neox == is_neox) {
+        // use cache.
+        return;
+    }
+
     int64_t theta_scale_length = src0->ne[0] / 2;
     int64_t theta_scale_ne[] = {theta_scale_length, 1, 1, 1};
     size_t theta_scale_nb[] = {sizeof(float), sizeof(float), sizeof(float),
@@ -2316,8 +2320,6 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         ctx.rope_cache.freq_scale != freq_scale) {
 
         ctx.rope_cache.theta_scale_length = theta_scale_length;
-        ctx.rope_cache.theta_scale = theta_scale;
-        ctx.rope_cache.freq_scale = freq_scale;
 
         if (ctx.rope_cache.theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(ctx.rope_cache.theta_scale_cache));
@@ -2342,7 +2344,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
             // return MIN(1, MAX(0, y)) - 1;
             yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
             void* yarn_ramp_buffer = yarn_ramp_allocator.get();
-            acl_yarn_ramp_tensor = ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float_t),
+            acl_yarn_ramp_tensor = ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float),
                                            theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
             float zero_value = 0, one_value = 1;
             float denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
@@ -2411,6 +2413,20 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         ggml_cann_release_resources(ctx, acl_freq_factors_tensor, acl_freq_fac_res_tensor);
     }
 
+    // init sin_repeat && cos_repeat, only to accelerate first layer on each device
+    if (position_length > ctx.rope_cache.position_length) {
+        ctx.rope_cache.position_length = position_length;
+        if (ctx.rope_cache.sin_cache != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_cache.sin_cache));
+        }
+        if (ctx.rope_cache.cos_cache != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_cache.cos_cache));
+        }
+        int64_t repeat_theta_length = theta_scale_length * position_length * 2;
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cache.sin_cache, repeat_theta_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cache.cos_cache, repeat_theta_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
+    }
+
     // position
     aclTensor* acl_position_tensor = ggml_cann_create_tensor(
         src1->data, ggml_cann_type_mapping(src1->type),
@@ -2462,10 +2478,10 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_repeat_tensor =
-        ggml_cann_create_tensor(sin_tensor_buffer, ACL_FLOAT, sizeof(float),
+        ggml_cann_create_tensor(ctx.rope_cache.sin_cache, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_repeat_tensor =
-        ggml_cann_create_tensor(cos_tensor_buffer, ACL_FLOAT, sizeof(float),
+        ggml_cann_create_tensor(ctx.rope_cache.cos_cache, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
     // repeat
@@ -2483,6 +2499,14 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                                 num_repeats, output_size);
     }
 
+    // Other layers use cache except first layer.
+    ctx.rope_cache.cached = true;
+    ctx.rope_cache.ext_factor = ext_factor;
+    ctx.rope_cache.theta_scale = theta_scale;
+    ctx.rope_cache.freq_scale = freq_scale;
+    ctx.rope_cache.attn_factor = attn_factor;
+    ctx.rope_cache.is_neox = is_neox;
+
     ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
         acl_theta_tensor, acl_sin_tensor, acl_sin_repeat_tensor, acl_cos_tensor,
         acl_cos_repeat_tensor);
@@ -2504,10 +2528,7 @@ aclnnStatus aclnnRotaryPositionEmbedding(void* workspace,
 #endif
 
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
-    // TODO: use ascendc
-    // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
-    ggml_tensor* src1 = dst->src[1];
 
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
@@ -2538,15 +2559,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
-    // sin/cos tensor length.
-    int64_t repeat_theta_length = src0->ne[0] * src1->ne[0];
-    ggml_cann_pool_alloc sin_tensor_allocator(ctx.pool(), repeat_theta_length * sizeof(float));
-    ggml_cann_pool_alloc cos_tensor_allocator(ctx.pool(), repeat_theta_length * sizeof(float));
-    void *sin_tensor_buffer = sin_tensor_allocator.get();
-    void *cos_tensor_buffer = cos_tensor_allocator.get();
-
     // init ctx.rope_cos/rope_sin cache
-    aclnn_cache_init(ctx, dst, sin_tensor_buffer, cos_tensor_buffer, corr_dims, ext_factor,
+    aclnn_cache_init(ctx, dst, corr_dims, ext_factor,
                     theta_scale, freq_scale, attn_factor, is_neox);
 
     int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
@@ -2556,10 +2570,10 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_reshape_tensor =
-        ggml_cann_create_tensor(sin_tensor_buffer, ACL_FLOAT, sizeof(float),
+        ggml_cann_create_tensor(ctx.rope_cache.sin_cache, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_reshape_tensor =
-        ggml_cann_create_tensor(cos_tensor_buffer, ACL_FLOAT, sizeof(float),
+        ggml_cann_create_tensor(ctx.rope_cache.cos_cache, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
     aclTensor* acl_src = ggml_cann_create_tensor(src0);

ggml/src/ggml-cann/common.h

@@ -38,6 +38,7 @@
 #include <unistd.h>
 #include <functional>
 #include <optional>
+#include <list>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -106,6 +107,7 @@ int32_t ggml_cann_get_device();
 
 std::optional<std::string> get_env(const std::string& name);
 bool parse_bool(const std::string& value);
+int parse_integer(const std::string& value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -350,7 +352,7 @@ struct ggml_graph_node_properties {
 struct ggml_cann_graph {
     ~ggml_cann_graph() {
         if (graph != nullptr) {
-            aclmdlRIDestroy(graph);
+            ACL_CHECK(aclmdlRIDestroy(graph));
         }
     }
 
@@ -358,6 +360,64 @@ struct ggml_cann_graph {
 
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
 };
+
+/**
+ * @brief LRU cache for managing ggml_cann_graph objects.
+ *
+ * This class maintains a list of shared_ptr to ggml_cann_graph objects
+ * and enforces a maximum capacity. It provides methods to push new graphs,
+ * move existing graphs to the front (most recently used), and clear the cache.
+ */
+struct ggml_cann_graph_lru_cache {
+    size_t capacity;  /**< Maximum number of graphs in the cache. */
+
+    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
+
+    ggml_cann_graph_lru_cache() {
+        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
+    }
+
+    /**
+     * @brief Push a new graph to the front of the cache.
+     * If the cache exceeds capacity, the least recently used graph is deleted.
+     * @param new_node Pointer to the new ggml_cann_graph to cache.
+     *        Ownership is transferred to the cache (cache will delete it).
+     */
+    void push(ggml_cann_graph* new_node) {
+        if (cache_list.size() >= capacity) {
+            ggml_cann_graph* old = cache_list.back();
+            cache_list.pop_back();
+            delete old; // free the old graph
+        }
+        cache_list.push_front(new_node);
+    }
+
+    /**
+     * @brief Move an existing graph to the front of the cache.
+     * @param node Pointer to the ggml_cann_graph to move.
+     */
+    void move_to_front(ggml_cann_graph* node) {
+        cache_list.remove(node);
+        cache_list.push_front(node);
+    }
+
+    /**
+     * @brief Clear all graphs from the cache (also frees memory).
+     */
+    void clear() {
+        for (auto ptr : cache_list) {
+            delete ptr;
+        }
+        cache_list.clear();
+    }
+
+    /**
+     * @brief Destructor that clears the cache and frees all cached graphs.
+     */
+    ~ggml_cann_graph_lru_cache() {
+        clear();
+    }
+};
 #endif  // USE_ACL_GRAPH
 
 struct ggml_cann_rope_cache {
@@ -365,12 +425,27 @@ struct ggml_cann_rope_cache {
         if(theta_scale_cache != nullptr) {
             ACL_CHECK(aclrtFree(theta_scale_cache));
         }
+        if(sin_cache != nullptr) {
+            ACL_CHECK(aclrtFree(sin_cache));
+        }
+        if(cos_cache != nullptr) {
+            ACL_CHECK(aclrtFree(cos_cache));
+        }
     }
 
     void* theta_scale_cache = nullptr;
     int64_t theta_scale_length = 0;
+    // sin/cos cache, used only to accelerate first layer on each device
+    void* sin_cache = nullptr;
+    void* cos_cache = nullptr;
+    int64_t position_length = 0;
+    // Properties to check before reusing the sincos cache
+    bool cached = false;
+    float ext_factor = 0.0f;
     float theta_scale = 0.0f;
     float freq_scale = 0.0f;
+    float attn_factor = 0.0f;
+    bool is_neox = false;
 };
 
 struct ggml_cann_tensor_cache {
@@ -394,7 +469,7 @@ struct ggml_backend_cann_context {
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
-    std::unique_ptr<ggml_cann_graph> cann_graph;
+    ggml_cann_graph_lru_cache graph_lru_cache;
     bool acl_graph_mode = true;
 #endif
     cann_task_queue task_queue;
@@ -451,7 +526,10 @@ struct ggml_backend_cann_context {
      */
     aclrtStream stream(int stream) {
         if (streams[stream] == nullptr) {
-            ggml_cann_set_device(device);
+            // If the device is not set here, destroying the stream later may cause a mismatch
+            // between the thread contexts where the stream was created and destroyed.
+            // However, I printed the device_id, thread_id, and stream, and they are all consistent.
+            ACL_CHECK(aclrtSetDevice(device));
             ACL_CHECK(aclrtCreateStream(&streams[stream]));
         }
         return streams[stream];

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

@@ -75,13 +75,12 @@
  * @param device The device ID to set.
  */
 void ggml_cann_set_device(const int32_t device) {
-    // TODO: uncomment these lines after empty context has fixed.
-    // int current_device;
-    // ACL_CHECK(aclrtGetDevice(&current_device));
+    int current_device = -1;
+    aclrtGetDevice(&current_device);
 
-    // if (device == current_device) {
-    //   return;
-    // }
+    if (device == current_device) {
+      return;
+    }
     ACL_CHECK(aclrtSetDevice(device));
 }
 
@@ -116,6 +115,24 @@ bool parse_bool(const std::string& value) {
     return valid_values.find(value) != valid_values.end();
 }
 
+/**
+ * @brief Parse a string as an integer, returning 0 if invalid.
+ *
+ * This function attempts to convert the input string `value` to an `int`.
+ * If the string is not a valid integer or is out of the `int` range,
+ * it returns 0.
+ *
+ * @param value The string to parse.
+ * @return The parsed integer, or 0 if conversion fails.
+ */
+int parse_integer(const std::string& value) {
+    try {
+        return std::stoi(value);
+    } catch (...) {
+        return 0;
+    }
+}
+
 /**
  * @brief Initialize the CANN device information.
  *
@@ -1711,6 +1728,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_get_rows(ctx, dst);
             break;
         case GGML_OP_SET_ROWS:
+            std::cout << "lcg GGML_OP_SET_ROWS"<< std::endl;
             ggml_cann_set_rows(ctx, dst);
             break;
         case GGML_OP_DUP:
@@ -2092,16 +2110,17 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
-
         // record event on src stream after the copy
-        if (!cann_ctx_src->copy_event) {
-            ACL_CHECK(aclrtCreateEventWithFlag(&cann_ctx_src->copy_event, ACL_EVENT_SYNC));
-        }
-        ACL_CHECK(aclrtRecordEvent(cann_ctx_src->copy_event, cann_ctx_src->stream()));
+        // TODO: this event is not effective with acl graph mode, change to use aclrtSynchronizeStream
+        // if (!cann_ctx_src->copy_event) {
+        //     ACL_CHECK(aclrtCreateEventWithFlag(&cann_ctx_src->copy_event, ACL_EVENT_SYNC));
+        // }
+        // ACL_CHECK(aclrtRecordEvent(cann_ctx_src->copy_event, cann_ctx_src->stream()));
 
-        // wait on dst stream for the copy to complete
-        ggml_cann_set_device(cann_ctx_dst->device);
-        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx_dst->stream(), cann_ctx_src->copy_event));
+        // // wait on dst stream for the copy to complete
+        // ggml_cann_set_device(cann_ctx_dst->device);
+        // ACL_CHECK(aclrtStreamWaitEvent(cann_ctx_dst->stream(), cann_ctx_src->copy_event));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx_src->stream()));
     } else {
         // src and dst are on the same backend
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
@@ -2130,30 +2149,52 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
 
 #ifdef USE_ACL_GRAPH
 /**
- * @brief Populate the internal CANN graph node properties from the ggml computation graph.
+ * @brief Add a new CANN graph to the LRU cache by populating node properties from the ggml graph.
  *
- * This function copies all node attributes (operation type, dimensions, strides, input sources,
- * and operation parameters) into the cached CANN graph structure for later reuse or comparison.
+ * This function creates a new ggml_cann_graph object and fills its node properties
+ * (operation type, dimensions, strides, input sources, and operation parameters)
+ * based on the current ggml computation graph.
  *
- * @param cann_ctx  The CANN backend context.
- * @param cgraph    The ggml computational graph.
+ * Each node in the ggml graph is mapped to a property entry in the new CANN graph:
+ * - node address
+ * - operation type
+ * - shape (ne) and strides (nb)
+ * - source tensor addresses
+ * - operation parameters
+ *
+ * After initialization, the new graph is pushed into the LRU cache owned by the
+ * CANN backend context. The cache takes ownership of the graph and manages its
+ * lifetime (including deletion upon eviction).
+ *
+ * @param cann_ctx  The CANN backend context containing the graph cache.
+ * @param cgraph    The current ggml computation graph.
  */
-static void set_ggml_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
-        ggml_tensor * node = cgraph->nodes[node_idx];
-        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_address = node->data;
-        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_op = node->op;
+static void add_lru_matched_graph_node_properties(
+        ggml_backend_cann_context * cann_ctx,
+        ggml_cgraph * cgraph) {
+    // Create a new ggml_cann_graph object on the heap (its lifetime is managed by the cache).
+    ggml_cann_graph * new_graph = new ggml_cann_graph();
+    new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
 
-        for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].ne[dim] = node->ne[dim];
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].nb[dim] = node->nb[dim];
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
+        ggml_tensor * node = cgraph->nodes[node_idx];
+        auto & prop = new_graph->ggml_graph_properties[node_idx];
+
+        prop.node_address = node->data;
+        prop.node_op      = node->op;
+
+        std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
+        std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
+
+        for (int src = 0; src < GGML_MAX_SRC; ++src) {
+            prop.src_address[src] = node->src[src] ? node->src[src]->data : nullptr;
         }
-        for (int src = 0; src < GGML_MAX_SRC; src++) {
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].src_address[src] =
-                node->src[src] ? node->src[src]->data : nullptr;
-        }
-        memcpy(cann_ctx->cann_graph->ggml_graph_properties[node_idx].op_params, node->op_params, GGML_MAX_OP_PARAMS);
+
+        memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
     }
+
+    // Insert into the LRU cache (cache takes ownership and will delete it when evicted).
+    cann_ctx->graph_lru_cache.push(new_graph);
 }
 
 /**
@@ -2198,30 +2239,45 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
 }
 
 /**
- * @brief Determine if the CANN graph needs to be rebuilt due to graph changes.
+ * @brief Check whether there is a cached CANN graph that matches the current ggml graph.
  *
- * This checks whether the number or properties of ggml graph nodes have changed
- * compared to the last captured CANN graph. If so, the CANN graph must be re-captured.
+ * This function iterates through the cached CANN graphs stored in the LRU cache and
+ * compares them against the given ggml computation graph. A match requires that the
+ * number of nodes is the same and that each node’s properties (operation type,
+ * dimensions, strides, inputs, and operation parameters) are identical.
  *
- * @param cann_ctx  The CANN backend context.
+ * If a matching graph is found, it is promoted to the front of the LRU cache and the
+ * function returns true. Otherwise, the function returns false, indicating that a new
+ * CANN graph needs to be captured.
+ *
+ * @param cann_ctx  The CANN backend context containing the graph cache.
  * @param cgraph    The current ggml computation graph.
- * @return true if an update is required; false otherwise.
+ * @return true if a matching cached graph exists; false otherwise.
  */
-static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    // The number of nodes is different, so the graph needs to be reconstructed.
-    if (cann_ctx->cann_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
-        cann_ctx->cann_graph->ggml_graph_properties.resize(cgraph->n_nodes);
-        return true;
-    }
+static bool is_matched_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    ggml_cann_graph_lru_cache &lru_cache = cann_ctx->graph_lru_cache;
+    for (auto &graph_ptr : lru_cache.cache_list) {
+        // Skip graphs with a different number of nodes.
+        if (graph_ptr->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
+            continue;
+        }
 
-    // The number of nodes is the same; iterate over each node to check whether they match.
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        bool has_matching_properties = ggml_graph_node_has_matching_properties(
-            cgraph->nodes[i], &cann_ctx->cann_graph->ggml_graph_properties[i]);
-        if(!has_matching_properties) {
+        // Check if all nodes match.
+        bool all_match = true;
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            if (!ggml_graph_node_has_matching_properties(cgraph->nodes[i], &graph_ptr->ggml_graph_properties[i])) {
+                all_match = false;
+                break;
+            }
+        }
+
+        if (all_match) {
+            // update cache_list && renturn graph_ptr
+            lru_cache.move_to_front(graph_ptr);
             return true;
         }
     }
+
     return false;
 }
 #endif  // USE_ACL_GRAPH
@@ -2240,17 +2296,13 @@ static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx,
  * @param cann_graph_update_required Whether graph capture is needed due to graph changes.
  */
 static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph,
-    bool & use_cann_graph, bool & cann_graph_update_required) {
+    bool & use_cann_graph,  bool & cann_graph_update_required) {
 #ifdef USE_ACL_GRAPH
+    ggml_cann_graph* matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
     if (use_cann_graph && cann_graph_update_required) {
-        if (cann_ctx->cann_graph->graph != nullptr) {
-            ACL_CHECK(aclmdlRIDestroy(cann_ctx->cann_graph->graph));
-            cann_ctx->cann_graph->graph = nullptr;
-        }
         ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
     }
 #endif // USE_ACL_GRAPH
-
     // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
     // With the use of CANN graphs, the execution will be performed by the graph launch.
     if (!use_cann_graph || cann_graph_update_required) {
@@ -2271,12 +2323,12 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
 
 #ifdef USE_ACL_GRAPH
     if (use_cann_graph && cann_graph_update_required) { // End CANN graph capture
-        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &cann_ctx->cann_graph->graph));
+        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &matched_graph->graph));
     }
 
     if (use_cann_graph) {
         // Execute graph
-        ACL_CHECK(aclmdlRIExecuteAsync(cann_ctx->cann_graph->graph, cann_ctx->stream()));
+        ACL_CHECK(aclmdlRIExecuteAsync(matched_graph->graph, cann_ctx->stream()));
     }
 #endif // USE_ACL_GRAPH
 }
@@ -2301,28 +2353,44 @@ static enum ggml_status ggml_backend_cann_graph_compute(
     ggml_cann_set_device(cann_ctx->device);
     g_nz_workspaces[cann_ctx->device].clear();
 
+    // calculate rope cache for fist layer in current device.
+    cann_ctx->rope_cache.cached = false;
+
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph = true;
     bool cann_graph_update_required = false;
 
+    static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
+    if (!prefill_use_graph) {
+        // Do not use acl_graph for prefill.
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+            // TODO: Optimize here. Currently, we can only
+            // get seq_len by FA's input.
+            if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+                // Q -> src[0], shape: [B, S, N, D]
+                use_cann_graph = (node->src[0]->ne[1] == 1);
+                break;
+            }
+        }
+    }
+
     if (!cann_ctx->acl_graph_mode) {
         use_cann_graph = false;
     }
 
     if (use_cann_graph) {
-        if (cann_ctx->cann_graph == nullptr) {
-            cann_ctx->cann_graph.reset(new ggml_cann_graph());
-            cann_graph_update_required = true;
+        // If no matching graph is found, the graph needs to be recaptured.
+        cann_graph_update_required = !is_matched_graph(cann_ctx, cgraph);
+        if (cann_graph_update_required) {
+            // If no matching graph is found, add a new ACL graph.
+            add_lru_matched_graph_node_properties(cann_ctx, cgraph);
         }
-
-        cann_graph_update_required = is_cann_graph_update_required(cann_ctx, cgraph);
-        set_ggml_graph_node_properties(cann_ctx, cgraph);
     }
 #else
     bool use_cann_graph = false;
     bool cann_graph_update_required = false;
 #endif  // USE_ACL_GRAPH
-
     evaluate_and_capture_cann_graph(
         cann_ctx,
         cgraph,
@@ -2689,6 +2757,7 @@ static const ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
+    /* .optimize_graph          = */ NULL,
 };
 
 /**

ggml/src/ggml-cpu/CMakeLists.txt

@@ -224,7 +224,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
                     string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
                     if (NOT ${feature_pos} EQUAL -1)
-                        message(STATUS "ARM feature ${feature} enabled")
+                        # Special handling for MATMUL_INT8 when machine doesn't support i8mm
+                        if ("${feature}" STREQUAL "MATMUL_INT8" AND GGML_MACHINE_SUPPORTS_noi8mm)
+                            message(STATUS "ARM feature ${feature} detected but unsetting due to machine not supporting i8mm")
+                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_MATMUL_INT8)
+                        else()
+                            message(STATUS "ARM feature ${feature} enabled")
+                        endif()
                     endif()
                 endforeach()
             endif()
@@ -457,7 +463,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
-            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -479,11 +484,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
             list(APPEND ARCH_DEFINITIONS GGML_VXE)
         endif()
-
-        if (GGML_NNPA)
-            message(STATUS "NNPA enabled")
-            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
-        endif()
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -53,9 +53,9 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__VXE__) || defined(__VXE2__)
     for (int i = 0; i < nb; i++) {
-        __vector float srcv [8];
-        __vector float asrcv[8];
-        __vector float amaxv[8];
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
 
         for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
         for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
@@ -74,8 +74,8 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
-            const __vector float v = vec_mul(srcv[j], vec_splats(id));
-            const __vector int32_t vi = vec_signed(v);
+            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
+            const int32x4_t vi = vec_signed(v);
 
             y[i].qs[4*j + 0] = vec_extract(vi, 0);
             y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -98,9 +98,9 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__VXE__) || defined(__VXE2__)
     for (int i = 0; i < nb; i++) {
-        __vector float srcv [8];
-        __vector float asrcv[8];
-        __vector float amaxv[8];
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
 
         for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
         for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
@@ -118,11 +118,11 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
-        __vector int32_t acc = vec_splats(0);
+        int32x4_t acc = vec_splats(0);
 
         for (int j = 0; j < 8; j++) {
-            const __vector float v = vec_mul(srcv[j], vec_splats(id));
-            const __vector int32_t vi = vec_signed(v);
+            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
+            const int32x4_t vi = vec_signed(v);
 
             y[i].qs[4*j + 0] = vec_extract(vi, 0);
             y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -162,37 +162,36 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
 #if defined(__VXE__) || defined(__VXE2__)
-    __vector float acc = vec_splats(0.0f);
+    float32x4_t acc = vec_splats(0.0f);
 
-    const __vector uint8_t v_m = vec_splats((const uint8_t)0x0F);
-    const __vector int8_t  v_s = vec_splats( (const int8_t)0x08);
+    const uint8x16_t v_m = vec_splats((const uint8_t)0x0F);
+    const int8x16_t  v_s = vec_splats( (const int8_t)0x08);
 
     for (; ib < nb; ++ib) {
-        const __vector uint8_t v_x = vec_xl(0, x[ib].qs);
-        const __vector int8_t v_xl = (const __vector int8_t)(v_x & v_m);
-        const __vector int8_t v_xh = (const __vector int8_t)(v_x >> 4);
+        const uint8x16_t v_x = vec_xl(0, x[ib].qs);
+        const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
+        const int8x16_t v_xh = (const int8x16_t)(v_x >> 4);
 
-        const __vector int8_t v_xls = vec_sub(v_xl, v_s);
-        const __vector int8_t v_xhs = vec_sub(v_xh, v_s);
+        const int8x16_t v_xls = vec_sub(v_xl, v_s);
+        const int8x16_t v_xhs = vec_sub(v_xh, v_s);
 
-        const __vector int8_t v_yl = vec_xl(0      , y[ib].qs);
-        const __vector int8_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
 
-        const __vector int16_t v_xylso = vec_mulo(v_xls, v_yl);
-        const __vector int16_t v_xylse = vec_mule(v_xls, v_yl);
-        const __vector int16_t v_xyhso = vec_mulo(v_xhs, v_yh);
-        const __vector int16_t v_xyhse = vec_mule(v_xhs, v_yh);
+        const int16x8_t v_xylso = vec_mulo(v_xls, v_yl);
+        const int16x8_t v_xylse = vec_mule(v_xls, v_yl);
+        const int16x8_t v_xyhso = vec_mulo(v_xhs, v_yh);
+        const int16x8_t v_xyhse = vec_mule(v_xhs, v_yh);
 
-        __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
+        int16x8_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
 
-        const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
-        const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_xy = vec_float(vec_unpackh(v_xy_));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3];
-
+    sumf = vec_hsum_f32x4(acc);
     *s = sumf;
 #else
     UNUSED(nb);
@@ -249,8 +248,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
-
+    sumf = vec_hsum_f32x4(acc) + summs;
     *s = sumf;
 #else
     UNUSED(nb);
@@ -351,7 +349,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
     }
 
-    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1);
+    sumf += vec_hsum_f32x4(v_sum0) + vec_hsum_f32x4(v_sum1);
 
     #pragma GCC unroll 4
     for (; ib < nb; ++ib) {
@@ -390,7 +388,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
         const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));
 
-        sumf += vec_hsum(v_acc);
+        sumf += vec_hsum_f32x4(v_acc);
     }
 
     *s = sumf;
@@ -502,7 +500,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
     }
 
-    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1) + summs0 + summs1;
+    sumf += vec_hsum_f32x4(v_sum0) + vec_hsum_f32x4(v_sum1) + summs0 + summs1;
 
     #pragma GCC unroll 4
     for (; ib < nb; ++ib) {
@@ -543,7 +541,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
         const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
 
-        sumf += vec_hsum(v_acc) + summs;
+        sumf += vec_hsum_f32x4(v_acc) + summs;
     }
 
     *s = sumf;
@@ -575,7 +573,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
 #if defined(__VXE__) || defined(__VXE2__)
-    __vector float acc = vec_splats(0.0f);
+    float32x4_t acc = vec_splats(0.0f);
 
 #pragma GCC unroll 8
     for (; ib < nb; ++ib) {
@@ -594,7 +592,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3];
+    sumf = vec_hsum_f32x4(acc);
 
     *s = sumf;
 #else
@@ -718,10 +716,10 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
             isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
 
-            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
-            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
-            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
-            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+            isum += vec_hsum_i32x4(isum0) * scale[0];
+            isum += vec_hsum_i32x4(isum1) * scale[1];
+            isum += vec_hsum_i32x4(isum2) * scale[2];
+            isum += vec_hsum_i32x4(isum3) * scale[3];
 
             scale += 4;
 
@@ -819,7 +817,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             v_xl[1] = (int8x16_t)vec_and(v_x[1], v_lm);
 
             const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
-            sumi1 += (p1[0] + p1[1] + p1[2] + p1[3]) * scales[2*j+0];
+            sumi1 += vec_hsum_i32x4(p1) * scales[2*j+0];
 
             v_y[0] = vec_xl(0 , y0);
             v_y[1] = vec_xl(16, y0);
@@ -829,7 +827,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             v_xl[1] = (int8x16_t)vec_sr(v_x[1], 4);
 
             const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
-            sumi2 += (p2[0] + p2[1] + p2[2] + p2[3]) * scales[2*j+1];
+            sumi2 += vec_hsum_i32x4(p2) * scales[2*j+1];
         }
 
         sumf += d * (sumi1 + sumi2);
@@ -911,7 +909,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh);
         const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh);
         const int32x4_t v_mins = vec_add(v_minsho, v_minshe);
-        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+        const int32_t mins = vec_hsum_i32x4(v_mins);
 
         const uint8_t * scales = (const uint8_t *)utmp;
         const uint8_t * GGML_RESTRICT x0l = x[i].qs;
@@ -948,8 +946,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int32x4_t sumi0 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[0], v_y[0]), q5b[1], v_y[1]);
             int32x4_t sumi1 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[2], v_y[2]), q5b[3], v_y[3]);
 
-            sumi += (sumi0[0] + sumi0[1] + sumi0[2] + sumi0[3]) * *scales++;
-            sumi += (sumi1[0] + sumi1[1] + sumi1[2] + sumi1[3]) * *scales++;
+            sumi += vec_hsum_i32x4(sumi0) * *scales++;
+            sumi += vec_hsum_i32x4(sumi1) * *scales++;
         }
 
         sumf += d * sumi - dmin * mins;
@@ -1020,7 +1018,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh);
         const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe;
 
-        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+        const int32_t mins = vec_hsum_i32x4(v_mins);
 
         int32_t isum = 0;
         for (int j = 0; j < QK_K/128; ++j) {
@@ -1060,10 +1058,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int32x4_t summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
             int32x4_t summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
 
-            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
-                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
-                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
-                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+            isum += vec_hsum_i32x4(summs0) * scale[0] +
+                    vec_hsum_i32x4(summs1) * scale[1] +
+                    vec_hsum_i32x4(summs2) * scale[2] +
+                    vec_hsum_i32x4(summs3) * scale[3];
 
             scale += 4;
 
@@ -1094,10 +1092,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
             summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
 
-            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
-                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
-                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
-                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+            isum += vec_hsum_i32x4(summs0) * scale[0] +
+                    vec_hsum_i32x4(summs1) * scale[1] +
+                    vec_hsum_i32x4(summs2) * scale[2] +
+                    vec_hsum_i32x4(summs3) * scale[3];
 
             scale += 4;
         }
@@ -1285,7 +1283,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
         const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
 
-        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * vec_hsum_i32x4(v_xy);
     }
 
     *s = sumf;
@@ -1354,8 +1352,8 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
             h >>= 4;
 
-            sumi1 += (vsumi0[0] + vsumi0[1] + vsumi0[2] + vsumi0[3]) * ls1;
-            sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
+            sumi1 += vec_hsum_i32x4(vsumi0) * ls1;
+            sumi2 += vec_hsum_i32x4(vsumi1) * ls2;
         }
 
         sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);

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

@@ -68,12 +68,6 @@ struct ggml_compute_params {
 #endif  // __VXE2__
 #endif  // __s390x__ && __VEC__
 
-#if defined(__s390x__) && defined(GGML_NNPA)
-#ifndef __NNPA__
-#define __NNPA__
-#endif  // __NNPA__
-#endif  // __s390x__ && GGML_NNPA
-
 #if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
 #endif
@@ -489,11 +483,16 @@ inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
 /**
  * @see https://github.com/ggml-org/llama.cpp/pull/14037
  */
-inline static float vec_hsum(float32x4_t v) {
+inline static float vec_hsum_f32x4(float32x4_t v) {
     float32x4_t v_temp = v + vec_reve(v);
     return v_temp[0] + v_temp[1];
 }
 
+inline static int32_t vec_hsum_i32x4(int32x4_t v) {
+    int32x4_t v_temp = v + vec_reve(v);
+    return v_temp[0] + v_temp[1];
+}
+
 inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
     const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
     return acc + (vec_unpackh(p) + vec_unpackl(p));

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

@@ -373,6 +373,9 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_I32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_i32,
+    },
 };
 
 const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type) {
@@ -2696,7 +2699,10 @@ struct ggml_cplan ggml_graph_plan(
                         if (ggml_is_quantized(node->type) ||
                             // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
                             (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
-                            (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
+                            (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16) ||
+                            // conversion between F32 and I32
+                            (node->src[0]->type == GGML_TYPE_F32 && node->src[1] && node->src[1]->type == GGML_TYPE_I32) ||
+                            (node->src[0]->type == GGML_TYPE_I32 && node->src[1] && node->src[1]->type == GGML_TYPE_F32)) {
                             cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                         }
                     } break;
@@ -3211,21 +3217,6 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
         _mm_storel_epi64((__m128i *)(y + i), y_vec);
     }
-#elif defined(__NNPA__)
-    for (; i + 7 < n; i += 8) {
-        float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
-        float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
-        uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
-        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
-        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
-    }
-    for (; i + 3 < n; i += 4) {
-        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
-        float32x4_t v_zero = vec_splats(0.0f);
-        uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
-        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
-        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
-    }
 #elif defined(__riscv_zvfh)
     for (int vl; i < n; i += vl) {
         vl = __riscv_vsetvl_e32m2(n - i);
@@ -3259,21 +3250,6 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
-#elif defined(__NNPA__)
-    for (; i + 7 < n; i += 8) {
-        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
-        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
-        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
-        float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
-        vec_xst(v_yh, 0, (float *)(y + i + 0));
-        vec_xst(v_yl, 0, (float *)(y + i + 4));
-    }
-    for (; i + 3 < n; i += 4) {
-        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
-        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
-        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
-        vec_xst(v_yh, 0, (float *)(y + i));
-    }
 #endif
 
     for (; i < n; ++i) {
@@ -3288,6 +3264,13 @@ void ggml_cpu_fp32_to_bf16(const float * x, ggml_bf16_t * y, int64_t n) {
     }
 }
 
+void ggml_cpu_fp32_to_i32(const float * x, int32_t * y, int64_t n) {
+    int64_t i = 0;
+    for (; i < n; ++i) {
+        y[i] = x[i];
+    }
+}
+
 void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
     int64_t i = 0;
 #if defined(__AVX2__)
@@ -3477,14 +3460,6 @@ int ggml_cpu_has_vxe(void) {
 #endif
 }
 
-int ggml_cpu_has_nnpa(void) {
-#if defined(GGML_NNPA)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return 1;

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

@@ -190,6 +190,7 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
     /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .optimize_graph          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_cpu_guid(void) {
@@ -348,8 +349,10 @@ static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t *
     long pages = sysconf(_SC_PHYS_PAGES);
     long page_size = sysconf(_SC_PAGE_SIZE);
     *total = pages * page_size;
+
+    // "free" system memory is ill-defined, for practical purposes assume that all of it is free:
     *free = *total;
-#endif
+#endif // _WIN32
 
     GGML_UNUSED(dev);
 }
@@ -576,9 +579,6 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_vxe()) {
             features.push_back({ "VXE", "1" });
         }
-        if (ggml_cpu_has_nnpa()) {
-            features.push_back({ "NNPA", "1" });
-        }
         if (ggml_cpu_has_wasm_simd()) {
             features.push_back({ "WASM_SIMD", "1" });
         }

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

@@ -154,7 +154,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             if (dst->src[0]->type == GGML_TYPE_Q4_0) {
                 return compute_forward_q4_0(params, dst);
             } else if (dst->src[0]->type == GGML_TYPE_F16) {
-                return compute_forward_kv_cache(params, dst);
+                return compute_forward_fp16(params, dst);
             }
         } else if (dst->op == GGML_OP_GET_ROWS) {
             if (dst->src[0]->type == GGML_TYPE_Q4_0) {
@@ -164,7 +164,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         return false;
     }
 
-    bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) {
+    bool compute_forward_fp16(ggml_compute_params * params, struct ggml_tensor * dst) {
         static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
 
         const ggml_tensor * src0 = dst->src[0];
@@ -515,9 +515,6 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             op->src[0]->buffer &&
             (ggml_n_dims(op->src[0]) == 2) &&
             op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
-            if (op->op == GGML_OP_GET_ROWS && op->src[1]->ne[0] != 8) {
-                return false;
-            }
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
@@ -534,13 +531,8 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
-            else if (ggml_kleidiai_select_kernels(ctx.features, op) &&
-                     op->src[0]->op == GGML_OP_VIEW &&
-                     (op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op ==  GGML_OP_SOFT_MAX) &&
-                     op->src[1]->ne[1] > 1) {
-                if ((op->src[0]->nb[0] != 2) ||
-                    (op->src[1]->nb[0] != 4) ||
-                    (op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
+            else if (ggml_kleidiai_select_kernels(ctx.features, op) && op->src[1]->ne[1] > 1) {
+                if ((op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
                     (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
                     return nullptr;
                 }

ggml/src/ggml-cpu/ops.cpp

@@ -776,6 +776,24 @@ static void ggml_compute_forward_dup_f32(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
+            } else if (dst->type == GGML_TYPE_I32) {
+                size_t id = 0;
+                int32_t * dst_ptr = (int32_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = *src0_ptr;
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
             } else {
                 GGML_ABORT("fatal error"); // TODO: implement
             }
@@ -947,6 +965,144 @@ static void ggml_compute_forward_dup_f32(
                 }
             }
         }
+    } else if (dst->type == GGML_TYPE_I32) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(int32_t *) dst_ptr = *(const float *) src0_ptr;
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    } else {
+        GGML_ABORT("fatal error"); // TODO: implement
+    }
+}
+
+static void ggml_compute_forward_dup_i32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const int ith = params->ith; // thread index
+    const int nth = params->nth; // number of threads
+
+    // parallelize by rows
+    const int nr = ne01;
+    // number of rows per thread
+    const int dr = (nr + nth - 1) / nth;
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    // dst counters
+
+    int64_t i10 = 0;
+    int64_t i11 = 0;
+    int64_t i12 = 0;
+    int64_t i13 = 0;
+
+    // TODO: not optimal, but works
+    if (dst->type == GGML_TYPE_F32) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(float *) dst_ptr = *(const int32_t *) src0_ptr;
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
     } else {
         GGML_ABORT("fatal error"); // TODO: implement
     }
@@ -1177,6 +1333,10 @@ void ggml_compute_forward_dup(
             {
                 ggml_compute_forward_dup_f32(params, dst);
             } break;
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_dup_i32(params, dst);
+            } break;
         default:
             {
                 if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) {
@@ -8438,7 +8598,7 @@ static void ggml_compute_forward_timestep_embedding_f32(
             embed_data[j + half] = sinf(arg);
         }
         if (dim % 2 != 0 && ith == 0) {
-            embed_data[dim] = 0.f;
+            embed_data[2 * half] = 0.f;
         }
     }
 }

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

@@ -114,26 +114,6 @@ extern "C" {
     #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
     #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
     #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
-#elif defined(__NNPA__)
-    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
-    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x)
-
-    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
-    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
-
-    static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) {
-        uint16x8_t v_h = vec_splats(h);
-        uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
-        return vec_extend_to_fp32_hi(v_hd, 0)[0];
-    }
-
-    static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) {
-        float32x4_t v_f = vec_splats(f);
-        float32x4_t v_zero = vec_splats(0.0f);
-        uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
-        uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
-        return vec_extract(v_h, 0);
-    }
 #endif
 
 // precomputed f32 table for f16 (256 KB)
@@ -1156,11 +1136,6 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_EPR  GGML_F32_EPR
 
 static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
-#if defined(__NNPA__)
-    uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x);
-    uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
-    return vec_extend_to_fp32_hi(v_xd, 0);
-#else
     float tmp[4];
 
     for (int i = 0; i < 4; i++) {
@@ -1170,20 +1145,9 @@ static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
     // note: keep type-cast here to prevent compiler bugs
     // see: https://github.com/ggml-org/llama.cpp/issues/12846
     return vec_xl(0, (const float *)(tmp));
-#endif
 }
 
 static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
-#if defined(__NNPA__)
-    float32x4_t v_zero = vec_splats(0.0f);
-    uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0);
-    uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0);
-
-    x[0] = vec_extract(v_x, 0);
-    x[1] = vec_extract(v_x, 1);
-    x[2] = vec_extract(v_x, 2);
-    x[3] = vec_extract(v_x, 3);
-#else
     float arr[4];
 
     // note: keep type-cast here to prevent compiler bugs
@@ -1193,7 +1157,6 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
     for (int i = 0; i < 4; i++) {
         x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
     }
-#endif
 }
 
 #define GGML_F16_VEC                GGML_F32x4

ggml/src/ggml-cuda/CMakeLists.txt

@@ -44,6 +44,8 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/mmq*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmf*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
 
     if (GGML_CUDA_FA_ALL_QUANTS)
         file(GLOB   SRCS "template-instances/fattn-vec*.cu")

ggml/src/ggml-cuda/binbcast.cu

@@ -23,28 +23,44 @@ static __device__ __forceinline__ float op_div(const float a, const float b) {
     return a / b;
 }
 
+template <float (*bin_op)(const float, const float),
+          typename src0_t,
+          typename src1_t,
+          typename dst_t,
+          typename... src1_ptrs>
+static __global__ void k_bin_bcast(const src0_t *         src0,
+                                   const src1_t *         src1,
+                                   dst_t *                dst,
+                                   const int              ne0,
+                                   const int              ne1,
+                                   const int              ne2,
+                                   const uint3            ne3,
+                                   const uint3            ne10,
+                                   const uint3            ne11,
+                                   const uint3            ne12,
+                                   const uint3            ne13,
+                                   /*int s0, */ const int s1,
+                                   const int              s2,
+                                   const int              s3,
+                                   /*int s00,*/ const int s01,
+                                   const int              s02,
+                                   const int              s03,
+                                   /*int s10,*/ const int s11,
+                                   const int              s12,
+                                   const int              s13,
+                                   src1_ptrs... src1s) {
+    const uint32_t i0s = blockDim.x * blockIdx.x + threadIdx.x;
+    const uint32_t i1  = (blockDim.y * blockIdx.y + threadIdx.y);
+    const uint32_t i2  = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
+    const uint32_t i3  = (blockDim.z * blockIdx.z + threadIdx.z) - (i2 * ne3.z);
 
-
-template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t, typename... src1_ptrs>
-static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        const int ne0, const int ne1, const int ne2, const int ne3,
-        const int ne10, const int ne11, const int ne12, const int ne13,
-        /*int s0, */ const int s1, const int s2, const int s3,
-        /*int s00,*/ const int s01, const int s02, const int s03,
-        /*int s10,*/ const int s11, const int s12, const int s13,
-        src1_ptrs... src1s) {
-    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
-    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
-    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3.z) {
         return;
     }
 
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
+    const uint32_t i11 = fastmodulo(i1, ne11);
+    const uint32_t i12 = fastmodulo(i2, ne12);
+    const uint32_t i13 = fastmodulo(i3, ne13);
 
     const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
     const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
@@ -53,8 +69,8 @@ static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst
     const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
     dst_t * dst_row = dst + i_dst;
 
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
-        const int i10 = i0 % ne10;
+    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
+        const uint32_t i10 = fastmodulo(i0, ne10);
 
         float result = src0_row ? (float) src0_row[i0] : 0.0f;
         if constexpr (sizeof...(src1_ptrs) > 0) {
@@ -67,28 +83,48 @@ static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst
     }
 }
 
-template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t, typename... src1_ptrs>
-static __global__ void k_bin_bcast_unravel(const src0_t *   src0, const src1_t *   src1, dst_t *          dst,
-        const int ne0, const int ne1, const int ne2,const int ne3,
-        const int ne10, const int ne11, const int ne12, const int ne13,
-        /*int s0, */ const int s1, const int s2, const int s3,
-        /*int s00,*/ const int s01, const int s02, const int s03,
-        /*int s10,*/ const int s11, const int s12, const int s13,
-        src1_ptrs ... src1s) {
+template <float (*bin_op)(const float, const float),
+          typename src0_t,
+          typename src1_t,
+          typename dst_t,
+          typename... src1_ptrs>
+static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
+                                           const src1_t *         src1,
+                                           dst_t *                dst,
+                                           const uint3            ne0,
+                                           const uint3            ne1,
+                                           const uint3            ne2,
+                                           const uint32_t         ne3,
+                                           const uint3            prod_012,
+                                           const uint3            prod_01,
+                                           const uint3            ne10,
+                                           const uint3            ne11,
+                                           const uint3            ne12,
+                                           const uint3            ne13,
+                                           /*int s0, */ const int s1,
+                                           const int              s2,
+                                           const int              s3,
+                                           /*int s00,*/ const int s01,
+                                           const int              s02,
+                                           const int              s03,
+                                           /*int s10,*/ const int s11,
+                                           const int              s12,
+                                           const int              s13,
+                                           src1_ptrs... src1s) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
+    const uint32_t i3 = fastdiv(i, prod_012);
+    const uint32_t i2 = fastdiv(i - i3 * prod_012.z, prod_01);
+    const uint32_t i1 = fastdiv(i - i3 * prod_012.z - i2 * prod_01.z, ne0);
+    const uint32_t i0 = i - i3 * prod_012.z - i2 * prod_01.z - i1 * ne0.z;
 
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+    if (i0 >= ne0.z || i1 >= ne1.z || i2 >= ne2.z || i3 >= ne3) {
         return;
     }
 
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
+    const int i11 = fastmodulo(i1, ne11);
+    const int i12 = fastmodulo(i2, ne12);
+    const int i13 = fastmodulo(i3, ne13);
 
     const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
     const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
@@ -97,7 +133,7 @@ static __global__ void k_bin_bcast_unravel(const src0_t *   src0, const src1_t *
     const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
     dst_t * dst_row = dst + i_dst;
 
-    const int i10 = i0 % ne10;
+    const int i10 = fastmodulo(i0, ne10);
 
     float result = src0_row ? (float) src0_row[i0] : 0.0f;
     if constexpr (sizeof...(src1_ptrs) > 0) {
@@ -170,11 +206,6 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
         //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
 
-        int64_t ne10 = cne1[0];
-        int64_t ne11 = cne1[1];
-        int64_t ne12 = cne1[2];
-        int64_t ne13 = cne1[3];
-
         size_t nb0 = cnb[0];
         size_t nb1 = cnb[1];
         size_t nb2 = cnb[2];
@@ -233,48 +264,51 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
         block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
         block_dims.z = std::min(std::min<unsigned int>(ne2 * ne3, block_size / block_dims.x / block_dims.y), 64U);
 
-        dim3 block_nums((hne0 + block_dims.x - 1) / block_dims.x,
-                        (ne1 + block_dims.y - 1) / block_dims.y,
+        dim3 block_nums((hne0 + block_dims.x - 1) / block_dims.x, (ne1 + block_dims.y - 1) / block_dims.y,
                         (ne2 * ne3 + block_dims.z - 1) / block_dims.z);
 
+        const uint3 ne10 = init_fastdiv_values((uint32_t) cne1[0]);
+        const uint3 ne11 = init_fastdiv_values((uint32_t) cne1[1]);
+        const uint3 ne12 = init_fastdiv_values((uint32_t) cne1[2]);
+        const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);
+
         if (block_nums.z > 65535) {
-            int block_num = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+            int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+            const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
+            const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));
+            const uint3 ne0_fastdiv = init_fastdiv_values((uint32_t) ne0);
+            const uint3 ne1_fastdiv = init_fastdiv_values((uint32_t) ne1);
+            const uint3 ne2_fastdiv = init_fastdiv_values((uint32_t) ne2);
+
             if constexpr (sizeof...(I) > 0) {
-                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
-                    <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd,
-                        ne0, ne1, ne2, ne3,
-                        ne10, ne11, ne12, ne13,
-                        /* s0, */ s1, s2, s3,
-                        /* s00,*/ s01, s02, s03,
-                        /* s10,*/ s11, s12,s13,
-                        (const src1_t *) dst->src[I + 1]->data...);
+                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
+                    ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00,*/ s01, s02, s03,
+                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
                 k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
-                    <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd,
-                        ne0, ne1, ne2, ne3,
-                        ne10, ne11, ne12, ne13,
-                        /* s0, */ s1, s2, s3,
-                        /* s00,*/ s01, s02, s03,
-                        /* s10,*/ s11, s12,s13);
+                    <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
+                                                           ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
+                                                           /* s0, */ s1, s2, s3,
+                                                           /* s00,*/ s01, s02, s03,
+                                                           /* s10,*/ s11, s12, s13);
             }
         } else {
+            const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
             if constexpr (sizeof...(I) > 0) {
-                k_bin_bcast<bin_op, src0_t, src1_t, dst_t>
-                    <<<block_nums, block_dims, 0, stream>>>(src0_dd, src1_dd, dst_dd,
-                        ne0, ne1, ne2, ne3,
-                        ne10, ne11, ne12, ne13,
-                        /* s0, */ s1, s2, s3,
-                        /* s00,*/ s01, s02, s03,
-                        /* s10,*/ s11, s12,s13,
-                        (const src1_t *) dst->src[I + 1]->data...);
+                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00,*/ s01, s02, s03,
+                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
             } else {
-                k_bin_bcast<bin_op, src0_t, src1_t, dst_t>
-                    <<<block_nums, block_dims, 0, stream>>>(src0_dd, src1_dd, dst_dd,
-                        ne0, ne1, ne2, ne3,
-                        ne10, ne11, ne12, ne13,
-                        /* s0, */ s1, s2, s3,
-                        /* s00,*/ s01, s02, s03,
-                        /* s10,*/ s11, s12,s13);
+                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00,*/ s01, s02, s03,
+                    /* s10,*/ s11, s12, s13);
             }
         }
     }

ggml/src/ggml-cuda/common.cuh

@@ -545,6 +545,45 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP)
 }
 
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float v, const float u) {
+    acc += v*u;
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v, const float2 u) {
+    acc += v.x*u.x;
+    acc += v.y*u.y;
+}
+
+static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
+#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+    asm volatile("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(acc) : "v"(v), "v"(u));
+#else
+#ifdef FAST_FP16_AVAILABLE
+    const float2 tmp = __half22float2(v*u);
+    acc += tmp.x + tmp.y;
+#else
+    const float2 tmpv = __half22float2(v);
+    const float2 tmpu = __half22float2(u);
+    acc += tmpv.x * tmpu.x;
+    acc += tmpv.y * tmpu.y;
+#endif // FAST_FP16_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
+}
+
+// Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
+template <int nbytes>
+static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
+    if constexpr (nbytes == 4) {
+        *(int *) dst = *(const int *) src;
+    } else if constexpr (nbytes == 8) {
+        *(int2 *) dst = *(const int2 *) src;
+    } else if constexpr (nbytes == 16) {
+        *(int4 *) dst = *(const int4 *) src;
+    } else {
+        static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
+    }
+}
+
 static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 #if CUDART_VERSION >= 12080
     const nv_bfloat16 e = __nv_cvt_e8m0_to_bf16raw(x);
@@ -570,6 +609,8 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 //
 // n/d = (mulhi(n, mp) + n) >> L;
 static const uint3 init_fastdiv_values(uint32_t d) {
+    GGML_ASSERT(d != 0);
+
     // compute L = ceil(log2(d));
     uint32_t L = 0;
     while (L < 32 && (uint32_t{ 1 } << L) < d) {

ggml/src/ggml-cuda/convert.cuh

@@ -38,6 +38,8 @@ template<typename dst_t, typename src_t>
         return __float2bfloat16(float(x));
     } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
         return __bfloat162float(x);
+    } else if constexpr(std::is_same_v<dst_t, int32_t>) {
+        return int32_t(x);
     } else {
         return float(x);
     }

ggml/src/ggml-cuda/cpy.cu

@@ -374,6 +374,10 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
+        ggml_cpy_flt_cuda<float, int32_t> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_flt_cuda<int32_t, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

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

@@ -1,371 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-#include "fattn-tile-f16.cuh"
-
-#define FATTN_KQ_STRIDE_TILE_F16 64
-
-template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !defined(GGML_USE_HIP)
-__launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !defined(GGML_USE_HIP)
-static __global__ void flash_attn_tile_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
-
-    // Skip unused kernel variants for faster compilation:
-#ifdef FP16_MMA_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FP16_MMA_AVAILABLE
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        NO_DEVICE_CODE;
-        return;
-    }
-
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2   = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2  * K_h2   = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2  * V_h2   = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half   * maskh  = (const half   *) (mask  + nb33*(sequence % ne33)                          + nb31*ic0);
-    const float  * sinksf = (const float  *) (sinks);
-
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-    const half  slopeh = __float2half(slopef);
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-
-    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
-    half2 * KQ2 = (half2 *) KQ;
-
-    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
-
-    half kqmax[ncols/nwarps];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        kqmax[j0/nwarps] = -HALF_MAX_HALF;
-    }
-    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
-
-    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
-
-    // Convert Q to half2 and store in registers:
-    __shared__ half2 Q_h2[ncols][D/2];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        const int j = j0 + threadIdx.y;
-
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-            const int i = i0 + threadIdx.x;
-
-            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
-            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
-        }
-    }
-
-    __syncthreads();
-
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        half kqmax_new[ncols/nwarps];
-#pragma unroll
-        for (int j = 0; j < ncols/nwarps; ++j) {
-            kqmax_new[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
-                const int k_KQ = k_KQ_0 + threadIdx.x;
-
-                KV_tmp[i_KQ][k_KQ] = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-            }
-        }
-
-        __syncthreads();
-
-        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
-
-#pragma unroll
-        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
-            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
-            half2 Q_k[ncols/nwarps];
-
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
-                const int i_KQ = i_KQ_0 + threadIdx.x;
-
-                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
-            }
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                const int j_KQ = j_KQ_0 + threadIdx.y;
-
-                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
-            }
-
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
-#pragma unroll
-                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
-                }
-            }
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
-            const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                const int j_KQ = j_KQ_0 + threadIdx.y;
-
-                half sum;
-                if (use_logit_softcap) {
-                    const float2 tmp = __half22float2(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
-                    sum = logit_softcap * tanhf(tmp.x + tmp.y);
-                } else {
-                    sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
-                }
-                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
-
-                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
-
-                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            const int j = j0 + threadIdx.y;
-
-            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
-            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
-            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
-
-#pragma unroll
-            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
-                const half2 val = h2exp(diff);
-                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
-                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
-            const int k = k0 + threadIdx.y;
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                KV_tmp[k][i] = V_h2[int64_t(k_VKQ_0 + k)*stride_KV2 + i];
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
-            half2  V_k[(D/2)/WARP_SIZE][2];
-            half2 KQ_k[ncols/nwarps];
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
-                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
-            }
-#pragma unroll
-            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-                const int j = j0 + threadIdx.y;
-
-                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-#pragma unroll
-                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
-                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
-                }
-            }
-        }
-
-        __syncthreads();
-    }
-
-    //Attention sink: adjust running max and sum once per head
-    if (sinksf && blockIdx.y == 0) {
-        const half sink = __float2half(sinksf[head]);
-
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            half kqmax_new_j = fmaxf(kqmax[j0/nwarps], sink);
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new_j));
-            kqmax[j0/nwarps] = kqmax_new_j;
-
-            const half val = hexp(sink - kqmax[j0/nwarps]);
-            kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
-            if (threadIdx.x == 0) {
-                kqsum[j0/nwarps].x = __hadd(__low2half(kqsum[j0/nwarps]), val);
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
-            }
-        }
-    }
-
-    float2 * dst2 = (float2 *) dst;
-
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
-        const int j_VKQ = j_VKQ_0 + threadIdx.y;
-
-        if (ic0 + j_VKQ >= ne01) {
-            return;
-        }
-
-        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
-        kqsum_j = warp_reduce_sum((float)kqsum_j);
-
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
-
-#pragma unroll
-        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
-            const int i0 = i00 + threadIdx.x;
-
-            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
-            if (gridDim.y == 1) {
-                dst_val /= __half2half2(kqsum_j);
-            }
-            dst2[j_dst_unrolled*(D/2) + i0] = __half22float2(dst_val);
-        }
-
-        if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
-        }
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
-}
-
-template <int cols_per_block, bool use_logit_softcap>
-void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    switch (Q->ne[0]) {
-        case  64: {
-            constexpr int    D             = 64;
-            constexpr int    nwarps        = 8;
-            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
-        } break;
-        case 128: {
-            constexpr int    D             = 128;
-            constexpr int    nwarps        = 8;
-            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
-        } break;
-        default: {
-            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
-        } break;
-    }
-}
-
-void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
-
-    const int32_t precision = KQV->op_params[3];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    if (Q->ne[1] <= 16) {
-        constexpr int cols_per_block = 16;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 32;
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-    }
-}

ggml/src/ggml-cuda/fattn-tile-f16.cuh

@@ -1,3 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -1,379 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-#include "fattn-tile-f32.cuh"
-
-#define FATTN_KQ_STRIDE_TILE_F32 32
-
-template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
-#if !defined(GGML_USE_HIP)
-__launch_bounds__(nwarps*WARP_SIZE, 2)
-#endif // !defined(GGML_USE_HIP)
-static __global__ void flash_attn_tile_ext_f32(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#ifdef FLASH_ATTN_AVAILABLE
-
-    // Skip unused kernel variants for faster compilation:
-#ifdef FP16_MMA_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FP16_MMA_AVAILABLE
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-            max_bias, m0, m1, n_head_log2, logit_softcap,
-            ne00, ne01, ne02, ne03,
-                  nb01, nb02, nb03,
-            ne10, ne11, ne12, ne13,
-                  nb11, nb12, nb13,
-                  nb21, nb22, nb23,
-                  ne31, ne32, ne33,
-                  nb31, nb32, nb33);
-        NO_DEVICE_CODE;
-        return;
-    }
-
-    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2   = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2  * K_h2   = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2  * V_h2   = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half   * maskh  = (const half   *) (mask  + nb33*(sequence % ne33)                          + nb31*ic0);
-    const float  * sinksf = (const float  *) (sinks);
-
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-
-    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
-
-    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
-    float2 * KV_tmp2 = (float2 *) KV_tmp;
-
-    float kqmax[ncols/nwarps];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
-    }
-    float kqsum[ncols/nwarps] = {0.0f};
-
-    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
-
-    // Convert Q to half2 and store in registers:
-    __shared__ float Q_f[ncols][D];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        const int j = j0 + threadIdx.y;
-
-#pragma unroll
-        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
-            float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
-            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
-            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
-        }
-    }
-
-    __syncthreads();
-
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        float kqmax_new[ncols/nwarps];
-#pragma unroll
-        for (int j = 0; j < ncols/nwarps; ++j) {
-            kqmax_new[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
-                const half2 tmp = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
-                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
-                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
-            }
-        }
-
-        __syncthreads();
-
-        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
-
-#pragma unroll
-        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
-            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
-            float Q_k[ncols/nwarps];
-
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
-                const int i_KQ = i_KQ_0 + threadIdx.x;
-
-                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
-            }
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                const int j_KQ = j_KQ_0 + threadIdx.y;
-
-                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
-            }
-
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
-#pragma unroll
-                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
-                }
-            }
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
-            const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
-                const int j_KQ = j_KQ_0 + threadIdx.y;
-
-                if (use_logit_softcap) {
-                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
-                }
-
-                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
-
-                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
-
-                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            const int j = j0 + threadIdx.y;
-
-            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
-            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
-            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
-
-            float kqsum_add = 0.0f;
-#pragma unroll
-            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
-                const float val = expf(diff);
-                kqsum_add += val;
-                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
-            }
-            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
-                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
-            const int k = k0 + threadIdx.y;
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                const half2 tmp = V_h2[int64_t(k_VKQ_0 + k)*stride_KV2 + i];
-                KV_tmp2[k*(D/2) + i].x =  __low2float(tmp);
-                KV_tmp2[k*(D/2) + i].y = __high2float(tmp);
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
-            float2 V_k[(D/2)/WARP_SIZE];
-            float  KQ_k[ncols/nwarps];
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
-            }
-#pragma unroll
-            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-                const int j = j0 + threadIdx.y;
-
-                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-#pragma unroll
-                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
-                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
-                }
-            }
-        }
-
-        __syncthreads();
-    }
-
-
-    //Attention sink: adjust running max and sum once per head
-    if (sinksf && blockIdx.y == 0) {
-        const float sink = sinksf[head];
-
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            float kqmax_new_j = fmaxf(kqmax[j0/nwarps], sink);
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new_j);
-            kqmax[j0/nwarps] = kqmax_new_j;
-
-            const float val = expf(sink - kqmax[j0/nwarps]);
-            kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
-            if (threadIdx.x == 0) {
-                kqsum[j0/nwarps] += val;
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
-                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
-            }
-        }
-    }
-
-    float2 * dst2 = (float2 *) dst;
-
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
-        const int j_VKQ = j_VKQ_0 + threadIdx.y;
-
-        if (ic0 + j_VKQ >= ne01) {
-            return;
-        }
-
-        float kqsum_j = kqsum[j_VKQ_0/nwarps];
-        kqsum_j = warp_reduce_sum(kqsum_j);
-
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
-
-#pragma unroll
-        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
-            const int i0 = i00 + threadIdx.x;
-
-            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
-            if (gridDim.y == 1) {
-                dst_val.x /= kqsum_j;
-                dst_val.y /= kqsum_j;
-            }
-            dst2[j_dst_unrolled*(D/2) + i0] = dst_val;
-        }
-
-        if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
-        }
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // FLASH_ATTN_AVAILABLE
-}
-
-template <int cols_per_block, bool use_logit_softcap>
-void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    switch (Q->ne[0]) {
-        case  64: {
-            constexpr int    D             = 64;
-            constexpr int    nwarps        = 8;
-            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
-        } break;
-        case 128: {
-            constexpr int    D             = 128;
-            constexpr int    nwarps        = 8;
-            constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
-        } break;
-        default: {
-            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
-        } break;
-    }
-}
-
-void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q = dst->src[0];
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    if (Q->ne[1] <= 16) {
-        constexpr int cols_per_block = 16;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 32;
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
-    }
-}

ggml/src/ggml-cuda/fattn-tile-f32.cuh

@@ -1,3 +0,0 @@
-#include "common.cuh"
-
-void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/fattn-tile.cu

@@ -0,0 +1,660 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile.cuh"
+
+#define FATTN_TILE_NTHREADS 256
+
+static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int cc, const int warp_size) {
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+        switch (D) {
+            case 64:
+                return 64;
+            case 128:
+            case 256:
+                if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
+                    return ncols <= 16 ? 64 : 32;
+                } else {
+                    return 64;
+                }
+            default:
+                GGML_ABORT("fatal error");
+                return -1;
+        }
+    }
+    if (fast_fp16_available(cc)) {
+        switch (D) {
+            case 64:
+            case 128:
+                return 128;
+            case 256:
+                return ncols <= 16 ? 128 : 64;
+            default:
+                GGML_ABORT("fatal error");
+                return -1;
+        }
+    }
+    switch (D) {
+        case 64:
+            return ncols <= 16 ? 128 : 64;
+        case 128:
+            return ncols <= 16 ? 64 : 32;
+        case 256:
+            return 32;
+        default:
+            GGML_ABORT("fatal error");
+            return -1;
+    }
+    GGML_UNUSED(warp_size);
+}
+
+static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols, int warp_size) {
+#ifdef GGML_USE_HIP
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+#if defined(GCN) || defined(CDNA)
+            return ncols <= 16 ? 64 : 32;
+#else
+            return 64;
+#endif // defined(GCN) || defined(CDNA)
+        case 256:
+#if defined(GCN) || defined(CDNA)
+            return ncols <= 16 ? 64 : 32;
+#else
+            return 64;
+#endif // defined(GCN) || defined(CDNA)
+        default:
+            return -1;
+    }
+#else
+#ifdef FAST_FP16_AVAILABLE
+    switch (D) {
+        case 64:
+        case 128:
+            return 128;
+        case 256:
+            return ncols <= 16 ? 128 : 64;
+        default:
+            return -1;
+    }
+#else
+    switch (D) {
+        case 64:
+            return ncols <= 16 ? 128 : 64;
+        case 128:
+            return ncols <= 16 ? 64 : 32;
+        case 256:
+            return 32;
+        default:
+            return -1;
+    }
+#endif // FAST_FP16_AVAILABLE
+#endif // GGML_USE_HIP
+    GGML_UNUSED_VARS(ncols, warp_size);
+}
+
+static constexpr __device__ int fattn_tile_get_kq_nbatch_device(int D, int ncols, int warp_size) {
+#ifdef GGML_USE_HIP
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+#if defined(GCN) || defined(CDNA)
+            return ncols <= 16 ? 64 : 128;
+#else
+            return 64;
+#endif // defined(GCN) || defined(CDNA)
+        case 256:
+#if defined(GCN) || defined(CDNA)
+            return ncols <= 16 ? 64 : 128;
+#else
+            return ncols <= 16 ? 64 : 256;
+#endif // defined(GCN) || defined(CDNA)
+        default:
+            return -1;
+    }
+#else
+#ifdef FAST_FP16_AVAILABLE
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+            return ncols <= 16 ? 128 : 64;
+        case 256:
+            return ncols <= 16 ? 64 : 128;
+        default:
+            return -1;
+    }
+#else
+    switch (D) {
+        case 64:
+            return 64;
+        case 128:
+            return 128;
+        case 256:
+            return ncols <= 16 ? 128 : 64;
+        default:
+            return -1;
+    }
+#endif // FAST_FP16_AVAILABLE
+#endif // GGML_USE_HIP
+    GGML_UNUSED_VARS(ncols, warp_size);
+}
+
+template<int D, int ncols, bool use_logit_softcap> // D == head size
+#ifdef GGML_USE_HIP
+__launch_bounds__(FATTN_TILE_NTHREADS, 1)
+#else
+__launch_bounds__(FATTN_TILE_NTHREADS, 2)
+#endif // GGML_USE_HIP
+static __global__ void flash_attn_tile(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        const char * __restrict__ sinks,
+        const int  * __restrict__ KV_max,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+#ifdef FLASH_ATTN_AVAILABLE
+
+    // Skip unused kernel variants for faster compilation:
+#ifdef FP16_MMA_AVAILABLE
+    NO_DEVICE_CODE;
+    return;
+#endif // FP16_MMA_AVAILABLE
+
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int warp_size = 32;
+    constexpr int nwarps    = FATTN_TILE_NTHREADS / warp_size;
+    constexpr int kq_stride = fattn_tile_get_kq_stride_device(D, ncols, warp_size);
+    static_assert(kq_stride % warp_size == 0, "kq_stride not divisable by warp_size.");
+    constexpr int kq_nbatch = fattn_tile_get_kq_nbatch_device(D, ncols, warp_size);
+    static_assert(kq_nbatch % (2*warp_size) == 0, "bad kq_nbatch");
+
+    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
+
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2   = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2   = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2   = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh  = (const half   *) (mask  + nb33*(sequence % ne33)                          + nb31*ic0);
+    const float  * sinksf = (const float  *) (sinks);
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
+
+#if defined(GGML_USE_HIP)
+    constexpr int cpy_nb = 16;
+#else
+    constexpr int cpy_nb = 8;
+#endif // defined(GGML_USE_HIP) && defined(GCN)
+    constexpr int cpy_ne = cpy_nb / 4;
+
+    __shared__ float KQ[ncols][kq_stride];
+#ifdef FAST_FP16_AVAILABLE
+    __shared__ half2 Q_tmp[ncols][D/2];
+    __shared__ half2 KV_tmp_h2[kq_stride * (kq_nbatch/2 + cpy_ne)]; // Padded to avoid memory bank conflicts.
+    half2 VKQ[ncols/nwarps][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
+#else
+    __shared__ float Q_tmp[ncols][D];
+    __shared__ float KV_tmp_f[kq_stride * (kq_nbatch + cpy_ne)]; // Padded to avoid memory bank conflicts.
+    float2 * KV_tmp_f2 = (float2 *) KV_tmp_f;
+    float2 VKQ[ncols/nwarps][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
+#endif // FAST_FP16_AVAILABLE
+
+
+    float kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols/nwarps] = {0.0f};
+
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0 + threadIdx.x] : make_float2(0.0f, 0.0f);
+#ifdef FAST_FP16_AVAILABLE
+            Q_tmp[j][i0 + threadIdx.x] = make_half2(tmp.x * scale, tmp.y * scale);
+#else
+            Q_tmp[j][2*i0             + threadIdx.x] = tmp.x * scale;
+            Q_tmp[j][2*i0 + warp_size + threadIdx.x] = tmp.y * scale;
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    __syncthreads();
+
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    for (int k_VKQ_0 = blockIdx.y*kq_stride; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*kq_stride) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+        float sum[kq_stride/warp_size][ncols/nwarps] = {{0.0f}};
+
+#pragma unroll
+        for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += kq_nbatch) {
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += nwarps) {
+                const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += warp_size) {
+                    const half2 tmp_h2 = K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1 + threadIdx.x];
+#ifdef FAST_FP16_AVAILABLE
+                    KV_tmp_h2[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1 + threadIdx.x] = tmp_h2;
+#else
+                    const float2 tmp_f2 = __half22float2(tmp_h2);
+                    KV_tmp_f[i_KQ*(kq_nbatch + cpy_ne) + 2*k_KQ_1             + threadIdx.x] = tmp_f2.x;
+                    KV_tmp_f[i_KQ*(kq_nbatch + cpy_ne) + 2*k_KQ_1 + warp_size + threadIdx.x] = tmp_f2.y;
+#endif // FAST_FP16_AVAILABLE
+                }
+            }
+
+            __syncthreads();
+
+#ifdef FAST_FP16_AVAILABLE
+#pragma unroll
+            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += cpy_ne) {
+                half2 K_k[kq_stride/warp_size][cpy_ne];
+                half2 Q_k[ncols/nwarps][cpy_ne];
+#else
+#pragma unroll
+            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += cpy_ne) {
+                float K_k[kq_stride/warp_size][cpy_ne];
+                float Q_k[ncols/nwarps][cpy_ne];
+#endif // FAST_FP16_AVAILABLE
+
+#pragma unroll
+                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+                    const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#ifdef FAST_FP16_AVAILABLE
+                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp_h2[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1]);
+#else
+                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp_f [i_KQ*(kq_nbatch   + cpy_ne) + k_KQ_1]);
+#endif // FAST_FP16_AVAILABLE
+                }
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    const int j_KQ = j_KQ_0 + threadIdx.y;
+
+#ifdef FAST_FP16_AVAILABLE
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0/nwarps], &Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1]);
+#else
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0/nwarps], &Q_tmp[j_KQ][k_KQ_0   + k_KQ_1]);
+#endif // FAST_FP16_AVAILABLE
+                }
+
+#pragma unroll
+                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+#pragma unroll
+                    for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+#pragma unroll
+                        for (int k = 0; k < cpy_ne; ++k) {
+                            ggml_cuda_mad(sum[i_KQ_0/warp_size][j_KQ_0/nwarps], K_k[i_KQ_0/warp_size][k], Q_k[j_KQ_0/nwarps][k]);
+                        }
+                    }
+                }
+            }
+
+            if (k_KQ_0 + kq_nbatch < D) {
+                __syncthreads(); // Sync not needed on last iteration.
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                if (use_logit_softcap) {
+                    sum[i_KQ_0/warp_size][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/warp_size][j_KQ_0/nwarps]);
+                }
+
+                sum[i_KQ_0/warp_size][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/warp_size][j_KQ_0/nwarps]);
+
+                KQ[j_KQ][i_KQ] = sum[i_KQ_0/warp_size][j_KQ_0/nwarps];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max<warp_size>(kqmax_new[j0/nwarps]);
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+            float kqsum_add = 0.0f;
+            if (kq_stride % (4*warp_size) == 0 && cpy_ne % 4 == 0) {
+#pragma unroll
+                for (int i0 = 0; i0 < kq_stride; i0 += 4*warp_size) {
+                    const int i = i0 + 4*threadIdx.x;
+
+                    float4 val = *(const float4 *) &KQ[j][i];
+                    val.x = expf(val.x - kqmax[j0/nwarps]);
+                    val.y = expf(val.y - kqmax[j0/nwarps]);
+                    val.z = expf(val.z - kqmax[j0/nwarps]);
+                    val.w = expf(val.w - kqmax[j0/nwarps]);
+                    kqsum_add += val.x + val.y + val.z + val.w;
+
+#ifdef FAST_FP16_AVAILABLE
+                    const half2 tmp[2] = {make_half2(val.x, val.y), make_half2(val.z, val.w)};
+                    ggml_cuda_memcpy_1<sizeof(tmp)>(&KQ[j][i/2], &tmp);
+#else
+                    ggml_cuda_memcpy_1<sizeof(val)>(&KQ[j][i], &val);
+#endif // FAST_FP16_AVAILABLE
+                }
+            } else if (kq_stride % (2*warp_size) == 0 && cpy_ne % 2 == 0) {
+#pragma unroll
+                for (int i0 = 0; i0 < kq_stride; i0 += 2*warp_size) {
+                    const int i = i0 + 2*threadIdx.x;
+
+                    float2 val = *(const float2 *) &KQ[j][i];
+                    val.x = expf(val.x - kqmax[j0/nwarps]);
+                    val.y = expf(val.y - kqmax[j0/nwarps]);
+                    kqsum_add += val.x + val.y;
+#ifdef FAST_FP16_AVAILABLE
+                    const half2 tmp = make_half2(val.x, val.y);
+                    ggml_cuda_memcpy_1<sizeof(tmp)>(&KQ[j][i/2], &tmp);
+#else
+                    ggml_cuda_memcpy_1<sizeof(val)>(&KQ[j][i], &val);
+#endif // FAST_FP16_AVAILABLE
+                }
+            } else {
+                for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
+                    const int i = i0 + threadIdx.x;
+
+                    const float diff = KQ[j][i] - kqmax[j0/nwarps];
+                    const float val = expf(diff);
+                    kqsum_add += val;
+#ifdef FAST_FP16_AVAILABLE
+                    ((half *) KQ[j])[i] = val;
+#else
+                    KQ[j][i] = val;
+#endif // FAST_FP16_AVAILABLE
+                }
+            }
+            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
+
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0/nwarps][i0/warp_size] *= KQ_max_scale_h2;
+            }
+#else
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0/nwarps][i0/warp_size].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/warp_size].y *= KQ_max_scale;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+
+        constexpr int V_cols_per_iter = kq_stride*kq_nbatch / D;
+        static_assert(kq_stride % V_cols_per_iter == 0, "bad V_cols_per_iter");
+#pragma unroll
+        for (int k0 = 0; k0 < kq_stride; k0 += V_cols_per_iter) {
+#pragma unroll
+            for (int k1 = 0; k1 < V_cols_per_iter; k1 += nwarps) {
+                const int k_tile = k1 + threadIdx.y;
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    const int i = i0 + threadIdx.x;
+
+                    const half2 tmp = V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i];
+#ifdef FAST_FP16_AVAILABLE
+                    KV_tmp_h2[k_tile*(D/2) + i] = tmp;
+#else
+                    KV_tmp_f2[k_tile*(D/2) + i] = __half22float2(tmp);
+#endif // FAST_FP16_AVAILABLE
+                }
+            }
+
+            __syncthreads();
+
+#pragma unroll
+            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
+#ifdef FAST_FP16_AVAILABLE
+                half2 V_k[(D/2)/warp_size];
+                half2 KQ_k[ncols/nwarps];
+#else
+                float2 V_k[(D/2)/warp_size];
+                float  KQ_k[ncols/nwarps];
+#endif // FAST_FP16_AVAILABLE
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                    const int i = i0 + threadIdx.x;
+
+#ifdef FAST_FP16_AVAILABLE
+                    V_k[i0/warp_size] = KV_tmp_h2[k1*(D/2) + i];
+#else
+                    V_k[i0/warp_size] = KV_tmp_f2[k1*(D/2) + i];
+#endif // FAST_FP16_AVAILABLE
+                }
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    const int j = j0 + threadIdx.y;
+
+#ifdef FAST_FP16_AVAILABLE
+                    KQ_k[j0/nwarps] = __half2half2(((const half *)KQ[j])[k0 + k1]);
+#else
+                    KQ_k[j0/nwarps] = KQ[j][k0 + k1];
+#endif // FAST_FP16_AVAILABLE
+                }
+
+#pragma unroll
+                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+#pragma unroll
+                    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+#ifdef FAST_FP16_AVAILABLE
+                        VKQ[j0/nwarps][i0/warp_size]   += V_k[i0/warp_size]  *KQ_k[j0/nwarps];
+#else
+                        VKQ[j0/nwarps][i0/warp_size].x += V_k[i0/warp_size].x*KQ_k[j0/nwarps];
+                        VKQ[j0/nwarps][i0/warp_size].y += V_k[i0/warp_size].y*KQ_k[j0/nwarps];
+#endif // FAST_FP16_AVAILABLE
+                    }
+                }
+            }
+
+            __syncthreads();
+        }
+    }
+
+
+    // Attention sink: adjust running max and sum once per head
+    if (sinksf && blockIdx.y == 0) {
+        const float sink = sinksf[head];
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            float kqmax_new_j = fmaxf(kqmax[j0/nwarps], sink);
+            kqmax_new_j = warp_reduce_max<warp_size>(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new_j);
+            kqmax[j0/nwarps] = kqmax_new_j;
+
+            const float val = expf(sink - kqmax[j0/nwarps]);
+            kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
+            if (threadIdx.x == 0) {
+                kqsum[j0/nwarps] += val;
+            }
+
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0/nwarps][i0/warp_size] *= KQ_max_scale_h2;
+            }
+#else
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
+                VKQ[j0/nwarps][i0/warp_size].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/warp_size].y *= KQ_max_scale;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    float2 * dst2 = (float2 *) dst;
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        float kqsum_j = kqsum[j_VKQ_0/nwarps];
+        kqsum_j = warp_reduce_sum<warp_size>(kqsum_j);
+
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
+
+#pragma unroll
+        for (int i00 = 0; i00 < D/2; i00 += warp_size) {
+            const int i0 = i00 + threadIdx.x;
+
+#ifdef FAST_FP16_AVAILABLE
+            float2 dst_val = __half22float2(VKQ[j_VKQ_0/nwarps][i0/warp_size]);
+#else
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/warp_size];
+#endif // FAST_FP16_AVAILABLE
+
+            if (gridDim.y == 1) {
+                dst_val.x /= kqsum_j;
+                dst_val.y /= kqsum_j;
+            }
+            dst2[j_dst_unrolled*(D/2) + i0] = dst_val;
+        }
+
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
+    NO_DEVICE_CODE;
+#endif // FLASH_ATTN_AVAILABLE
+}
+
+template <int D, bool use_logit_softcap>
+static void launch_fattn_tile_switch_ncols(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    const int id                 = ggml_cuda_get_device();
+    const int cc                 = ggml_cuda_info().devices[id].cc;
+    const int warp_size          = 32;
+    const int nwarps             = FATTN_TILE_NTHREADS / warp_size;
+
+    constexpr size_t nbytes_shared = 0;
+
+    if (Q->ne[1] > 16) {
+        constexpr int cols_per_block = 32;
+        fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
+        const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
+        launch_fattn<D, cols_per_block, 1>
+            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
+        return;
+    }
+
+    constexpr int cols_per_block = 16;
+    fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
+    const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
+    launch_fattn<D, cols_per_block, 1>
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
+}
+
+template <bool use_logit_softcap>
+static void launch_fattn_tile_switch_head_size(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            launch_fattn_tile_switch_ncols< 64, use_logit_softcap>(ctx, dst);
+        } break;
+        case 128: {
+            launch_fattn_tile_switch_ncols<128, use_logit_softcap>(ctx, dst);
+        } break;
+        case 256: {
+            launch_fattn_tile_switch_ncols<256, use_logit_softcap>(ctx, dst);
+        } break;
+        default: {
+            GGML_ABORT("Unsupported head size");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
+    }
+}

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/fattn.cu

@@ -1,8 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 #include "fattn-mma-f16.cuh"
-#include "fattn-tile-f16.cuh"
-#include "fattn-tile-f32.cuh"
+#include "fattn-tile.cuh"
 #include "fattn-vec-f16.cuh"
 #include "fattn-vec-f32.cuh"
 #include "fattn-wmma-f16.cuh"
@@ -271,8 +270,7 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
 // Best FlashAttention kernel for a specific GPU:
 enum best_fattn_kernel {
     BEST_FATTN_KERNEL_NONE     =   0,
-    BEST_FATTN_KERNEL_TILE_F32 = 200,
-    BEST_FATTN_KERNEL_TILE_F16 = 210,
+    BEST_FATTN_KERNEL_TILE     = 200,
     BEST_FATTN_KERNEL_VEC_F32  = 100,
     BEST_FATTN_KERNEL_VEC_F16  = 110,
     BEST_FATTN_KERNEL_WMMA_F16 = 300,
@@ -411,10 +409,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
-    if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
-        return BEST_FATTN_KERNEL_TILE_F16;
-    }
-    return BEST_FATTN_KERNEL_TILE_F32;
+    return BEST_FATTN_KERNEL_TILE;
 }
 
 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -422,11 +417,8 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     switch (ggml_cuda_get_best_fattn_kernel(ggml_cuda_get_device(), dst)) {
         case BEST_FATTN_KERNEL_NONE:
             GGML_ABORT("fatal error");
-        case BEST_FATTN_KERNEL_TILE_F32:
-            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
-            break;
-        case BEST_FATTN_KERNEL_TILE_F16:
-            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+        case BEST_FATTN_KERNEL_TILE:
+            ggml_cuda_flash_attn_ext_tile(ctx, dst);
             break;
         case BEST_FATTN_KERNEL_VEC_F32:
             ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);

ggml/src/ggml-cuda/getrows.cu

@@ -2,39 +2,39 @@
 #include "dequantize.cuh"
 #include "convert.cuh"
 
-#define MAX_GRIDDIM_Y 65535
-
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
         const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
         const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
-        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const int64_t ne10,*/ const int64_t ne11, const int64_t ne12, /*const int64_t ne13,*/
         /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    for (int64_t i00 = 2*(blockIdx.y*blockDim.x + threadIdx.x); i00 < ne00; i00 += gridDim.y*blockDim.x) {
-        // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
-        const int i10 =  blockIdx.x;
-        const int i11 =  blockIdx.z / ne12;
-        const int i12 =  blockIdx.z % ne12;
+    for (int64_t z = blockIdx.z; z < ne11*ne12; z += gridDim.z) {
+        for (int64_t i00 = 2*(blockIdx.y*blockDim.x + threadIdx.x); i00 < ne00; i00 += gridDim.y*blockDim.x) {
+            // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+            const int i10 =  blockIdx.x;
+            const int i11 =  z / ne12; // TODO fastdiv
+            const int i12 =  z % ne12;
 
-        const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+            const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
 
-        dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-        const void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;
+            dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+            const void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;
 
-        const int ib   =  i00/qk;      // block index
-        const int iqs  = (i00%qk)/qr;  // quant index
-        const int iybs = i00 - i00%qk; // dst block start index
-        const int y_offset = qr == 1 ? 1 : qk/2;
+            const int ib   =  i00/qk;      // block index
+            const int iqs  = (i00%qk)/qr;  // quant index
+            const int iybs = i00 - i00%qk; // dst block start index
+            const int y_offset = qr == 1 ? 1 : qk/2;
 
-        // dequantize
-        float2 v;
-        dequantize_kernel(src0_row, ib, iqs, v);
+            // dequantize
+            float2 v;
+            dequantize_kernel(src0_row, ib, iqs, v);
 
-        dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);
-        dst_row[iybs + iqs + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+            dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);
+            dst_row[iybs + iqs + y_offset] = ggml_cuda_cast<dst_t>(v.y);
+        }
     }
 }
 
@@ -42,27 +42,29 @@ template<typename src0_t, typename dst_t>
 static __global__ void k_get_rows_float(
         const src0_t * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
         const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
-        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
+        /*const int64_t ne10,*/ const int64_t ne11, const int64_t ne12, /*const int64_t ne13,*/
         /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
         /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
 
-    for (int64_t i00 = blockIdx.y*blockDim.x + threadIdx.x; i00 < ne00; i00 += gridDim.y*blockDim.x) {
-        // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
-        const int i10 = blockIdx.x;
-        const int i11 = blockIdx.z / ne12;
-        const int i12 = blockIdx.z % ne12;
+    for (int64_t z = blockIdx.z; z < ne11*ne12; z += gridDim.z) {
+        for (int64_t i00 = blockIdx.y*blockDim.x + threadIdx.x; i00 < ne00; i00 += gridDim.y*blockDim.x) {
+            // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+            const int i10 = blockIdx.x;
+            const int i11 = z / ne12; // TODO fastdiv
+            const int i12 = z % ne12;
 
-        if (i00 >= ne00) {
-            return;
+            if (i00 >= ne00) {
+                return;
+            }
+
+            const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+            dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+            const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+            dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
         }
-
-        const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-
-        dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-        const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
-
-        dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
     }
 }
 
@@ -98,7 +100,7 @@ static void get_rows_cuda_q(
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(ne10, MIN(block_num_y, MAX_GRIDDIM_Y), ne11*ne12);
+    const dim3 block_nums(ne10, MIN(block_num_y, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);
@@ -116,7 +118,7 @@ static void get_rows_cuda_q(
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
         src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
-        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /*ne10,*/ ne11, ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
@@ -131,7 +133,7 @@ static void get_rows_cuda_float(
         cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
-    const dim3 block_nums(ne10, MIN(block_num_y, MAX_GRIDDIM_Y), ne11*ne12);
+    const dim3 block_nums(ne10, MIN(block_num_y, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);
@@ -147,7 +149,7 @@ static void get_rows_cuda_float(
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
         src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
-        /*ne10, ne11,*/ ne12, /*ne13,*/
+        /*ne10,*/ ne11, ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);

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

@@ -2109,6 +2109,11 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
             return;
         }
+
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2])) {
+            ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
+            return;
+        }
     }
 
     cudaStream_t stream = ctx.stream();
@@ -3135,6 +3140,7 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
     /* .event_record            = */ ggml_backend_cuda_event_record,
     /* .event_wait              = */ ggml_backend_cuda_event_wait,
+    /* .optimize_graph          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -3204,6 +3210,7 @@ struct ggml_backend_cuda_device_context {
     int device;
     std::string name;
     std::string description;
+    std::string pci_bus_id;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -3228,9 +3235,12 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
 }
 
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+
     props->name        = ggml_backend_cuda_device_get_name(dev);
     props->description = ggml_backend_cuda_device_get_description(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
+    props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
     bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
@@ -3461,6 +3471,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_I32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_I32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3574,9 +3590,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
+        case GGML_OP_PAD:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
-        case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3792,6 +3808,10 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;
 
+                char pci_bus_id[16] = {};
+                snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
+                dev_ctx->pci_bus_id = pci_bus_id;
+
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface   = */ ggml_backend_cuda_device_interface,
                     /* .reg     = */ &reg,

ggml/src/ggml-cuda/im2col.cu

@@ -122,11 +122,14 @@ static  __global__ void im2col_3d_kernel(
         int64_t OH_OW, int64_t KD_KH_KW, int64_t ID_IH_IW, int64_t KH_KW, int64_t IH_IW, int64_t IC_ID_IH_IW,
         int64_t IC_KD_KH_KW, int64_t OW_KD_KH_KW, int64_t OD_OH_OW_IC_KD_KH_KW, int64_t OH_OW_IC_KD_KH_KW,
         int64_t OW_IC_KD_KH_KW, int64_t N_OD_OH, int64_t OD_OH,
+        int64_t stride_q, int64_t stride_z, int64_t stride_y, int64_t stride_x,
         int s0, int s1, int s2, int p0, int p1, int p2, int d0, int d1, int d2) {
     const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
     if (i >= IC_KD_KH_KW) {
         return;
     }
+    GGML_UNUSED(N); GGML_UNUSED(OC); GGML_UNUSED(OH_OW); GGML_UNUSED(OD); GGML_UNUSED(OW); GGML_UNUSED(KD); GGML_UNUSED(KH);
+    GGML_UNUSED(ID_IH_IW); GGML_UNUSED(IH_IW); GGML_UNUSED(IC_ID_IH_IW); GGML_UNUSED(OW_KD_KH_KW);
 
     const int64_t iic = i / KD_KH_KW;
     const int64_t ikd = (i - iic * KD_KH_KW) / KH_KW;
@@ -148,7 +151,7 @@ static  __global__ void im2col_3d_kernel(
         if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
             dst[offset_dst] = 0.0f;
         } else {
-            const int64_t offset_src = in*IC_ID_IH_IW + iic*ID_IH_IW + iid*IH_IW + iih*IW + iiw;
+            const int64_t offset_src = ((in * IC + iic) * stride_q) + (iid * stride_z) + (iih * stride_y) + (iiw * stride_x);
             dst[offset_dst] = src[offset_src];
         }
     }
@@ -159,6 +162,7 @@ template <typename T>
 static void im2col_3d_cuda(const float * src, T* dst,
     int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW, int64_t OC,
     int64_t KD, int64_t KH, int64_t KW, int64_t OD, int64_t OH, int64_t OW,
+    int64_t stride_q, int64_t stride_z, int64_t stride_y, int64_t stride_x,
     int s0, int s1, int s2, int p0, int p1, int p2, int d0, int d1, int d2, cudaStream_t stream) {
     const int64_t OH_OW = OH*OW;
     const int64_t KD_KH_KW = KD*KH*KW;
@@ -179,23 +183,30 @@ static void im2col_3d_cuda(const float * src, T* dst,
                                                                                            OH_OW, KD_KH_KW, ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW,
                                                                                            IC_KD_KH_KW, OW_KD_KH_KW, OD_OH_OW_IC_KD_KH_KW,
                                                                                            OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,
+                                                                                           stride_q, stride_z, stride_y, stride_x,
                                                                                            s0, s1, s2, p0, p1, p2, d0, d1, d2);
 }
 
 static void im2col_3d_cuda_f16(const float * src, half * dst,
     int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW, int64_t OC,
     int64_t KD, int64_t KH, int64_t KW, int64_t OD, int64_t OH, int64_t OW,
+    int64_t stride_q, int64_t stride_z, int64_t stride_y, int64_t stride_x,
     int s0, int s1, int s2, int p0, int p1, int p2, int d0, int d1, int d2, cudaStream_t stream) {
 
-    im2col_3d_cuda<half>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+    im2col_3d_cuda<half>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                         stride_q, stride_z, stride_y, stride_x,
+                         s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
 }
 
 static void im2col_3d_cuda_f32(const float * src, float * dst,
     int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW, int64_t OC,
     int64_t KD, int64_t KH, int64_t KW, int64_t OD, int64_t OH, int64_t OW,
+    int64_t stride_q, int64_t stride_z, int64_t stride_y, int64_t stride_x,
     int s0, int s1, int s2, int p0, int p1, int p2, int d0, int d1, int d2, cudaStream_t stream) {
 
-    im2col_3d_cuda<float>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+    im2col_3d_cuda<float>(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                          stride_q, stride_z, stride_y, stride_x,
+                          s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
 }
 
 void ggml_cuda_op_im2col_3d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -235,9 +246,19 @@ void ggml_cuda_op_im2col_3d(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
     const int64_t OH = ne2;
     const int64_t OW = ne1;
 
+    const size_t  es       = ggml_element_size(src1);
+    const int64_t stride_x = src1->nb[0] / es;
+    const int64_t stride_y = src1->nb[1] / es;
+    const int64_t stride_z = src1->nb[2] / es;
+    const int64_t stride_q = src1->nb[3] / es;
+
     if(dst->type == GGML_TYPE_F16) {
-        im2col_3d_cuda_f16(src1_d, (half *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+        im2col_3d_cuda_f16(src1_d, (half *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                           stride_q, stride_z, stride_y, stride_x,
+                           s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
     } else {
-        im2col_3d_cuda_f32(src1_d, (float *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
+        im2col_3d_cuda_f32(src1_d, (float *) dst_d, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW,
+                           stride_q, stride_z, stride_y, stride_x,
+                           s0, s1, s2, p0, p1, p2, d0, d1, d2, stream);
     }
 }

ggml/src/ggml-cuda/mma.cuh

@@ -1,3 +1,4 @@
+#pragma once
 // This file contains primitives that expose the tensor core PTX instructions for CUDA code.
 // The primitives can be used in a similar way as the nvcuda::wmma interface but with a well-defined memory layout.
 // The documentation for the PTX instructions can be found under:

ggml/src/ggml-cuda/mmf.cu

@@ -1,343 +1,12 @@
 #include "ggml.h"
-#include "common.cuh"
-#include "mma.cuh"
 #include "mmf.cuh"
 
-using namespace ggml_cuda_mma;
-
-#define MMF_ROWS_PER_BLOCK 32
-
-template <typename T, int rows_per_block, int cols_per_block, int nwarps>
-__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
-static __global__ void mul_mat_f(
-        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
-        const int ncols, const int nchannels_y, const int stride_row, const int stride_col_y, const int stride_col_dst,
-        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
-
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    constexpr int tile_k_padded = warp_size + 4;
-    constexpr int ntA = rows_per_block / tile_A::I;
-    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
-
-    const int row0        = blockIdx.x * rows_per_block;
-    const int channel_dst = blockIdx.y;
-    const int channel_x   = channel_dst / channel_ratio;
-    const int channel_y   = channel_dst;
-    const int sample_dst  = blockIdx.z;
-    const int sample_x    = sample_dst / sample_ratio;
-    const int sample_y    = sample_dst;
-
-    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row0*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;
-
-    const float2 * y2 = (const float2 *) y;
-
-    extern __shared__ char data_mmv[];
-
-    tile_C C[ntA][ntB];
-
-    T * tile_xy = (T *) data_mmv + threadIdx.y*(tile_A::I * tile_k_padded);
-
-    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
-        tile_A A[ntA][warp_size / tile_A::J];
-#pragma unroll
-        for (int itA = 0; itA < ntA; ++itA) {
-#pragma unroll
-            for (int i = 0; i < tile_A::I; ++i) {
-                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
-            }
-#pragma unroll
-            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
-                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
-            }
-        }
-
-#pragma unroll
-        for (int itB = 0; itB < ntB; ++itB) {
-            if constexpr (std::is_same_v<T, float>) {
-#pragma unroll
-                for (int j0 = 0; j0 < tile_B::I; ++j0) {
-                    const int j = j0 + itB*tile_B::I;
-
-                    tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[j*stride_col_y + col] : 0.0f;
-                }
-            } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
-#pragma unroll
-                for (int j0 = 0; j0 < tile_B::I; ++j0) {
-                    const int j = j0 + itB*tile_B::I;
-
-                    const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
-                    tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
-                }
-            } else {
-                static_assert(std::is_same_v<T, void>, "unsupported type");
-            }
-#pragma unroll
-            for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
-                tile_B B;
-                load_ldmatrix(B, tile_xy + k0, tile_k_padded);
-#pragma unroll
-                for (int itA = 0; itA < ntA; ++itA) {
-                    mma(C[itA][itB], A[itA][k0/tile_B::J], B);
-                }
-            }
-        }
-    }
-
-    float * buf_iw = (float *) data_mmv;
-    constexpr int kiw = nwarps*rows_per_block + 4;
-
-    if (nwarps > 1) {
-        __syncthreads();
-    }
-#pragma unroll
-    for (int itB = 0; itB < ntB; ++itB) {
-#pragma unroll
-        for (int itA = 0; itA < ntA; ++itA) {
-#pragma unroll
-            for (int l = 0; l < tile_C::ne; ++l) {
-                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
-                const int j = itB*tile_C::J + tile_C::get_j(l);
-                buf_iw[j*kiw + i] = C[itA][itB].x[l];
-            }
-        }
-    }
-
-    if (nwarps > 1) {
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
-        const int j = j0 + threadIdx.y;
-
-        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
-            return;
-        }
-
-        float sum = 0.0f;
-        static_assert(rows_per_block == warp_size, "need loop/check");
-#pragma unroll
-        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
-            const int i = i0 + threadIdx.x;
-
-            sum += buf_iw[j*kiw + i];
-        }
-        dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
-    }
-#else
-    GGML_UNUSED_VARS(x, y, ids, 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);
-    NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-}
-
-template <typename T, int cols_per_block>
-static void mul_mat_f_cuda(
-        const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols_x, const int64_t nrows_x,
-        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_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) {
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-
-    GGML_ASSERT(!ids && "mul_mat_id not implemented");
-
-    GGML_ASSERT(ncols_x      % 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 int64_t channel_ratio = nchannels_dst / nchannels_x;
-    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
-
-    const int device = ggml_cuda_get_device();
-    const int warp_size = ggml_cuda_info().devices[device].warp_size;
-
-    int64_t nwarps_best     = 1;
-    int64_t niter_best      = (ncols_x + warp_size*2 - 1) / (warp_size*2);
-    int64_t max_block_size  = 256;
-    for (int64_t nwarps = 2; nwarps <= max_block_size/warp_size; nwarps++) {
-        const int64_t niter = (ncols_x + nwarps*warp_size*2 - 1) / (nwarps*warp_size*2);
-        if (niter < niter_best) {
-            niter_best  = niter;
-            nwarps_best = nwarps;
-        }
-    }
-
-    constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
-    const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
-    const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
-    const dim3 block_nums(nrows_x/rows_per_block, nchannels_dst, nsamples_dst);
-    const dim3 block_dims(warp_size, nwarps_best, 1);
-    switch (nwarps_best) {
-        case 1: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 1><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 2: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 2><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 3: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 3><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 4: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 4><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 5: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 5><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 6: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 6><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 7: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 7><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        case 8: {
-            mul_mat_f<T, rows_per_block, cols_per_block, 8><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, dst, ncols_x, 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);
-        } break;
-        default: {
-            GGML_ABORT("fatal error");
-        } break;
-    }
-}
-
-template <typename T>
-static void mul_mat_f_switch_cols_per_block(
-        const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
-        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_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) {
-    switch (ncols_dst) {
-        case  1: {
-            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  2: {
-            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  3: {
-            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  4: {
-            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  5: {
-            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  6: {
-            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  7: {
-            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  8: {
-            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case  9: {
-            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 10: {
-            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 11: {
-            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 12: {
-            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 13: {
-            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 14: {
-            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 15: {
-            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        case 16: {
-            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, 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);
-        } break;
-        default: {
-            GGML_ABORT("fatal error");
-        } break;
-    }
-}
-
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
     GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
     GGML_ASSERT(         dst->type == GGML_TYPE_F32);
 
+
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     const size_t ts_src0 = ggml_type_size(src0->type);
@@ -365,55 +34,72 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     const int64_t s13 = src1->nb[3] / ts_src1;
     const int64_t s3  =  dst->nb[3] / ts_dst;
 
+    const int64_t ids_s0 = ids ? ids->nb[0] / ggml_type_size(ids->type) : 0;
+    const int64_t ids_s1 = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
+
     // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
     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_channel_dst = ids ? s1   : s2;
-    const int64_t stride_channel_y   = ids ? s11  : s12;
+    const int64_t nchannels_dst      = ids ? ne1 : ne2;
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
+    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;
+
+    int64_t stride_channel_y         = ids ? s11  : s12;
+    int64_t nchannels_y              = ids ? ne11 : ne12;
+
+    //mul_mat_id: handle broadcast
+    if (ids && nchannels_y == 1) {
+        stride_channel_y = 0;
+        nchannels_y      = ids->ne[0];
+    }
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
             constexpr int vals_per_T = 1;
             mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
-                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+                ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half2 * src0_d = (const half2 *) src0->data;
             constexpr int vals_per_T = 2;
             mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
-                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+                ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
             constexpr int vals_per_T = 2;
             mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, s11/vals_per_T, s1,
-                ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03/vals_per_T, s13,              s3,                 ctx.stream());
+                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+                ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
+                ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream());
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
     }
 }
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, int64_t ne11) {
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols) {
+
+    if (ggml_is_quantized(type)) {
+        return false;
+    }
+
     if (src0_ne[0] % (warp_size * (4/ggml_type_size(type))) != 0) {
         return false;
     }
     if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
         return false;
     }
-    if (ne11 > 16) {
+    if (src1_ncols > 16) {
         return false;
     }
+
     switch (type) {
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);

ggml/src/ggml-cuda/mmf.cuh

@@ -1,5 +1,461 @@
+#pragma once
+
+#include "mma.cuh"
 #include "common.cuh"
 
+using namespace ggml_cuda_mma;
+
+#define MMF_ROWS_PER_BLOCK 32
+
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, int64_t ne11);
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols);
+
+template <typename T, int rows_per_block, int cols_per_block, int nwarps, bool has_ids>
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
+static __global__ void mul_mat_f(
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int stride_col_id, const int stride_row_id,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+    typedef tile<16, 8, float> tile_C;
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int ntA = rows_per_block / tile_A::I;
+    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
+
+    const int row0        = blockIdx.x * rows_per_block;
+
+    const int expert_idx  = has_ids ? blockIdx.y : 0;
+    const int channel_dst = has_ids ? 0 : blockIdx.y;
+
+    const int channel_x   = has_ids ? expert_idx : (channel_dst / channel_ratio);
+    const int channel_y   = channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x  + row0*stride_row ;
+    y   += int64_t(sample_y)  *stride_sample_y   + (has_ids ? 0 : channel_y  *stride_channel_y);
+    dst += int64_t(sample_dst)*stride_sample_dst + (has_ids ? 0 : channel_dst*stride_channel_dst);
+
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+
+    char * shmem_base = data_mmv;
+    int  * slot_map   = (int *) shmem_base;
+    char * compute_base = has_ids ? (shmem_base + GGML_PAD(cols_per_block, 16) * sizeof(int)) : shmem_base;
+
+    tile_C C[ntA][ntB];
+
+    T * tile_xy = (T *) compute_base + threadIdx.y*(tile_A::I * tile_k_padded);
+
+    if constexpr (has_ids) {
+        int found = 0;
+
+        for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+            const int32_t * __restrict__ id_row = ids + j*stride_row_id;
+
+            if (threadIdx.x == 0) {
+                slot_map[j] = -1;
+            }
+
+            for (int k = threadIdx.x; k < nchannels_dst; k += warp_size) {
+                int match = id_row[k*stride_col_id] == expert_idx;
+
+                if (match) {
+                    slot_map[j] = k;
+                    found = 1;
+                    break;
+                }
+            }
+        }
+
+        if (!__syncthreads_or(found)) {
+            return;
+        }
+    }
+
+
+    for (int col = threadIdx.y*warp_size + threadIdx.x; col < ncols; col += nwarps*warp_size) {
+        tile_A A[ntA][warp_size / tile_A::J];
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int i = 0; i < tile_A::I; ++i) {
+                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
+                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
+            }
+        }
+
+#pragma unroll
+        for (int itB = 0; itB < ntB; ++itB) {
+            if constexpr (std::is_same_v<T, float>) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + itB*tile_B::I;
+
+                    if constexpr (!has_ids) {
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[j*stride_col_y + col] : 0.0f;
+                    } else {
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[slot_map[j]*stride_channel_y + j*stride_col_y + col] : 0.0f;
+                    }
+                }
+            } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
+#pragma unroll
+                for (int j0 = 0; j0 < tile_B::I; ++j0) {
+                    const int j = j0 + itB*tile_B::I;
+
+                    if constexpr (!has_ids) {
+                        const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                    } else {
+                        float2 tmp = j < cols_per_block && slot_map[j] >= 0 ? *(const float2*) &y[slot_map[j]*stride_channel_y + 2*(j*stride_col_y + col)] : make_float2(0.0f, 0.0f);
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
+                    }
+                }
+            } else {
+                static_assert(std::is_same_v<T, void>, "unsupported type");
+            }
+#pragma unroll
+            for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
+                tile_B B;
+                load_ldmatrix(B, tile_xy + k0, tile_k_padded);
+#pragma unroll
+                for (int itA = 0; itA < ntA; ++itA) {
+                    mma(C[itA][itB], A[itA][k0/tile_B::J], B);
+                }
+            }
+        }
+    }
+
+    float * buf_iw = (float *) compute_base;
+    constexpr int kiw = nwarps*rows_per_block + 4;
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+#pragma unroll
+    for (int itB = 0; itB < ntB; ++itB) {
+#pragma unroll
+        for (int itA = 0; itA < ntA; ++itA) {
+#pragma unroll
+            for (int l = 0; l < tile_C::ne; ++l) {
+                const int i = threadIdx.y*rows_per_block + itA*tile_C::I + tile_C::get_i(l);
+                const int j = itB*tile_C::J + tile_C::get_j(l);
+                buf_iw[j*kiw + i] = C[itA][itB].x[l];
+            }
+        }
+    }
+
+    if (nwarps > 1) {
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j0 + nwarps > cols_per_block && j >= cols_per_block) {
+            return;
+        }
+
+        float sum = 0.0f;
+        static_assert(rows_per_block == warp_size, "need loop/check");
+#pragma unroll
+        for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
+            const int i = i0 + threadIdx.x;
+
+            sum += buf_iw[j*kiw + i];
+        }
+
+        if constexpr (!has_ids) {
+            dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
+        } else {
+            const int slot = (j < cols_per_block) ? slot_map[j] : -1;
+            if (slot >= 0) {
+                dst[slot*stride_channel_dst + j*stride_col_dst + row0 + threadIdx.x] = sum;
+            }
+        }
+    }
+#else
+    GGML_UNUSED_VARS(x, y, ids, dst,
+        ncols, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        stride_col_id, stride_row_id,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+    NO_DEVICE_CODE;
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
+template<typename T, int cols_per_block, int nwarps>
+static inline void mul_mat_f_switch_ids(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nchannels_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t stride_col_id, const int64_t stride_row_id,
+        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
+    if (ids) {
+        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums, block_dims, nbytes_shared_total, stream>>>
+             (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+             stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+             sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+    } else {
+        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, false><<<block_nums, block_dims, nbytes_shared_total, stream>>>
+            (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+             stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+             sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+    }
+}
+
+template <typename T, int cols_per_block>
+void mul_mat_f_cuda(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t stride_col_id, const int64_t stride_row_id,
+        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) {
+    typedef tile<16, 8, T>     tile_A;
+    typedef tile< 8, 8, T>     tile_B;
+
+    GGML_ASSERT(ncols_x      % 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 int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
+
+    const int device = ggml_cuda_get_device();
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
+
+    int64_t nwarps_best     = 1;
+    int64_t niter_best      = (ncols_x + warp_size*2 - 1) / (warp_size*2);
+    int64_t max_block_size  = 256;
+    for (int64_t nwarps = 2; nwarps <= max_block_size/warp_size; nwarps++) {
+        const int64_t niter = (ncols_x + nwarps*warp_size*2 - 1) / (nwarps*warp_size*2);
+        if (niter < niter_best) {
+            niter_best  = niter;
+            nwarps_best = nwarps;
+        }
+    }
+
+    constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
+    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
+    const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
+    const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
+    const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
+    const int64_t grid_y = ids ? nchannels_x : nchannels_dst; // per expert when ids present
+
+    const dim3 block_nums(nrows_x/rows_per_block, grid_y, nsamples_dst);
+    const dim3 block_dims(warp_size, nwarps_best, 1);
+
+    switch (nwarps_best) {
+        case 1: {
+            mul_mat_f_switch_ids<T, cols_per_block, 1>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 2: {
+            mul_mat_f_switch_ids<T, cols_per_block, 2>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 3: {
+            mul_mat_f_switch_ids<T, cols_per_block, 3>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 4: {
+            mul_mat_f_switch_ids<T, cols_per_block, 4>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 5: {
+            mul_mat_f_switch_ids<T, cols_per_block, 5>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 6: {
+            mul_mat_f_switch_ids<T, cols_per_block, 6>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 7: {
+            mul_mat_f_switch_ids<T, cols_per_block, 7>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        case 8: {
+            mul_mat_f_switch_ids<T, cols_per_block, 8>(
+                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+
+    GGML_UNUSED_VARS(nchannels_y);
+}
+
+template <typename T>
+static void mul_mat_f_switch_cols_per_block(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t stride_col_id, const int stride_row_id,
+        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) {
+    switch (ncols_dst) {
+        case  1: {
+            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  2: {
+            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  3: {
+            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  4: {
+            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  5: {
+            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  6: {
+            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  7: {
+            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  8: {
+            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case  9: {
+            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 10: {
+            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 11: {
+            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 12: {
+            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 13: {
+            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 14: {
+            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 15: {
+            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        case 16: {
+            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, 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);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+#define DECL_MMF_CASE_HELPER(T, ncols_dst) \
+    template void mul_mat_f_cuda<T, ncols_dst>( \
+        const T * x, const float * y, const int32_t * ids, float * dst, \
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, \
+        const int64_t stride_col_id, const int64_t stride_row_id, \
+        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);
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#define DECL_MMF_CASE_EXTERN(ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(float, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(half2, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst)
+
+#define DECL_MMF_CASE(ncols_dst) \
+    DECL_MMF_CASE_HELPER(float, ncols_dst) \
+    DECL_MMF_CASE_HELPER(half2, ncols_dst) \
+    DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst)
+
+DECL_MMF_CASE_EXTERN(1);
+DECL_MMF_CASE_EXTERN(2);
+DECL_MMF_CASE_EXTERN(3);
+DECL_MMF_CASE_EXTERN(4);
+DECL_MMF_CASE_EXTERN(5);
+DECL_MMF_CASE_EXTERN(6);
+DECL_MMF_CASE_EXTERN(7);
+DECL_MMF_CASE_EXTERN(8);
+DECL_MMF_CASE_EXTERN(9);
+DECL_MMF_CASE_EXTERN(10);
+DECL_MMF_CASE_EXTERN(11);
+DECL_MMF_CASE_EXTERN(12);
+DECL_MMF_CASE_EXTERN(13);
+DECL_MMF_CASE_EXTERN(14);
+DECL_MMF_CASE_EXTERN(15);
+DECL_MMF_CASE_EXTERN(16);
+#else
+#define DECL_MMF_CASE(ncols_dst)
+#endif

ggml/src/ggml-cuda/mmvq.cu

@@ -141,9 +141,10 @@ template <ggml_type type, int ncols_dst>
 __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, float * __restrict__ dst,
-        const int ncols_x, const int nchannels_y, const int stride_row_x, const int stride_col_y, const int stride_col_dst,
-        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_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) {
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
@@ -161,12 +162,12 @@ static __global__ void mul_mat_vec_q(
     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 int channel_dst = blockIdx.y;
-    const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
-    const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
-    const int sample_dst  = blockIdx.z;
-    const int sample_x    = sample_dst / sample_ratio;
-    const int sample_y    = sample_dst;
+    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;
+    const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
+    const uint32_t sample_y    = sample_dst;
 
     // partial sum for each thread
     float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
@@ -247,8 +248,9 @@ static void mul_mat_vec_q_switch_ncols_dst(
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
     GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
 
-    const int channel_ratio = nchannels_dst / nchannels_x;
-    const int sample_ratio  = nsamples_dst  / nsamples_x;
+    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);
 
     const int device = ggml_cuda_get_device();
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
@@ -256,86 +258,70 @@ static void mul_mat_vec_q_switch_ncols_dst(
 
     GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
-        case 1:
-        {
+        case 1: {
             constexpr int c_ncols_dst = 1;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 2:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 2: {
             constexpr int c_ncols_dst = 2;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 3:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 3: {
             constexpr int c_ncols_dst = 3;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 4:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 4: {
             constexpr int c_ncols_dst = 4;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 5:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 5: {
             constexpr int c_ncols_dst = 5;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 6:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 6: {
             constexpr int c_ncols_dst = 6;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 7:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 7: {
             constexpr int c_ncols_dst = 7;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
-        case 8:
-        {
+                (vx, vy, ids, 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);
+        } break;
+        case 8: {
             constexpr int c_ncols_dst = 8;
             std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
-                (vx, vy, ids, 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);
-            break;
-        }
+                (vx, vy, ids, 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);
+        } break;
         default:
             GGML_ABORT("fatal error");
             break;

ggml/src/ggml-cuda/quantize.cu

@@ -1,26 +1,27 @@
 #include "quantize.cuh"
 #include <cstdint>
 
+__launch_bounds__(CUDA_QUANTIZE_BLOCK_SIZE, 1)
 static __global__ void quantize_q8_1(
         const float * __restrict__ x, void * __restrict__ vy,
         const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t ne0, const int ne1, const int ne2) {
+        const int64_t ne0, const uint32_t ne1, const uint3 ne2) {
     const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i0 >= ne0) {
         return;
     }
 
+    const int64_t i3 = fastdiv(blockIdx.z, ne2);
+    const int64_t i2 = blockIdx.z - i3*ne2.z;
     const int64_t i1 = blockIdx.y;
-    const int64_t i2 = blockIdx.z % ne2;
-    const int64_t i3 = blockIdx.z / ne2;
 
     const int64_t & i00 = i0;
     const int64_t & i01 = i1;
     const int64_t & i02 = i2;
     const int64_t & i03 = i3;
 
-    const int64_t i_cont = ((i3*ne2 + i2) * ne1 + i1) * ne0 + i0;
+    const int64_t i_cont = ((i3*ne2.z + i2) * ne1 + i1) * ne0 + i0;
 
     block_q8_1 * y = (block_q8_1 *) vy;
 
@@ -31,10 +32,10 @@ static __global__ void quantize_q8_1(
     float amax = fabsf(xi);
     float sum = xi;
 
-    amax = warp_reduce_max(amax);
-    sum  = warp_reduce_sum(sum);
+    amax = warp_reduce_max<QK8_1>(amax);
+    sum  = warp_reduce_sum<QK8_1>(sum);
 
-    const float  d = amax / 127;
+    const float  d = amax / 127.0f;
     const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
 
     y[ib].qs[iqs] = q;
@@ -43,8 +44,7 @@ static __global__ void quantize_q8_1(
         return;
     }
 
-    reinterpret_cast<half&>(y[ib].ds.x) = d;
-    reinterpret_cast<half&>(y[ib].ds.y) = sum;
+    y[ib].ds = make_half2(d, sum);
 }
 
 template <mmq_q8_1_ds_layout ds_layout>
@@ -152,10 +152,12 @@ void quantize_row_q8_1_cuda(
     GGML_ASSERT(!ids);
     GGML_ASSERT(ne0 % QK8_1 == 0);
 
+    const uint3 ne2_fastdiv = init_fastdiv_values(ne2);
+
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
     const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2_fastdiv);
     GGML_UNUSED(type_src0);
 }
 

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -24,7 +24,7 @@ TYPES_MMQ = [
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
     "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
     "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",
-    "GGML_TYPE_IQ1_S", "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS"
+    "GGML_TYPE_IQ1_S", "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS", "GGML_TYPE_MXFP4"
 ]
 
 SOURCE_MMQ = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
@@ -34,6 +34,13 @@ SOURCE_MMQ = """// This file has been autogenerated by generate_cu_files.py, do
 DECL_MMQ_CASE({type});
 """
 
+SOURCE_MMF = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE({type});
+"""
+
 
 def get_short_name(long_quant_name):
     return long_quant_name.replace("GGML_TYPE_", "").lower()
@@ -76,3 +83,7 @@ for ncols in [8, 16, 32, 64]:
 for type in TYPES_MMQ:
     with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f:
         f.write(SOURCE_MMQ.format(type=type))
+
+for type in range(1, 17):
+    with open(f"mmf-instance-ncols_{type}.cu", "w") as f:
+        f.write(SOURCE_MMF.format(type=type))

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(1);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(10);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(11);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(12);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(13);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(14);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(15);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(16);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(2);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(3);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(4);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(5);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(6);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(7);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(8);

ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmf.cuh"
+
+DECL_MMF_CASE(9);

ggml/src/ggml-cuda/tsembd.cu

@@ -7,11 +7,11 @@ static __global__ void timestep_embedding_f32(const float * timesteps, float * d
     int j = threadIdx.x + blockIdx.x * blockDim.x;
     float * embed_data = (float *)((char *)dst +  i*nb1);
 
-    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
-        embed_data[dim] = 0.f;
+    int half = dim / 2;
+    if (dim % 2 != 0 && j == half) {
+        embed_data[2 * half] = 0.f;
     }
 
-    int half = dim / 2;
     if (j >= half) {
         return;
     }

ggml/src/ggml-cuda/vendors/hip.h

@@ -162,6 +162,14 @@
 #define GCN
 #endif
 
+#if defined(__gfx900__) || defined(__gfx906__)
+#define GCN5
+#endif
+
+#if defined(__gfx803__)
+#define GCN4
+#endif
+
 #if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
 #define CDNA // For the entire family
 #endif

ggml/src/ggml-metal/CMakeLists.txt

@@ -6,6 +6,7 @@ message(STATUS "Metal framework found")
 
 ggml_add_backend_library(ggml-metal
                          ggml-metal.m
+                         ggml-metal-common.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE

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

@@ -0,0 +1,458 @@
+#include "ggml-metal-common.h"
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include <vector>
+
+// represents a memory range (i.e. an interval from a starting address p0 to an ending address p1 in a given buffer pb)
+// the type indicates whether it is a source range (i.e. ops read data from it) or a destination range (i.e. ops write data to it)
+struct ggml_mem_range {
+    uint64_t pb; // buffer id
+
+    uint64_t p0; // begin
+    uint64_t p1; // end
+
+    ggml_mem_range_type pt;
+};
+
+struct ggml_mem_ranges {
+    std::vector<ggml_mem_range> ranges;
+
+    int debug = 0;
+};
+
+struct ggml_mem_ranges * ggml_mem_ranges_init(int debug) {
+    auto * res = new ggml_mem_ranges;
+
+    res->ranges.reserve(256);
+    res->debug = debug;
+
+    return res;
+}
+
+void ggml_mem_ranges_free(ggml_mem_ranges * mrs) {
+    delete mrs;
+}
+
+void ggml_mem_ranges_reset(ggml_mem_ranges * mrs) {
+    mrs->ranges.clear();
+}
+
+static bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, ggml_mem_range mr) {
+    mrs->ranges.push_back(mr);
+
+    return true;
+}
+
+static ggml_mem_range ggml_mem_range_from_tensor(const ggml_tensor * tensor, ggml_mem_range_type pt) {
+    // always use the base tensor
+    tensor = tensor->view_src ? tensor->view_src : tensor;
+
+    GGML_ASSERT(!tensor->view_src);
+
+    ggml_mem_range mr;
+
+    if (tensor->buffer) {
+        // when the tensor is allocated, use the actual memory address range in the buffer
+        //
+        // take the actual allocated size with ggml_backend_buft_get_alloc_size()
+        // this can be larger than the tensor size if the buffer type allocates extra memory
+        // ref: https://github.com/ggml-org/llama.cpp/pull/15966
+        mr = {
+            /*.pb =*/ (uint64_t) tensor->buffer,
+            /*.p0 =*/ (uint64_t) tensor->data,
+            /*.p1 =*/ (uint64_t) tensor->data + ggml_backend_buft_get_alloc_size(tensor->buffer->buft, tensor),
+            /*.pt =*/ pt,
+        };
+    } else {
+        // otherwise, the pointer address is used as an unique id of the memory ranges
+        //   that the tensor will be using when it is allocated
+        mr = {
+            /*.pb =*/ (uint64_t) tensor,
+            /*.p0 =*/ 0,    //
+            /*.p1 =*/ 1024, // [0, 1024) is a dummy range, not used
+            /*.pt =*/ pt,
+        };
+    };
+
+    return mr;
+}
+
+static ggml_mem_range ggml_mem_range_from_tensor_src(const ggml_tensor * tensor) {
+    return ggml_mem_range_from_tensor(tensor, MEM_RANGE_TYPE_SRC);
+}
+
+static ggml_mem_range ggml_mem_range_from_tensor_dst(const ggml_tensor * tensor) {
+    return ggml_mem_range_from_tensor(tensor, MEM_RANGE_TYPE_DST);
+}
+
+static bool ggml_mem_ranges_add_src(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
+
+    ggml_mem_range mr = ggml_mem_range_from_tensor_src(tensor);
+
+    if (mrs->debug > 2) {
+        GGML_LOG_DEBUG("%s: add src range buf=%lld, [%lld, %lld)\n", __func__, mr.pb, mr.p0, mr.p1);
+    }
+
+    return ggml_mem_ranges_add(mrs, mr);
+}
+
+static bool ggml_mem_ranges_add_dst(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
+
+    ggml_mem_range mr = ggml_mem_range_from_tensor_dst(tensor);
+
+    if (mrs->debug > 2) {
+        GGML_LOG_DEBUG("%s: add dst range buf=%lld, [%lld, %lld)\n", __func__, mr.pb, mr.p0, mr.p1);
+    }
+
+    return ggml_mem_ranges_add(mrs, mr);
+}
+
+bool ggml_mem_ranges_add(ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (tensor->src[i]) {
+            ggml_mem_ranges_add_src(mrs, tensor->src[i]);
+        }
+    }
+
+    return ggml_mem_ranges_add_dst(mrs, tensor);
+}
+
+static bool ggml_mem_ranges_check(const ggml_mem_ranges * mrs, ggml_mem_range mr) {
+    for (size_t i = 0; i < mrs->ranges.size(); i++) {
+        const auto & cmp = mrs->ranges[i];
+
+        // two memory ranges cannot intersect if they are in different buffers
+        if (mr.pb != cmp.pb) {
+            continue;
+        }
+
+        // intersecting source ranges are allowed
+        if (mr.pt == MEM_RANGE_TYPE_SRC && cmp.pt == MEM_RANGE_TYPE_SRC) {
+            continue;
+        }
+
+        if (mr.p0 < cmp.p1 && mr.p1 >= cmp.p0) {
+            if (mrs->debug > 2) {
+                GGML_LOG_DEBUG("%s: the %s range buf=%lld, [%lld, %lld) overlaps with a previous %s range buf=%lld, [%lld, %lld)\n",
+                        __func__,
+                        mr.pt == MEM_RANGE_TYPE_SRC ? "src" : "dst",
+                        mr.pb, mr.p0, mr.p1,
+                        cmp.pt == MEM_RANGE_TYPE_SRC ? "src" : "dst",
+                        cmp.pb, cmp.p0, cmp.p1);
+            }
+
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool ggml_mem_ranges_check_src(const ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
+
+    ggml_mem_range mr = ggml_mem_range_from_tensor_src(tensor);
+
+    const bool res = ggml_mem_ranges_check(mrs, mr);
+
+    return res;
+}
+
+static bool ggml_mem_ranges_check_dst(const ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
+
+    ggml_mem_range mr = ggml_mem_range_from_tensor_dst(tensor);
+
+    const bool res = ggml_mem_ranges_check(mrs, mr);
+
+    return res;
+}
+
+bool ggml_mem_ranges_check(const ggml_mem_ranges * mrs, const ggml_tensor * tensor) {
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (tensor->src[i]) {
+            if (!ggml_mem_ranges_check_src(mrs, tensor->src[i])) {
+                return false;
+            }
+        }
+    }
+
+    return ggml_mem_ranges_check_dst(mrs, tensor);
+}
+
+// TODO: move to ggml.h?
+static bool is_empty(ggml_op op) {
+    switch (op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return true;
+        default:
+            return false;
+    }
+}
+
+struct node_info {
+    ggml_tensor * node;
+
+    std::vector<ggml_tensor *> fused;
+
+    ggml_op op() const {
+        return node->op;
+    }
+
+    const ggml_tensor * dst() const {
+        return fused.empty() ? node : fused.back();
+    }
+
+    bool is_empty() const {
+        return ::is_empty(node->op);
+    }
+
+    void add_fused(ggml_tensor * t) {
+        fused.push_back(t);
+    }
+};
+
+static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node_info> & nodes) {
+    // helper to add node src and dst ranges
+    const auto & h_add = [](ggml_mem_ranges * mrs, const node_info & node) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (node.node->src[i]) {
+                if (!ggml_mem_ranges_add_src(mrs, node.node->src[i])) {
+                    return false;
+                }
+            }
+        }
+
+        // keep track of the sources of the fused nodes as well
+        for (const auto * fused : node.fused) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (fused->src[i]) {
+                    if (!ggml_mem_ranges_add_src(mrs, fused->src[i])) {
+                        return false;
+                    }
+                }
+            }
+        }
+
+        return ggml_mem_ranges_add_dst(mrs, node.dst());
+    };
+
+    // helper to check if a node can run concurrently with the existing set of nodes
+    const auto & h_check = [](const ggml_mem_ranges * mrs, const node_info & node) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (node.node->src[i]) {
+                if (!ggml_mem_ranges_check_src(mrs, node.node->src[i])) {
+                    return false;
+                }
+            }
+        }
+
+        for (const auto * fused : node.fused) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (fused->src[i]) {
+                    if (!ggml_mem_ranges_check_src(mrs, fused->src[i])) {
+                        return false;
+                    }
+                }
+            }
+        }
+
+        return ggml_mem_ranges_check_dst(mrs, node.dst());
+    };
+
+    // perform reorders only across these types of ops
+    // can be expanded when needed
+    // IMPORTANT: do not add ops such as GGML_OP_CPY or GGML_OP_SET_ROWS
+    //            the dependencies from such ops are not always represented in the graph
+    const auto & h_safe = [](ggml_op op) {
+        switch (op) {
+            case GGML_OP_MUL_MAT:
+            case GGML_OP_MUL_MAT_ID:
+            case GGML_OP_ROPE:
+            case GGML_OP_NORM:
+            case GGML_OP_RMS_NORM:
+            case GGML_OP_GROUP_NORM:
+            case GGML_OP_SUM_ROWS:
+            case GGML_OP_MUL:
+            case GGML_OP_ADD:
+            case GGML_OP_DIV:
+            case GGML_OP_GLU:
+            case GGML_OP_SCALE:
+            case GGML_OP_GET_ROWS:
+                return true;
+            default:
+                return is_empty(op);
+        }
+    };
+
+    const int n = nodes.size();
+
+    std::vector<int> res;
+    res.reserve(n);
+
+    std::vector<bool> used(n, false);
+
+    // the memory ranges for the set of currently concurrent nodes
+    ggml_mem_ranges * mrs0 = ggml_mem_ranges_init(0);
+
+    // the memory ranges for the set of nodes that haven't been processed yet, when looking forward for a node to reorder
+    ggml_mem_ranges * mrs1 = ggml_mem_ranges_init(0);
+
+    for (int i0 = 0; i0 < n; i0++) {
+        if (used[i0]) {
+            continue;
+        }
+
+        const auto & node0 = nodes[i0];
+
+        // the node is not concurrent with the existing concurrent set, so we have to "put a barrier" (i.e reset mrs0)
+        // but before we do that, look forward for some other nodes that can be added to the concurrent set mrs0
+        //
+        // note: we can always add empty nodes to the concurrent set as they don't read nor write anything
+        if (!node0.is_empty() && !h_check(mrs0, node0)) {
+            // this will hold the set of memory ranges from the nodes that haven't been processed yet
+            // if a node is not concurrent with this set, we cannot reorder it
+            ggml_mem_ranges_reset(mrs1);
+
+            // initialize it with the current node
+            h_add(mrs1, node0);
+
+            // that many nodes forward to search for a concurrent node
+            constexpr int N_FORWARD = 8;
+
+            for (int i1 = i0 + 1; i1 < i0 + N_FORWARD && i1 < n; i1++) {
+                if (used[i1]) {
+                    continue;
+                }
+
+                const auto & node1 = nodes[i1];
+
+                // disallow reordering of certain ops
+                if (!h_safe(node1.op())) {
+                    break;
+                }
+
+                const bool is_empty = node1.is_empty();
+
+                // to reorder a node and add it to the concurrent set, it has to be:
+                //   + empty or concurrent with all nodes in the existing concurrent set (mrs0)
+                //   + concurrent with all nodes prior to it that haven't been processed yet (mrs1)
+                if ((is_empty || h_check(mrs0, node1)) && h_check(mrs1, node1)) {
+                    // add the node to the existing concurrent set (i.e. reorder it for early execution)
+                    h_add(mrs0, node1);
+                    res.push_back(i1);
+
+                    // mark as used, so we skip re-processing it later
+                    used[i1] = true;
+                } else {
+                    // expand the set of nodes that haven't been processed yet
+                    h_add(mrs1, node1);
+                }
+            }
+
+            // finalize the concurrent set and begin a new one
+            ggml_mem_ranges_reset(mrs0);
+        }
+
+        // expand the concurrent set with the current node
+        {
+            h_add(mrs0, node0);
+            res.push_back(i0);
+        }
+    }
+
+    ggml_mem_ranges_free(mrs0);
+    ggml_mem_ranges_free(mrs1);
+
+    return res;
+}
+
+void ggml_metal_graph_optimize(ggml_cgraph * gf) {
+    constexpr int MAX_FUSE = 16;
+
+    const int n = gf->n_nodes;
+
+    enum ggml_op ops[MAX_FUSE];
+
+    std::vector<node_info> nodes;
+    nodes.reserve(gf->n_nodes);
+
+    // fuse nodes:
+    // we don't want to make reorders that break fusing, so we first pack all fusable tensors
+    //   and perform the reorder over the fused nodes. after the reorder is done, we unfuse
+    for (int i = 0; i < n; i++) {
+        node_info node = {
+            /*.node =*/ gf->nodes[i],
+            /*.fused =*/ {},
+        };
+
+        // fuse only ops that start with these operations
+        // can be expanded when needed
+        if (node.op() == GGML_OP_ADD ||
+            node.op() == GGML_OP_RMS_NORM) {
+            ops[0] = node.op();
+
+            int f = i + 1;
+            while (f < n && f < i + MAX_FUSE) {
+                // conservatively allow fusing only these ops
+                // can be expanded when needed
+                if (gf->nodes[f]->op != GGML_OP_ADD &&
+                    gf->nodes[f]->op != GGML_OP_MUL &&
+                    gf->nodes[f]->op != GGML_OP_RMS_NORM) {
+                    break;
+                }
+                ops[f - i] = gf->nodes[f]->op;
+                f++;
+            }
+
+            f -= i;
+            for (; f > 1; f--) {
+                if (ggml_can_fuse(gf, i, ops, f)) {
+                    break;
+                }
+            }
+
+            // add the fused tensors into the node info so we can unfuse them later
+            for (int k = 1; k < f; k++) {
+                ++i;
+
+                // the .dst() becomes the last fused tensor
+                node.add_fused(gf->nodes[i]);
+            }
+        }
+
+        nodes.push_back(std::move(node));
+    }
+
+#if 1
+    // reorder to improve concurrency
+    const auto order = ggml_metal_graph_optimize_reorder(nodes);
+#else
+    std::vector<int> order(nodes.size());
+    for (size_t i = 0; i < nodes.size(); i++) {
+        order[i] = i;
+    }
+#endif
+
+    // unfuse
+    {
+        int j = 0;
+        for (const auto i : order) {
+            const auto & node = nodes[i];
+
+            gf->nodes[j++] = node.node;
+
+            for (auto * fused : node.fused) {
+                gf->nodes[j++] = fused;
+            }
+        }
+    }
+}

ggml/src/ggml-metal/ggml-metal-common.h

@@ -0,0 +1,52 @@
+// helper functions for ggml-metal that are too difficult to implement in Objective-C
+
+#pragma once
+
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+enum ggml_mem_range_type {
+    MEM_RANGE_TYPE_SRC = 0,
+    MEM_RANGE_TYPE_DST = 1,
+};
+
+// a helper object that can be used for reordering operations to improve concurrency
+//
+// the fundamental idea is that a set of tasks (either ggml ops, or something else) can run concurrently if they
+//   don't write to a memory that is being read by another task or written to by another task in the set
+//
+// with this structure, we can add tasks to the set, setting memory constraints. we can also check if a new task
+//   can be added to the set without violating the constraints (i.e. if it can be executed concurrently with the
+//   tasks already in the set)
+//
+struct ggml_mem_ranges;
+
+struct ggml_mem_ranges * ggml_mem_ranges_init(int debug);
+void ggml_mem_ranges_free(struct ggml_mem_ranges * mrs);
+
+// remove all ranges from the set
+void ggml_mem_ranges_reset(struct ggml_mem_ranges * mrs);
+
+// add src or dst ranges to track
+bool ggml_mem_ranges_add(struct ggml_mem_ranges * mrs, const struct ggml_tensor * tensor);
+
+// return false if:
+// - new src range overlaps with any existing dst range
+// - new dst range overlaps with any existing range (src or dst)
+bool ggml_mem_ranges_check(const struct ggml_mem_ranges * mrs, const struct ggml_tensor * tensor);
+
+// reorder the nodes in the graph to improve concurrency, while respecting fusion
+//
+// note: this implementation is generic and not specific to metal
+//       if it proves to work well, we can start using it for other backends in the future
+void ggml_metal_graph_optimize(struct ggml_cgraph * gf);
+
+#ifdef __cplusplus
+}
+#endif

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

@@ -20,8 +20,8 @@
 #define N_R0_Q5_1 4
 #define N_SG_Q5_1 2
 
-#define N_R0_Q8_0 4
-#define N_SG_Q8_0 2
+#define N_R0_Q8_0 2
+#define N_SG_Q8_0 4
 
 #define N_R0_MXFP4 2
 #define N_SG_MXFP4 2
@@ -68,6 +68,11 @@
 #define N_R0_IQ4_XS 2
 #define N_SG_IQ4_XS 2
 
+// function constants offsets
+#define FC_FLASH_ATTN_EXT              100
+#define FC_FLASH_ATTN_EXT_VEC          200
+#define FC_FLASH_ATTN_EXT_VEC_REDUCE   300
+
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -236,9 +241,11 @@ typedef struct {
     int32_t  ne11;
     int32_t  ne_12_2; // assume K and V are same shape
     int32_t  ne_12_3;
+    int32_t  ns10;
     uint64_t nb11;
     uint64_t nb12;
     uint64_t nb13;
+    int32_t  ns20;
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
@@ -258,10 +265,43 @@ typedef struct {
     float    logit_softcap;
 } ggml_metal_kargs_flash_attn_ext;
 
+typedef struct {
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne_12_2; // assume K and V are same shape
+    int32_t  ne_12_3;
+    int32_t  ns10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ns20;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    float    scale;
+    float    max_bias;
+    float    m0;
+    float    m1;
+    int32_t  n_head_log2;
+    float    logit_softcap;
+} ggml_metal_kargs_flash_attn_ext_vec;
+
 typedef struct {
     int32_t  nrows;
-    int32_t  ne20;
-} ggml_metal_kargs_flash_attn_ext_reduce;
+} ggml_metal_kargs_flash_attn_ext_vec_reduce;
 
 typedef struct {
     int32_t  ne00;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -2838,6 +2838,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .optimize_graph          = */ NULL,
 };
 
 ggml_backend_t ggml_backend_opencl_init(void) {

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

@@ -26,8 +26,8 @@ kernel void kernel_timestep_embedding(
     local_half_dim = logical_dim / 2;
     local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
 
-    if (logical_dim % 2 != 0 && local_j == ((logical_dim + 1) / 2)) {
-        local_embed_data_ptr[logical_dim] = 0.0f;
+    if (logical_dim % 2 != 0 && local_j == local_half_dim) {
+        local_embed_data_ptr[2 * local_half_dim] = 0.0f;
     }
 
     if (local_j >= local_half_dim) {

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -795,6 +795,7 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .optimize_graph          = */ NULL,
 };
 
 ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {

ggml/src/ggml-sycl/binbcast.cpp

@@ -225,9 +225,9 @@ struct bin_bcast_sycl {
                     dpct::has_capability_or_fail(stream->get_device(),
                                                  {sycl::aspect::fp16});
 
-                    sycl_parallel_for(
-                        stream,
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size),
+                    stream->parallel_for(
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
+                                              sycl::range<3>(1, 1, block_size),
                                           sycl::range<3>(1, 1, block_size)),
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
@@ -246,8 +246,9 @@ struct bin_bcast_sycl {
                 dpct::has_capability_or_fail(stream->get_device(),
                                              {sycl::aspect::fp16});
 
-                sycl_parallel_for(
-                    stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                stream->parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
                                             s1, s2, s3, s01, s02, s03, s11, s12, s13,
@@ -302,6 +303,10 @@ inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, dst->src[0], dst->src[1], dst);
 }
 
+inline void ggml_sycl_op_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_count_equal>>(ctx, dst->src[0], dst->src[1], dst);
+}
+
 inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, dst->src[0], dst->src[1], dst);
@@ -327,6 +332,11 @@ void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_sub(ctx, dst);
 }
 
+void ggml_sycl_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_count_equal(ctx, dst);
+}
+
 void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     ggml_sycl_op_mul(ctx, dst);

ggml/src/ggml-sycl/binbcast.hpp

@@ -16,6 +16,12 @@ static __dpct_inline__ float op_sub(const float a, const float b) {
     return a - b;
 }
 
+static __dpct_inline__ float op_count_equal(const float a, const float b) {
+    return (a == b) ? 1.0f : 0.0f;
+}
+
+void ggml_sycl_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 static __dpct_inline__ float op_mul(const float a, const float b) {
     return a * b;
 }