server : add support for the RPC backend (#7305)

ref: #7292

.github/workflows/build.yml

@@ -340,6 +340,36 @@ jobs:
           cd build
           ctest -L main --verbose
 
+  ubuntu-latest-cmake-rpc:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake -DLLAMA_RPC=ON ..
+          cmake --build . --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose
+
   ubuntu-22-cmake-vulkan:
     runs-on: ubuntu-22.04
 
@@ -663,24 +693,28 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'noavx'
+          - build: 'rpc-x64'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_RPC=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'noavx-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx2'
+          - build: 'avx2-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx'
+          - build: 'avx-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx512'
+          - build: 'avx512-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'clblast'
+          - build: 'clblast-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
-          - build: 'openblas'
+          - build: 'openblas-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'kompute'
+          - build: 'kompute-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'vulkan'
+          - build: 'vulkan-x64'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'arm64'
-            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'llvm-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'msvc-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -691,13 +725,13 @@ jobs:
 
       - name: Clone Kompute submodule
         id: clone_kompute
-        if: ${{ matrix.build == 'kompute' }}
+        if: ${{ matrix.build == 'kompute-x64' }}
         run: |
           git submodule update --init kompute
 
       - name: Download OpenCL SDK
         id: get_opencl
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip"
           mkdir $env:RUNNER_TEMP/opencl
@@ -705,7 +739,7 @@ jobs:
 
       - name: Download CLBlast
         id: get_clblast
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
           curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
@@ -718,7 +752,7 @@ jobs:
 
       - name: Download OpenBLAS
         id: get_openblas
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
           curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
@@ -731,38 +765,41 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
 
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
       - name: Build
         id: cmake_build
         run: |
-          mkdir build
-          cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cmake -S . -B build ${{ matrix.defines }}
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Add clblast.dll
         id: add_clblast_dll
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
         run: |
           cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release
           cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
         run: |
           cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
           cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt
 
       - name: Check AVX512F support
         id: check_avx512f
-        if: ${{ matrix.build == 'avx512' }}
+        if: ${{ matrix.build == 'avx512-x64' }}
         continue-on-error: true
         run: |
           cd build
@@ -776,14 +813,14 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'clblast-x64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
 
       - name: Test (Intel SDE)
         id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
         run: |
           curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
           # for some weird reason windows tar doesn't like sde tar.xz
@@ -811,14 +848,14 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
-          name: llama-bin-win-${{ matrix.build }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
+          name: llama-bin-win-${{ matrix.build }}.zip
 
   windows-latest-cmake-cuda:
     runs-on: windows-latest
@@ -898,9 +935,9 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
-
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
         id: checkout
@@ -932,6 +969,17 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts

CMakeLists.txt

@@ -123,6 +123,7 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
+option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
@@ -296,7 +297,7 @@ if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
     endif()
-    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+    if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
         set(BLA_SIZEOF_INTEGER 8)
     endif()
 
@@ -494,6 +495,17 @@ if (LLAMA_MPI)
     endif()
 endif()
 
+if (LLAMA_RPC)
+    add_compile_definitions(GGML_USE_RPC)
+
+    if (WIN32)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ws2_32)
+    endif()
+
+    set(GGML_HEADERS_RPC ggml-rpc.h)
+    set(GGML_SOURCES_RPC ggml-rpc.cpp)
+endif()
+
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
@@ -995,6 +1007,11 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
         if (GGML_COMPILER_SUPPORT_DOTPROD)
             add_compile_definitions(__ARM_FEATURE_DOTPROD)
         endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
         check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
         if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
             add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
@@ -1176,6 +1193,7 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
             ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
             ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
             ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
@@ -1281,17 +1299,6 @@ install(
         WORLD_READ
         WORLD_EXECUTE
     DESTINATION ${CMAKE_INSTALL_BINDIR})
-install(
-    FILES convert-lora-to-ggml.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
 if (LLAMA_METAL)
     install(
         FILES ggml-metal.metal

CMakePresets.json

@@ -0,0 +1,45 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name":  "base",
+        "hidden": true,
+        "generator":   "Ninja",
+        "binaryDir":   "${sourceDir}/build-${presetName}",
+        "cacheVariables": {
+            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
+        }
+    },
+
+    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
+
+    {
+        "name": "arm64-windows-msvc", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
+        }
+    },
+
+    {
+        "name": "arm64-windows-llvm", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
+        }
+    },
+
+    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "release" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "release", "static" ] },
+
+    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] }
+  ]
+}

README.md

@@ -532,7 +532,7 @@ Building the program with BLAS support may lead to some performance improvements
         cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
         && cmake --build build --config Release -- -j 16
     ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
+    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
     However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
   - Using `make` (example for target gfx1030, build with 16 CPU threads):
@@ -712,6 +712,9 @@ Building the program with BLAS support may lead to some performance improvements
 
 ### Prepare and Quantize
 
+> [!NOTE]
+> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
+
 To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
 Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.

ci/run.sh

@@ -365,47 +365,6 @@ function gg_run_open_llama_3b_v2 {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/3B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
     set +e
 }
 
@@ -416,7 +375,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -429,11 +387,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # open_llama_7b_v2
@@ -549,48 +502,6 @@ function gg_run_open_llama_7b_v2 {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/7B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # currently not supported by the CUDA backend
-    # q8_0
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
     set +e
 }
 
@@ -601,7 +512,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -614,11 +524,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # bge-small

cmake/arm64-windows-llvm.cmake

@@ -0,0 +1,16 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

cmake/arm64-windows-msvc.cmake

@@ -0,0 +1,6 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )

common/common.cpp

@@ -1060,6 +1060,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
 #endif // GGML_USE_CUDA_SYCL_VULKAN
         return true;
     }
+    if (arg == "--rpc") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.rpc_servers = argv[i];
+        return true;
+    }
     if (arg == "--no-mmap") {
         params.use_mmap = false;
         return true;
@@ -1557,6 +1565,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
         printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
         printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
     }
+    printf("  --rpc SERVERS         comma separated list of RPC servers\n");
     printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
     printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
     printf("  -gan N, --grp-attn-n N\n");
@@ -1830,6 +1839,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
+    mparams.rpc_servers     = params.rpc_servers.c_str();
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;

common/common.h

@@ -82,6 +82,7 @@ struct gpt_params {
     float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
+    std::string rpc_servers       = "";    // comma separated list of RPC servers
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;

common/grammar-parser.cpp

@@ -26,7 +26,7 @@ namespace grammar_parser {
 
     static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
         uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
-        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
         return result.first->second;
     }
 

common/json-schema-to-grammar.cpp

@@ -272,7 +272,7 @@ private:
                     if (literal.empty()) {
                         return false;
                     }
-                    ret.push_back(std::make_pair(literal, true));
+                    ret.emplace_back(literal, true);
                     literal.clear();
                     return true;
                 };
@@ -298,7 +298,7 @@ private:
             while (i < length) {
                 char c = sub_pattern[i];
                 if (c == '.') {
-                    seq.push_back(std::make_pair(get_dot(), false));
+                    seq.emplace_back(get_dot(), false);
                     i++;
                 } else if (c == '(') {
                     i++;
@@ -307,7 +307,7 @@ private:
                             _warnings.push_back("Unsupported pattern syntax");
                         }
                     }
-                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
+                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
                 } else if (c == ')') {
                     i++;
                     if (start > 0 && sub_pattern[start - 1] != '(') {
@@ -331,9 +331,9 @@ private:
                     }
                     square_brackets += ']';
                     i++;
-                    seq.push_back(std::make_pair(square_brackets, false));
+                    seq.emplace_back(square_brackets, false);
                 } else if (c == '|') {
-                    seq.push_back(std::make_pair("|", false));
+                    seq.emplace_back("|", false);
                     i++;
                 } else if (c == '*' || c == '+' || c == '?') {
                     seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
@@ -417,7 +417,7 @@ private:
                         }
                     }
                     if (!literal.empty()) {
-                        seq.push_back(std::make_pair(literal, true));
+                        seq.emplace_back(literal, true);
                     }
                 }
             }

common/log.h

@@ -211,7 +211,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_FLF_FMT "%s"
@@ -224,7 +224,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
         #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
     #else
         #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
     #endif
 #else
     #define LOG_TEE_FLF_FMT "%s"
@@ -294,7 +294,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Main LOG macro.
 //  behaves like printf, and supports arguments the exact same way.
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
     #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
@@ -308,14 +308,14 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Secondary target can be changed just like LOG_TARGET
 //  by defining LOG_TEE_TARGET
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
     #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else

convert-hf-to-gguf-update.py

@@ -74,9 +74,9 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
-    {"name": "jina-en",        "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
-    {"name": "jina-es",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
-    {"name": "jina-de",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
 ]
 
 # make directory "models/tokenizers" if it doesn't exist

convert-hf-to-gguf.py

@@ -240,23 +240,6 @@ class Model:
         return False
 
     def write_tensors(self):
-        # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
-        def np_fp32_to_bf16(n: np.ndarray):
-            # force nan to quiet
-            n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
-            # flush subnormals to zero
-            n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
-            # round to nearest even
-            n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
-            return n.astype(np.int16)
-
-        # Doing this row-wise is much, much faster than element-wise, hence the signature
-        v_fp32_to_bf16 = np.vectorize(np_fp32_to_bf16, otypes=[np.int16], signature="(n)->(n)")
-        if self.lazy:
-            # TODO: find a way to implicitly wrap np.vectorize functions
-            # NOTE: the type is changed to reflect otypes passed to np.vectorize above
-            v_fp32_to_bf16 = gguf.LazyNumpyTensor._wrap_fn(v_fp32_to_bf16, meta_noop=np.int16)
-
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
 
         for name, data_torch in self.get_tensors():
@@ -309,27 +292,31 @@ class Model:
                 ))
 
                 if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                    if self.ftype == gguf.LlamaFileType.MOSTLY_F16:
+                    if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        data = gguf.quantize_bf16(data)
+                        assert data.dtype == np.int16
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
+                        data = gguf.quantize_q8_0(data)
+                        assert data.dtype == np.uint8
+                        data_qtype = gguf.GGMLQuantizationType.Q8_0
+
+                    else:  # default to float16 for quantized tensors
                         if data_dtype != np.float16:
                             data = data.astype(np.float16)
                         data_qtype = gguf.GGMLQuantizationType.F16
 
-                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        if data_dtype != np.float32:
-                            data = data.astype(np.float32)
-                        data = v_fp32_to_bf16(data.view(np.int32))
-                        assert data.dtype == np.int16
-                        data_qtype = gguf.GGMLQuantizationType.BF16
-
-                else:  # by default, convert to float32
+                if data_qtype is None:  # by default, convert to float32
                     if data_dtype != np.float32:
                         data = data.astype(np.float32)
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                assert data_qtype is not None
-
+                block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype]
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
+                shape_str = f"""{{{', '.join(str(n) for n in reversed(
+                    (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size))
+                )}}}"""
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -475,13 +462,13 @@ class Model:
             res = "dbrx"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
-            res = "jina-en"
+            res = "jina-v2-en"
         if chkhsh == "171aeeedd6fb548d418a7461d053f11b6f1f1fc9b387bd66640d28a4b9f5c643":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-es
-            res = "jina-es"
+            res = "jina-v2-es"
         if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
-            res = "jina-de"
+            res = "jina-v2-de"
 
         if res is None:
             logger.warning("\n")
@@ -859,6 +846,7 @@ class BaichuanModel(Model):
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "linear":
@@ -981,6 +969,7 @@ class XverseModel(Model):
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "linear":
@@ -1215,6 +1204,7 @@ class StableLMModel(Model):
         self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_file_type(self.ftype)
 
     _q_norms: list[dict[str, Tensor]] | None = None
     _k_norms: list[dict[str, Tensor]] | None = None
@@ -1591,6 +1581,7 @@ class QwenModel(Model):
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
         self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
 
 
 @Model.register("Qwen2ForCausalLM")
@@ -1828,6 +1819,7 @@ class PlamoModel(Model):
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_head_count_kv(5)  # hparams["num_key_value_heads"]) is wrong
         self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+        self.gguf_writer.add_file_type(self.ftype)
 
     def shuffle_attn_q_weight(self, data_torch):
         assert data_torch.size() == (5120, 5120)
@@ -2007,6 +1999,7 @@ in chat mode so that the conversation can end normally.")
         self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
         self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
         self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
+        self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
@@ -2415,25 +2408,15 @@ class LazyTorchTensor(gguf.LazyBase):
     def numpy(self) -> gguf.LazyNumpyTensor:
         dtype = self._dtype_map[self.dtype]
         return gguf.LazyNumpyTensor(
-            meta=np.lib.stride_tricks.as_strided(np.zeros(1, dtype), self.shape, (0 for _ in self.shape)),
+            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
             lazy=self._lazy,
             args=(self,),
             func=(lambda s: s[0].numpy())
         )
 
     @classmethod
-    def eager_to_meta(cls, t: Tensor) -> Tensor:
-        if t.is_meta:
-            return t
-        return t.detach().to("meta")
-
-    @classmethod
-    def meta_with_dtype(cls, m: Tensor, dtype: torch.dtype) -> Tensor:
-        m = m.detach()
-        if not m.is_meta:
-            m = m.to("meta")
-        m.dtype = dtype
-        return m
+    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: torch.Size) -> Tensor:
+        return torch.empty(size=shape, dtype=dtype, device="meta")
 
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):
@@ -2464,8 +2447,8 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -2523,6 +2506,7 @@ def main() -> None:
         "f32": gguf.LlamaFileType.ALL_F32,
         "f16": gguf.LlamaFileType.MOSTLY_F16,
         "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
         "auto": gguf.LlamaFileType.GUESSED,
     }
 

convert-lora-to-ggml.py

@@ -1,150 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any, BinaryIO, Sequence
-
-import numpy as np
-import torch
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-logging.basicConfig(level=logging.DEBUG)
-logger = logging.getLogger("lora-to-gguf")
-
-NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
-
-
-def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
-    fout.write(b"ggla"[::-1])  # magic (ggml lora)
-    fout.write(struct.pack("i", 1))  # file version
-    fout.write(struct.pack("i", params["r"]))
-    # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
-    # but some models ship a float value instead
-    # let's convert to int, but fail if lossless conversion is not possible
-    assert (
-        int(params["lora_alpha"]) == params["lora_alpha"]
-    ), "cannot convert float to int losslessly"
-    fout.write(struct.pack("i", int(params["lora_alpha"])))
-
-
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
-    sname = name.encode("utf-8")
-    fout.write(
-        struct.pack(
-            "iii",
-            len(shape),
-            len(sname),
-            NUMPY_TYPE_TO_FTYPE[data_type.name],
-        )
-    )
-    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-    fout.write(sname)
-    fout.seek((fout.tell() + 31) & -32)
-
-
-if __name__ == '__main__':
-    if len(sys.argv) < 2:
-        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
-        logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
-        logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-        sys.exit(1)
-
-    input_json = os.path.join(sys.argv[1], "adapter_config.json")
-    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
-
-    if os.path.exists(input_model):
-        model = torch.load(input_model, map_location="cpu")
-    else:
-        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-        model = load_file(input_model, device="cpu")
-
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
-
-    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-        logger.error(f"Error: unsupported architecture {arch_name}")
-        sys.exit(1)
-
-    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
-
-    with open(input_json, "r") as f:
-        params = json.load(f)
-
-    if params["peft_type"] != "LORA":
-        logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-        sys.exit(1)
-
-    if params["fan_in_fan_out"] is True:
-        logger.error("Error: param fan_in_fan_out is not supported")
-        sys.exit(1)
-
-    if params["bias"] is not None and params["bias"] != "none":
-        logger.error("Error: param bias is not supported")
-        sys.exit(1)
-
-    # TODO: these seem to be layers that have been trained but without lora.
-    # doesn't seem widely used but eventually should be supported
-    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-        logger.error("Error: param modules_to_save is not supported")
-        sys.exit(1)
-
-    with open(output_path, "wb") as fout:
-        fout.truncate()
-
-        write_file_header(fout, params)
-        for k, v in model.items():
-            orig_k = k
-            if k.endswith(".default.weight"):
-                k = k.replace(".default.weight", ".weight")
-            if k in ["llama_proj.weight", "llama_proj.bias"]:
-                continue
-            if k.endswith("lora_A.weight"):
-                if v.dtype != torch.float16 and v.dtype != torch.float32:
-                    v = v.float()
-                v = v.T
-            else:
-                v = v.float()
-
-            t = v.detach().numpy()
-
-            prefix = "base_model.model."
-            if k.startswith(prefix):
-                k = k[len(prefix) :]
-
-            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-            if k.endswith(lora_suffixes):
-                suffix = k[-len(lora_suffixes[0]):]
-                k = k[: -len(lora_suffixes[0])]
-            else:
-                logger.error(f"Error: unrecognized tensor name {orig_k}")
-                sys.exit(1)
-
-            tname = name_map.get_name(k)
-            if tname is None:
-                logger.error(f"Error: could not map tensor name {orig_k}")
-                logger.error(" Note: the arch parameter must be specified if the model is not llama")
-                sys.exit(1)
-
-            if suffix == ".lora_A.weight":
-                tname += ".weight.loraA"
-            elif suffix == ".lora_B.weight":
-                tname += ".weight.loraB"
-            else:
-                assert False
-
-            logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-            write_tensor_header(fout, tname, t.shape, t.dtype)
-            t.tofile(fout)
-
-    logger.info(f"Converted {input_json} and {input_model} to {output_path}")

convert.py

@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -344,10 +344,47 @@ class Params:
         return params
 
 
+@dataclass
+class Metadata:
+    name: Optional[str] = None
+    author: Optional[str] = None
+    version: Optional[str] = None
+    url: Optional[str] = None
+    description: Optional[str] = None
+    licence: Optional[str] = None
+    source_url: Optional[str] = None
+    source_hf_repo: Optional[str] = None
+
+    @staticmethod
+    def load(metadata_path: Path) -> Metadata:
+        if metadata_path is None or not metadata_path.exists():
+            return Metadata()
+
+        with open(metadata_path, 'r') as file:
+            data = json.load(file)
+
+        # Create a new Metadata instance
+        metadata = Metadata()
+
+        # Assigning values to Metadata attributes if they exist in the JSON file
+        # This is based on LLM_KV_NAMES mapping in llama.cpp
+        metadata.name = data.get("general.name")
+        metadata.author = data.get("general.author")
+        metadata.version = data.get("general.version")
+        metadata.url = data.get("general.url")
+        metadata.description = data.get("general.description")
+        metadata.license = data.get("general.license")
+        metadata.source_url = data.get("general.source.url")
+        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
+
+        return metadata
+
+
 #
 # vocab
 #
 
+
 @runtime_checkable
 class BaseVocab(Protocol):
     tokenizer_model: ClassVar[str]
@@ -1066,21 +1103,42 @@ class OutputFile:
     def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
-    def add_meta_arch(self, params: Params) -> None:
+    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
+        # Metadata About The Model And Its Provenence
         name = "LLaMA"
-
-        # TODO: better logic to determine model name
-        if params.n_ctx == 4096:
-            name = "LLaMA v2"
+        if metadata is not None and metadata.name is not None:
+            name = metadata.name
         elif params.path_model is not None:
-            name = str(params.path_model.parent).split('/')[-1]
+            name = params.path_model.name
+        elif params.n_ctx == 4096:
+            # Heuristic detection of LLaMA v2 model
+            name = "LLaMA v2"
 
-        self.gguf.add_name                (name)
-        self.gguf.add_vocab_size          (params.n_vocab)
-        self.gguf.add_context_length      (params.n_ctx)
-        self.gguf.add_embedding_length    (params.n_embd)
-        self.gguf.add_block_count         (params.n_layer)
-        self.gguf.add_feed_forward_length (params.n_ff)
+        self.gguf.add_name(name)
+
+        if metadata is not None:
+            if metadata.author is not None:
+                self.gguf.add_author(metadata.author)
+            if metadata.version is not None:
+                self.gguf.add_version(metadata.version)
+            if metadata.url is not None:
+                self.gguf.add_url(metadata.url)
+            if metadata.description is not None:
+                self.gguf.add_description(metadata.description)
+            if metadata.licence is not None:
+                self.gguf.add_licence(metadata.licence)
+            if metadata.source_url is not None:
+                self.gguf.add_source_url(metadata.source_url)
+            if metadata.source_hf_repo is not None:
+                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
+
+    def add_meta_arch(self, params: Params) -> None:
+        # Metadata About The Neural Architecture Itself
+        self.gguf.add_vocab_size(params.n_vocab)
+        self.gguf.add_context_length(params.n_ctx)
+        self.gguf.add_embedding_length(params.n_embd)
+        self.gguf.add_block_count(params.n_layer)
+        self.gguf.add_feed_forward_length(params.n_ff)
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
@@ -1183,13 +1241,14 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
+        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         of.add_meta_vocab(vocab)
         of.add_meta_special_vocab(svocab)
@@ -1216,12 +1275,14 @@ class OutputFile:
         fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
+        metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
         # meta data
+        of.add_meta_model(params, metadata)
         of.add_meta_arch(params)
         if isinstance(vocab, Vocab):
             of.add_meta_vocab(vocab)
@@ -1257,6 +1318,37 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
     raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
+def model_parameter_count(model: LazyModel) -> int:
+    total_model_parameters = 0
+    for i, (name, lazy_tensor) in enumerate(model.items()):
+        sum_weights_in_tensor = 1
+        for dim in lazy_tensor.shape:
+            sum_weights_in_tensor *= dim
+        total_model_parameters += sum_weights_in_tensor
+    return total_model_parameters
+
+
+def model_parameter_count_rounded_notation(model_params_count: int) -> str:
+    if model_params_count > 1e12 :
+        # Trillions Of Parameters
+        scaled_model_params = model_params_count * 1e-12
+        scale_suffix = "T"
+    elif model_params_count > 1e9 :
+        # Billions Of Parameters
+        scaled_model_params = model_params_count * 1e-9
+        scale_suffix = "B"
+    elif model_params_count > 1e6 :
+        # Millions Of Parameters
+        scaled_model_params = model_params_count * 1e-6
+        scale_suffix = "M"
+    else:
+        # Thousands Of Parameters
+        scaled_model_params = model_params_count * 1e-3
+        scale_suffix = "K"
+
+    return f"{round(scaled_model_params)}{scale_suffix}"
+
+
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
     return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
             for (name, tensor) in model.items()}
@@ -1436,13 +1528,35 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
-    namestr = {
-        GGMLFileType.AllF32:    "f32",
-        GGMLFileType.MostlyF16: "f16",
-        GGMLFileType.MostlyQ8_0:"q8_0",
+def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
+    quantization = {
+        GGMLFileType.AllF32:    "F32",
+        GGMLFileType.MostlyF16: "F16",
+        GGMLFileType.MostlyQ8_0: "Q8_0",
     }[file_type]
-    ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
+
+    parameters = model_parameter_count_rounded_notation(model_params_count)
+
+    expert_count = ""
+    if params.n_experts is not None:
+        expert_count = f"{params.n_experts}x"
+
+    version = ""
+    if metadata is not None and metadata.version is not None:
+        version = f"-{metadata.version}"
+
+    name = "ggml-model"
+    if metadata is not None and metadata.name is not None:
+        name = metadata.name
+    elif params.path_model is not None:
+        name = params.path_model.name
+
+    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
+
+
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
+    ret = model_paths[0].parent / f"{default_filename}.gguf"
     if ret in model_paths:
         logger.error(
             f"Error: Default output path ({ret}) would overwrite the input. "
@@ -1480,17 +1594,30 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
     parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
+    parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
 
     args = parser.parse_args(args_in)
 
     if args.verbose:
         logging.basicConfig(level=logging.DEBUG)
-    elif args.dump_single or args.dump:
+    elif args.dump_single or args.dump or args.get_outfile:
         # Avoid printing anything besides the dump output
         logging.basicConfig(level=logging.WARNING)
     else:
         logging.basicConfig(level=logging.INFO)
 
+    metadata = Metadata.load(args.metadata)
+
+    if args.get_outfile:
+        model_plus = load_some_model(args.model)
+        params = Params.load(model_plus)
+        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = model_parameter_count(model_plus.model)
+        ftype   = pick_output_type(model, args.outtype)
+        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
+        return
+
     if args.no_vocab and args.vocab_only:
         raise ValueError("--vocab-only does not make sense with --no-vocab")
 
@@ -1504,6 +1631,9 @@ def main(args_in: list[str] | None = None) -> None:
     else:
         model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
 
+    model_params_count = model_parameter_count(model_plus.model)
+    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
+
     if args.dump:
         do_dump_model(model_plus)
         return
@@ -1557,7 +1687,7 @@ def main(args_in: list[str] | None = None) -> None:
                 f_norm_eps = 1e-5,
             )
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess=endianess, pad_vocab=args.pad_vocab)
+                                    endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
         logger.info(f"Wrote {outfile}")
         return
 
@@ -1570,13 +1700,13 @@ def main(args_in: list[str] | None = None) -> None:
     model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")
 
     OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
     logger.info(f"Wrote {outfile}")
 
 

examples/CMakeLists.txt

@@ -49,4 +49,7 @@ else()
         add_subdirectory(server)
     endif()
     add_subdirectory(export-lora)
+    if (LLAMA_RPC)
+        add_subdirectory(rpc)
+    endif()
 endif()

examples/embedding/embedding.cpp

@@ -211,6 +211,7 @@ int main(int argc, char ** argv) {
 
     // clean up
     llama_print_timings(ctx);
+    llama_batch_free(batch);
     llama_free(ctx);
     llama_free_model(model);
     llama_backend_free();

examples/llava/llava-cli.cpp

@@ -300,14 +300,10 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    for (auto & image : params.image) {
+    if (prompt_contains_image(params.prompt)) {
         auto ctx_llava = llava_init_context(&params, model);
 
-        auto image_embed = load_image(ctx_llava, &params, image);
-        if (!image_embed) {
-            std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
-            return 1;
-        }
+        auto image_embed = load_image(ctx_llava, &params, "");
 
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
@@ -316,7 +312,26 @@ int main(int argc, char ** argv) {
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
+    } else {
+        for (auto & image : params.image) {
+            auto ctx_llava = llava_init_context(&params, model);
+
+            auto image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_print_timings(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
     }
+
     llama_free_model(model);
 
     return 0;

examples/llava/llava.cpp

@@ -88,7 +88,6 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
 // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
 static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
     struct {
-        struct ggml_tensor * newline;
         struct ggml_context * ctx;
     } model;
 
@@ -150,20 +149,6 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     model.ctx = ggml_init(params);
 
-    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
-    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
-    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
-        if (newline_tmp->buffer == NULL) {
-            LOG_TEE("newline_tmp tensor buffer is NULL\n");
-        }
-        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
-    } else {
-        model.newline->data = newline_tmp->data;
-        if (model.newline->data == NULL) {
-            LOG_TEE("newline_tmp tensor data is NULL\n");
-        }
-    }
-
     struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
     // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
     // fill it with the image embeddings, ignoring the base

examples/perplexity/README.md

@@ -7,6 +7,8 @@ Also note that finetunes typically result in a higher perplexity value even thou
 
 Within llama.cpp the perplexity of base models is used primarily to judge the quality loss from e.g. quantized models vs. FP16.
 The convention among contributors is to use the Wikitext-2 test set for testing unless noted otherwise (can be obtained with `scripts/get-wikitext-2.sh`).
+When numbers are listed all command line arguments and compilation options are left at their defaults unless noted otherwise.
+llama.cpp numbers are **not** directly comparable to those of other projects because the exact values depend strongly on the implementation details.
 
 By default only the mean perplexity value and the corresponding uncertainty is calculated.
 The uncertainty is determined empirically by assuming a Gaussian distribution of the "correct" logits per and then applying error propagation.
@@ -32,7 +34,13 @@ In addition to the KL divergence the following statistics are calculated with `-
 
 ## LLaMA 3 8b Scoreboard
 
-Results are sorted by Kullback-Leibler divergence relative to FP16.
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
+Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
 The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
 
 | Quantization | imatrix | Model size [GiB] | PPL                    | ΔPPL                   | KLD                   | Mean Δp           | RMS Δp           |
@@ -89,6 +97,12 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence
 
 ## LLaMA 2 vs. LLaMA 3 Quantization comparison
 
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
 | Metric          |          L2 7b q2_K |          L3 8b q2_K |        L2 7b q4_K_M |        L3 8b q4_K_M |          L2 7b q6_K |          L3 8b q6_K |          L2 7b q8_0 |          L3 8b q8_0 |
 |-----------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|
 | Mean PPL        | 5.794552 ± 0.032298 | 9.751568 ± 0.063312 | 5.877078 ± 0.032781 | 6.407115 ± 0.039119 | 5.808494 ± 0.032425 | 6.253382 ± 0.038078 | 5.798542 ± 0.032366 | 6.234284 ± 0.037878 |
@@ -107,6 +121,50 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence
 | RMS Δp          |     9.762 ± 0.053 % |    21.421 ± 0.079 % |     3.252 ± 0.024 % |     5.519 ± 0.050 % |     1.339 ± 0.010 % |     2.295 ± 0.019 % |     0.618 ± 0.011 % |     1.198 ± 0.007 % |
 | Same top p      |    85.584 ± 0.086 % |    71.138 ± 0.119 % |    94.665 ± 0.055 % |    91.901 ± 0.072 % |    97.520 ± 0.038 % |    96.031 ± 0.051 % |    98.846 ± 0.026 % |    97.674 ± 0.040 % |
 
+## LLaMA 3 BF16 vs. FP16 comparison
+
+| Revision | 83330d8c      |
+|:---------|:--------------|
+| Backend  | CPU           |
+| CPU      | AMD Epyc 7742 |
+| GPU      | N/A           |
+
+Results were calculated with LLaMA 3 8b BF16 as `--kl-divergence-base` and LLaMA 3 8b FP16 as the `--model` for comparison.
+
+| Metric                         |                    Value |
+|--------------------------------|--------------------------|
+| Mean PPL(Q)                    |      6.227711 ± 0.037833 |
+| Mean PPL(base)                 |      6.225194 ± 0.037771 |
+| Cor(ln(PPL(Q)), ln(PPL(base))) |                  99.990% |
+| Mean ln(PPL(Q)/PPL(base))      |      0.000404 ± 0.000086 |
+| Mean PPL(Q)/PPL(base)          |      1.000404 ± 0.000086 |
+| Mean PPL(Q)-PPL(base)          |      0.002517 ± 0.000536 |
+| Mean    KLD                    |  0.00002515 ± 0.00000020 |
+| Maximum KLD                    |                 0.012206 |
+| 99.9%   KLD                    |                 0.000799 |
+| 99.0%   KLD                    |                 0.000222 |
+| 99.0%   KLD                    |                 0.000222 |
+| Median  KLD                    |                 0.000013 |
+| 10.0%   KLD                    |                -0.000002 |
+| 5.0%   KLD                     |                -0.000008 |
+| 1.0%   KLD                     |                -0.000023 |
+| Minimum KLD                    |                -0.000059 |
+| Mean    Δp                     | -0.0000745 ± 0.0003952 % |
+| Maximum Δp                     |                   4.186% |
+| 99.9%   Δp                     |                   1.049% |
+| 99.0%   Δp                     |                   0.439% |
+| 95.0%   Δp                     |                   0.207% |
+| 90.0%   Δp                     |                   0.125% |
+| 75.0%   Δp                     |                   0.029% |
+| Median  Δp                     |                   0.000% |
+| 25.0%   Δp                     |                  -0.030% |
+| 10.0%   Δp                     |                  -0.126% |
+| 5.0%   Δp                      |                  -0.207% |
+| 1.0%   Δp                      |                  -0.434% |
+| 0.1%   Δp                      |                  -1.016% |
+| Minimum Δp                     |                  -4.672% |
+| RMS Δp                         |          0.150 ± 0.001 % |
+| Same top p                     |         99.739 ± 0.013 % |
 
 ## Old Numbers
 

examples/quantize/README.md

@@ -1,6 +1,8 @@
 # quantize
 
-TODO
+You can also use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to build your own quants without any setup.
+
+Note: It is synced from llama.cpp `main` every 6 hours.
 
 ## Llama 2 7B
 

examples/rpc/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_executable(rpc-server rpc-server.cpp)
+target_link_libraries(rpc-server PRIVATE ggml llama)

examples/rpc/README.md

@@ -0,0 +1,74 @@
+## Overview
+
+The `rpc-server` allows  running `ggml` backend on a remote host.
+The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
+This can be used for distributed LLM inference with `llama.cpp` in the following way:
+
+```mermaid
+flowchart TD
+    rpcb---|TCP|srva
+    rpcb---|TCP|srvb
+    rpcb-.-|TCP|srvn
+    subgraph hostn[Host N]
+    srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph hostb[Host B]
+    srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph hosta[Host A]
+    srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph host[Main Host]
+    ggml[llama.cpp]---rpcb[RPC backend]
+    end
+    style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
+```
+
+Each host can run a different backend, e.g. one with CUDA and another with Metal.
+You can also run multiple `rpc-server` instances on the same host, each with a different backend.
+
+## Usage
+
+On each host, build the corresponding backend with `cmake` and add `-DLLAMA_RPC=ON` to the build options.
+For example, to build the CUDA backend with RPC support:
+
+```bash
+mkdir build-rpc-cuda
+cd build-rpc-cuda
+cmake .. -DLLAMA_CUDA=ON -DLLAMA_RPC=ON
+cmake --build . --config Release
+```
+
+Then, start the `rpc-server` with the backend:
+
+```bash
+$ bin/rpc-server -p 50052
+create_backend: using CUDA backend
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
+ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA T1200 Laptop GPU, compute capability 7.5, VMM: yes
+Starting RPC server on 0.0.0.0:50052
+```
+
+When using the CUDA backend, you can specify the device with the `CUDA_VISIBLE_DEVICES` environment variable, e.g.:
+```bash
+$ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
+```
+This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
+
+
+On the main host build `llama.cpp` only with `-DLLAMA_RPC=ON`:
+
+```bash
+mkdir build-rpc
+cd build-rpc
+cmake .. -DLLAMA_RPC=ON
+cmake --build . --config Release
+```
+
+Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
+
+```bash
+$ bin/main -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
+```

examples/rpc/rpc-server.cpp

@@ -0,0 +1,130 @@
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include "ggml-rpc.h"
+#ifdef _WIN32
+#  include <windows.h>
+#else
+#  include <unistd.h>
+#endif
+#include <string>
+#include <stdio.h>
+
+struct rpc_server_params {
+    std::string host        = "0.0.0.0";
+    int         port        = 50052;
+    size_t      backend_mem = 0;
+};
+
+static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
+    fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
+    fprintf(stderr, "\n");
+}
+
+static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & params) {
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg == "-H" || arg == "--host") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.host = argv[i];
+        } else if (arg == "-p" || arg == "--port") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.port = std::stoi(argv[i]);
+            if (params.port <= 0 || params.port > 65535) {
+                return false;
+            }
+        } else if (arg == "-m" || arg == "--mem") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.backend_mem = std::stoul(argv[i]) * 1024 * 1024;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+    return true;
+}
+
+static ggml_backend_t create_backend() {
+    ggml_backend_t backend = NULL;
+#ifdef GGML_USE_CUDA
+    fprintf(stderr, "%s: using CUDA backend\n", __func__);
+    backend = ggml_backend_cuda_init(0); // init device 0
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+    }
+#elif GGML_USE_METAL
+    fprintf(stderr, "%s: using Metal backend\n", __func__);
+    backend = ggml_backend_metal_init();
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
+    }
+#endif
+
+    // if there aren't GPU Backends fallback to CPU backend
+    if (!backend) {
+        fprintf(stderr, "%s: using CPU backend\n", __func__);
+        backend = ggml_backend_cpu_init();
+    }
+    return backend;
+}
+
+static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
+#ifdef GGML_USE_CUDA
+    ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
+#else
+    #ifdef _WIN32
+        MEMORYSTATUSEX status;
+        status.dwLength = sizeof(status);
+        GlobalMemoryStatusEx(&status);
+        *total_mem = status.ullTotalPhys;
+        *free_mem = status.ullAvailPhys;
+    #else
+        long pages = sysconf(_SC_PHYS_PAGES);
+        long page_size = sysconf(_SC_PAGE_SIZE);
+        *total_mem = pages * page_size;
+        *free_mem = *total_mem;
+    #endif
+#endif
+}
+
+int main(int argc, char * argv[]) {
+    rpc_server_params params;
+    if (!rpc_server_params_parse(argc, argv, params)) {
+        fprintf(stderr, "Invalid parameters\n");
+        return 1;
+    }
+    ggml_backend_t backend = create_backend();
+    if (!backend) {
+        fprintf(stderr, "Failed to create backend\n");
+        return 1;
+    }
+    std::string endpoint = params.host + ":" + std::to_string(params.port);
+    size_t free_mem, total_mem;
+    if (params.backend_mem > 0) {
+        free_mem = params.backend_mem;
+        total_mem = params.backend_mem;
+    } else {
+        get_backend_memory(&free_mem, &total_mem);
+    }
+    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
+    start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
+    ggml_backend_free(backend);
+    return 0;
+}

examples/server/README.md

@@ -17,7 +17,8 @@ The project is under active development, and we are [looking for feedback and co
 
 **Command line options:**
 
-- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
+- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
+- `-t N`, `--threads N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
 - `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
@@ -36,9 +37,7 @@ The project is under active development, and we are [looking for feedback and co
 - `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems
 - `--numa distribute`: Spread execution evenly over all nodes
 - `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
-- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system
-page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
-
+- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
 - `--numa`: Attempt optimizations that may help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
@@ -48,7 +47,7 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/
 - `--path`: Path from which to serve static files. Default: disabled
 - `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
-- `--embedding`: Enable embedding extraction. Default: disabled
+- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
 - `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
 - `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)

examples/server/server.cpp

@@ -671,6 +671,13 @@ struct server_context {
             model = nullptr;
         }
 
+        // Clear any sampling context
+        for (server_slot & slot : slots) {
+            if (slot.ctx_sampling != nullptr) {
+                llama_sampling_free(slot.ctx_sampling);
+            }
+        }
+
         llama_batch_free(batch);
     }
 
@@ -2380,6 +2387,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     printf("  --lora-base FNAME         optional model to use as a base for the layers modified by the LoRA adapter\n");
     printf("  --host                    ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     printf("  --port PORT               port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --rpc SERVERS             comma separated list of RPC servers\n");
     printf("  --path PUBLIC_PATH        path from which to serve static files (default: disabled)\n");
     printf("  --api-key API_KEY         optional api key to enhance server security. If set, requests must include this key for access.\n");
     printf("  --api-key-file FNAME      path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
@@ -2432,6 +2440,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 break;
             }
             sparams.port = std::stoi(argv[i]);
+        } else if (arg == "--rpc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rpc_servers = argv[i];
         } else if (arg == "--host") {
             if (++i >= argc) {
                 invalid_param = true;

examples/server/tests/features/steps/steps.py

@@ -887,6 +887,7 @@ async def oai_chat_completions(user_prompt,
                                base_path,
                                async_client,
                                debug=False,
+                               temperature=None,
                                model=None,
                                n_predict=None,
                                enable_streaming=None,
@@ -913,7 +914,8 @@ async def oai_chat_completions(user_prompt,
         "model": model,
         "max_tokens": n_predict,
         "stream": enable_streaming,
-        "seed": seed
+        "temperature": temperature if temperature is not None else 0.0,
+        "seed": seed,
     }
     if response_format is not None:
         payload['response_format'] = response_format
@@ -978,7 +980,8 @@ async def oai_chat_completions(user_prompt,
                 max_tokens=n_predict,
                 stream=enable_streaming,
                 response_format=payload.get('response_format'),
-                seed=seed
+                seed=seed,
+                temperature=payload['temperature']
             )
         except openai.error.AuthenticationError as e:
             if expect_api_error is not None and expect_api_error:

examples/server/utils.hpp

@@ -371,7 +371,7 @@ static json oaicompat_completion_params_parse(
     llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
     llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
     llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       0.0);
+    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
     llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
 
     // Apply chat template to the list of messages

ggml-backend.c

@@ -1895,7 +1895,6 @@ void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * t
 
     tensor->buffer = buffer;
     tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    tensor->backend = tensor->view_src->backend;
     ggml_backend_buffer_init_tensor(buffer, tensor);
 }
 

ggml-cuda.cu

@@ -2558,7 +2558,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         }
 
         // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
-        if (cuda_graph_update_required) {
+        if (use_cuda_graph && cuda_graph_update_required) {
             cuda_ctx->cuda_graph->number_consecutive_updates++;
         } else {
             cuda_ctx->cuda_graph->number_consecutive_updates = 0;
@@ -2713,6 +2713,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -2840,8 +2841,16 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
-        case GGML_OP_FLASH_ATTN_EXT:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+#else
+            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;
     }

ggml-cuda/common.cuh

@@ -321,6 +321,10 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 
 #define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
+static bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
 static bool fp16_mma_available(const int cc) {
     return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
 }

ggml-cuda/fattn-common.cuh

@@ -0,0 +1,47 @@
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}

ggml-cuda/fattn-vec-f16.cu

@@ -0,0 +1,430 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        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 int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    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    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    half  slopeh = __float2half(1.0f);
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slopeh = __float2half(powf(base, exph));
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            half2 sum2[ncols] = {{0.0f, 0.0f}};
+#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;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum2[j] = warp_reduce_sum(sum2[j]);
+                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
+            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    constexpr int cols_per_block = 1;
+    constexpr int parallel_blocks = 4;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 256:
+            launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f16.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-vec-f32.cu

@@ -0,0 +1,384 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f32.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        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 int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    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    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = powf(base, exph);
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    float2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE].x *= scale;
+            Q_h2[j][i0/WARP_SIZE].y *= scale;
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            float sum[ncols] = {0.0f};
+#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;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum[j] +=  __low2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].x;
+                    sum[j] += __high2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].y;
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum[j] = warp_reduce_sum(sum[j]);
+                sum[j] += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum[j]);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum[j];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = __half2float(V_h[(k_VKQ_0 + k)*stride_KV + tid]);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn.cu

@@ -1,4 +1,7 @@
 #include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
 #include "fattn.cuh"
 
 #include <cstdint>
@@ -7,251 +10,6 @@
 #include <mma.h>
 #endif
 
-#define FATTN_KQ_STRIDE       256
-#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
-#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
-
-template<int D, int ncols, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_vec_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        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 int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-#if FP16_AVAILABLE
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
-
-    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    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic0;
-
-    const int stride_KV  = nb11 / sizeof(half);
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    half slopeh = __float2half(1.0f);
-
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const int h = blockIdx.y;
-
-        const float base = h < n_head_log2 ? m0 : m1;
-        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
-
-        slopeh = __float2half(powf(base, exph));
-    }
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ half KQ[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -HALF_MAX_HALF;
-    }
-    half2 * KQ2 = (half2 *) KQ;
-
-    half kqmax[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -HALF_MAX_HALF;
-    }
-    half kqsum[ncols] = {0.0f};
-
-    __shared__ half kqmax_shared[ncols][WARP_SIZE];
-    __shared__ half kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-    __syncthreads();
-
-    // Convert Q to half2 and store in registers:
-    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-            const int i = i0 + threadIdx.x;
-
-            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
-            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
-        }
-    }
-
-    half2 VKQ[ncols] = {{0.0f, 0.0f}};
-
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
-        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
-        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
-        half kqmax_new = kqmax[0];
-        half kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-            half2 sum2[ncols] = {{0.0f, 0.0f}};
-#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;
-
-                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-#pragma unroll
-                for (int j = 0; j < ncols; ++j) {
-                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
-                }
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                sum2[j] = warp_reduce_sum(sum2[j]);
-                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
-                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
-
-                if (ncols == 1) {
-                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
-                } else {
-                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
-                }
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum;
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
-
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= __half2half2(KQ_max_scale);
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < D; k0 += 2) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
-                break;
-            }
-
-            half2 V_k;
-            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
-            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
-            }
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum(kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
-
-        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
-    }
-
-    if (parallel_blocks != 1 && tid != 0) {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
-        }
-    }
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
 #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
@@ -655,54 +413,6 @@ static __global__ void flash_attn_ext_f16(
 #endif // FP16_MMA_AVAILABLE
 }
 
-template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-static __global__ void flash_attn_combine_results(
-        const float  * __restrict__ VKQ_parts,
-        const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-#if FP16_AVAILABLE
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
-
-    const int tid = threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float2 meta[parallel_blocks];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
-    }
-
-    __syncthreads();
-
-    float kqmax = meta[0].x;
-#pragma unroll
-    for (int l = 1; l < parallel_blocks; ++l) {
-        kqmax = max(kqmax, meta[l].x);
-    }
-
-    float VKQ_numerator   = 0.0f;
-    float VKQ_denominator = 0.0f;
-#pragma unroll
-    for (int l = 0; l < parallel_blocks; ++l) {
-        const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
-        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
-        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
-
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
-        VKQ_denominator += KQ_max_scale * meta[l].y;
-    }
-
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 constexpr int get_max_power_of_2(int x) {
     return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
 }
@@ -727,66 +437,6 @@ static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 
-template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
-
-    const uint32_t n_head      = Q->ne[2];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
-
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale, max_bias, m0, m1, n_head_log2,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
 template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename KQ_acc_t> void launch_fattn_f16_impl(
         const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
         ggml_cuda_pool & pool, cudaStream_t main_stream
@@ -891,95 +541,22 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     const int32_t precision = KQV->op_params[2];
 
+    if (!fast_fp16_available(cc)) {
+        ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        return;
+    }
+
     if (!fp16_mma_available(cc)) {
-        GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-        GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
-
-        if (Q->ne[1] == 1) {
-            constexpr int cols_per_block = 1;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] == 2) {
-            constexpr int cols_per_block = 2;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] <= 4) {
-            constexpr int cols_per_block = 4;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        if (Q->ne[1] <= 8) {
-            constexpr int cols_per_block = 8;
-            constexpr int parallel_blocks = 4;
-            switch (Q->ne[0]) {
-                case 64:
-                    launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                case 128:
-                    launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                    break;
-                default:
-                    GGML_ASSERT(false);
-                    break;
-            }
-            return;
-        }
-
-        constexpr int cols_per_block = 8;
-        constexpr int parallel_blocks = 1;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
         return;
     }
 
     if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] == 1 && (Q->ne[0] == 64 || Q->ne[0] == 128)) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+
         if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
             constexpr int cols_per_block = 16;
             constexpr int nwarps         =  4;
@@ -1037,22 +614,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
-        constexpr int cols_per_block = 1;
-        constexpr int parallel_blocks = 4;
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 256:
-                launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         return;
     }
 

ggml-cuda/upscale.cu

@@ -1,35 +1,36 @@
 #include "upscale.cuh"
 
-static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) {
-    // blockIdx.z: idx of ne02*ne03
-    // blockIdx.y: idx of ne01*scale_factor， aka ne1
-    // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE
-    // ne00xne01: ne00 * ne01
-    int ne0 = ne00 * scale_factor;
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
+static __global__ void upscale_f32(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3) {
+    int index = threadIdx.x + blockIdx.x * blockDim.x;
+    if (index >= ne10 * ne11 * ne12 * ne13) {
         return;
     }
-    // operation
-    int i00 = nidx / scale_factor;
-    int i01 = blockIdx.y / scale_factor;
-    int offset_src =
-        i00 +
-        i01 * ne00 +
-        blockIdx.z * ne00xne01;
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    dst[offset_dst] = x[offset_src];
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
 
-static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03,
-                             const int scale_factor, cudaStream_t stream) {
-    int ne0 = (ne00 * scale_factor);
-    int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03);
-    upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
+static void upscale_f32_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        cudaStream_t stream) {
+    int dst_size = ne10 * ne11 * ne12 * ne13;
+    int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
 }
 
 void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -39,10 +40,12 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const int scale_factor = dst->op_params[0];
+    const float sf0 = (float)dst->ne[0]/src0->ne[0];
+    const float sf1 = (float)dst->ne[1]/src0->ne[1];
+    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
 
-    upscale_f32_cuda(src0_d, dst_d, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, stream);
+    upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
 }

ggml-impl.h

@@ -120,9 +120,16 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
 #endif
 
 // 16-bit float

ggml-metal.m

@@ -1378,7 +1378,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
                         if (ne00%4 == 0) {
-                            while (nth < ne00/4 && nth < 256) {
+                            while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
                                 nth *= 2;
                             }
                             if (use_f16) {
@@ -1387,7 +1387,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline;
                             }
                         } else {
-                            while (nth < ne00 && nth < 1024) {
+                            while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
                                 nth *= 2;
                             }
                             if (use_f16) {
@@ -2353,7 +2353,10 @@ static enum ggml_status ggml_metal_graph_compute(
                     {
                         GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
-                        const int sf = dst->op_params[0];
+                        const float sf0 = (float)ne0/src0->ne[0];
+                        const float sf1 = (float)ne1/src0->ne[1];
+                        const float sf2 = (float)ne2/src0->ne[2];
+                        const float sf3 = (float)ne3/src0->ne[3];
 
                         const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
 
@@ -2376,7 +2379,10 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
                         [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
                         [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-                        [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
+                        [encoder setBytes:&sf0  length:sizeof(sf0)  atIndex:18];
+                        [encoder setBytes:&sf1  length:sizeof(sf1)  atIndex:19];
+                        [encoder setBytes:&sf2  length:sizeof(sf2)  atIndex:20];
+                        [encoder setBytes:&sf3  length:sizeof(sf3)  atIndex:21];
 
                         const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
@@ -2512,13 +2518,14 @@ static enum ggml_status ggml_metal_graph_compute(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
-                        GGML_ASSERT(ne00 % 4 == 0);
+                        GGML_ASSERT(ne00 % 4  == 0);
+                        GGML_ASSERT(ne11 % 32 == 0);
+
                         GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
-                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
+                        GGML_ASSERT(ggml_are_same_shape (src1, src2));
 
-                        GGML_ASSERT(ggml_are_same_shape(src1, src2));
-                        GGML_ASSERT(src3);
+                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
 
                         size_t offs_src3 = 0;
 
@@ -2528,6 +2535,11 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
                                 "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
 
+                        const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
+                        const uint64_t nb21 = src2 ? src2->nb[1] : 0;
+                        const uint64_t nb22 = src2 ? src2->nb[2] : 0;
+                        const uint64_t nb23 = src2 ? src2->nb[3] : 0;
+
                         const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
                       //const int64_t  ne31 = src3 ? src3->ne[1] : 0;
                         const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
@@ -2590,34 +2602,35 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
                         [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
                         [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
-                        [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                        if (id_src3) {
+                            [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                        } else {
+                            [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
+                        }
                         [encoder setBuffer:id_dst      offset:offs_dst            atIndex:4];
-                        [encoder setBytes:&ne00        length:sizeof( int64_t)    atIndex:5];
-                        [encoder setBytes:&ne01        length:sizeof( int64_t)    atIndex:6];
-                        [encoder setBytes:&ne02        length:sizeof( int64_t)    atIndex:7];
-                        [encoder setBytes:&ne03        length:sizeof( int64_t)    atIndex:8];
-                        [encoder setBytes:&nb00        length:sizeof(uint64_t)    atIndex:9];
-                        [encoder setBytes:&nb01        length:sizeof(uint64_t)    atIndex:10];
-                        [encoder setBytes:&nb02        length:sizeof(uint64_t)    atIndex:11];
-                        [encoder setBytes:&nb03        length:sizeof(uint64_t)    atIndex:12];
-                        [encoder setBytes:&ne10        length:sizeof( int64_t)    atIndex:13];
-                        [encoder setBytes:&ne11        length:sizeof( int64_t)    atIndex:14];
-                        [encoder setBytes:&ne12        length:sizeof( int64_t)    atIndex:15];
-                        [encoder setBytes:&ne13        length:sizeof( int64_t)    atIndex:16];
-                        [encoder setBytes:&nb10        length:sizeof(uint64_t)    atIndex:17];
-                        [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:18];
-                        [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:19];
-                        [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:20];
-                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:21];
-                        [encoder setBytes:&ne0         length:sizeof( int64_t)    atIndex:22];
-                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:23];
-                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:24];
-                        [encoder setBytes:&ne3         length:sizeof( int64_t)    atIndex:25];
-                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:26];
-                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:27];
-                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:28];
-                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:29];
-                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:30];
+                        [encoder setBytes:&ne01        length:sizeof( int64_t)    atIndex:5];
+                        [encoder setBytes:&ne02        length:sizeof( int64_t)    atIndex:6];
+                        [encoder setBytes:&ne03        length:sizeof( int64_t)    atIndex:7];
+                        [encoder setBytes:&nb01        length:sizeof(uint64_t)    atIndex:8];
+                        [encoder setBytes:&nb02        length:sizeof(uint64_t)    atIndex:9];
+                        [encoder setBytes:&nb03        length:sizeof(uint64_t)    atIndex:10];
+                        [encoder setBytes:&ne11        length:sizeof( int64_t)    atIndex:11];
+                        [encoder setBytes:&ne12        length:sizeof( int64_t)    atIndex:12];
+                        [encoder setBytes:&ne13        length:sizeof( int64_t)    atIndex:13];
+                        [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:14];
+                        [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:15];
+                        [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:16];
+                        [encoder setBytes:&nb21        length:sizeof(uint64_t)    atIndex:17];
+                        [encoder setBytes:&nb22        length:sizeof(uint64_t)    atIndex:18];
+                        [encoder setBytes:&nb23        length:sizeof(uint64_t)    atIndex:19];
+                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:20];
+                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:21];
+                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:22];
+                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:23];
+                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:24];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:25];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:26];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
 
                         if (!use_vec_kernel) {
                             // half8x8 kernel

ggml-metal.metal

@@ -1852,7 +1852,10 @@ kernel void kernel_upscale_f32(
     constant  uint64_t & nb1,
     constant  uint64_t & nb2,
     constant  uint64_t & nb3,
-    constant   int32_t & sf,
+    constant     float & sf0,
+    constant     float & sf1,
+    constant     float & sf2,
+    constant     float & sf3,
     uint3 tgpig[[threadgroup_position_in_grid]],
     uint3 tpitg[[thread_position_in_threadgroup]],
     uint3   ntg[[threads_per_threadgroup]]) {
@@ -1861,15 +1864,17 @@ kernel void kernel_upscale_f32(
     const int64_t i2 = tgpig.y;
     const int64_t i1 = tgpig.x;
 
-    const int64_t i03 = i3;
-    const int64_t i02 = i2;
-    const int64_t i01 = i1/sf;
-
-    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
-    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+    const int64_t i03 = i3/sf3;
+    const int64_t i02 = i2/sf2;
+    const int64_t i01 = i1/sf1;
 
     for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        dst_ptr[i0] = src0_ptr[i0/sf];
+        const int64_t i00 = i0/sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1  +  i0*nb0);
+
+        dst_ptr[0] = src0_ptr[0];
     }
 }
 
@@ -2049,27 +2054,24 @@ typedef void (flash_attn_ext_f16_t)(
         device const  char * v,
         device const  char * mask,
         device       float * dst,
-        constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant   int64_t & ne03,
-        constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant  uint64_t & nb03,
-        constant   int64_t & ne10,
         constant   int64_t & ne11,
         constant   int64_t & ne12,
         constant   int64_t & ne13,
-        constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
         constant  uint64_t & nb31,
-        constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
-        constant   int64_t & ne3,
         constant     float & scale,
         constant     float & max_bias,
         constant     float & m0,
@@ -2090,27 +2092,24 @@ kernel void kernel_flash_attn_ext_f16(
         device const  char * v,
         device const  char * mask,
         device       float * dst,
-        constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant   int64_t & ne03,
-        constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant  uint64_t & nb03,
-        constant   int64_t & ne10,
         constant   int64_t & ne11,
         constant   int64_t & ne12,
         constant   int64_t & ne13,
-        constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
         constant  uint64_t & nb31,
-        constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
-        constant   int64_t & ne3,
         constant     float & scale,
         constant     float & max_bias,
         constant     float & m0,
@@ -2180,10 +2179,6 @@ kernel void kernel_flash_attn_ext_f16(
         const short ne22 = ne12;
         const short ne23 = ne13;
 
-        const uint nb21 = nb11;
-        const uint nb22 = nb12;
-        const uint nb23 = nb13;
-
         // broadcast
         const short rk2 = ne02/ne12;
         const short rk3 = ne03/ne13;
@@ -2247,11 +2242,16 @@ kernel void kernel_flash_attn_ext_f16(
                         simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
                     }
 
-                    // mqk = mqk*scale + mask*slope
-                    simdgroup_half8x8 mm;
-                    simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
-                    simdgroup_multiply(mm, mslope, mm);
-                    simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
+                    if (mask != q) {
+                        // mqk = mqk*scale + mask*slope
+                        simdgroup_half8x8 mm;
+                        simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
+                        simdgroup_multiply(mm, mslope, mm);
+                        simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
+                    } else {
+                        // mqk = mqk*scale
+                        simdgroup_multiply(mqk, mscale, mqk);
+                    }
 
                     simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
                 }
@@ -2425,27 +2425,24 @@ kernel void kernel_flash_attn_ext_vec_f16(
         device const  char * v,
         device const  char * mask,
         device       float * dst,
-        constant   int64_t & ne00,
         constant   int64_t & ne01,
         constant   int64_t & ne02,
         constant   int64_t & ne03,
-        constant  uint64_t & nb00,
         constant  uint64_t & nb01,
         constant  uint64_t & nb02,
         constant  uint64_t & nb03,
-        constant   int64_t & ne10,
         constant   int64_t & ne11,
         constant   int64_t & ne12,
         constant   int64_t & ne13,
-        constant  uint64_t & nb10,
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
         constant  uint64_t & nb31,
-        constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
-        constant   int64_t & ne3,
         constant     float & scale,
         constant     float & max_bias,
         constant     float & m0,
@@ -2521,10 +2518,6 @@ kernel void kernel_flash_attn_ext_vec_f16(
         const short ne22 = ne12;
         const short ne23 = ne13;
 
-        const uint nb21 = nb11;
-        const uint nb22 = nb12;
-        const uint nb23 = nb13;
-
         // broadcast
         const short rk2 = ne02/ne12;
         const short rk3 = ne03/ne13;
@@ -2589,8 +2582,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
 
                     // mqk = mqk*scale + mask*slope
                     if (tiisg == 0) {
-                        float4 mm = (float4) mp4[ic/4 + cc];
-                        mqk = mqk*scale + mm*slope;
+                        mqk = mqk*scale + ((mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f);
 
                         ss4[cc] = mqk;
                     }

ggml-rpc.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_RPC_MAX_SERVERS       16
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+
+GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+
+GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-sycl.cpp

@@ -13987,6 +13987,10 @@ inline void ggml_sycl_op_upscale(const ggml_tensor *src0,
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
     GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
 
+#pragma message("TODO: generalize upscale operator")
+#pragma message("      https://github.com/ggerganov/ggml/pull/814")
+    GGML_ASSERT(false && "TODO: generalize upscale operator");
+
     const int scale_factor = dst->op_params[0];
 
     upscale_f32_sycl(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
@@ -15564,26 +15568,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-#if 0
-    // use syclGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                SYCL_CHECK(
-                        syclGemmEx(g_sycl_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , SYCL_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, SYCL_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
     if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
@@ -15595,7 +15579,6 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
             nb11 / nb10, nb12 / nb10, beta,
             (char *)dst_t, cu_data_type, ne01, nb2 / nb0,
             ne12 * ne13, cu_compute_type)));
-        g_sycl_handles[g_main_device]->wait();
     } else {
         const int ne23 = ne12*ne13;
 
@@ -15626,7 +15609,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
                                          nb02, nb03, nb12_scaled, nb13_scaled,
                                          nbd2, nbd3, r2, r3, item_ct1);
                                  });
-            }).wait();
+            });
         }
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
             *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans,
@@ -15637,9 +15620,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0,
             dpct::library_data_t::real_half, nb11 / nb10, beta,
             (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23,
             cu_compute_type)));
-        g_sycl_handles[g_main_device]->wait();
     }
-#endif
 
     if (no_mixed_dtypes) {
         const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);

ggml.h

@@ -565,7 +565,8 @@ extern "C" {
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type         type;
-        enum ggml_backend_type backend;
+
+        GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
 
         struct ggml_backend_buffer * buffer;
 
@@ -766,7 +767,8 @@ extern "C" {
     GGML_API           bool ggml_is_3d        (const struct ggml_tensor * tensor);
     GGML_API           int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
-    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
@@ -1673,12 +1675,24 @@ extern "C" {
             float                 p1);
 
     // nearest interpolate
+    // multiplies ne0 and ne1 by scale factor
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // nearest interpolate
+    // nearest interpolate to specified dimensions
+    // used in tortoise.cpp
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3);
+
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
             struct ggml_context * ctx,

gguf-py/gguf/__init__.py

@@ -2,5 +2,6 @@ from .constants import *
 from .lazy import *
 from .gguf_reader import *
 from .gguf_writer import *
+from .quants import *
 from .tensor_mapping import *
 from .vocab import *

gguf-py/gguf/gguf_writer.py

@@ -13,6 +13,7 @@ from string import ascii_letters, digits
 import numpy as np
 
 from .constants import (
+    GGML_QUANT_SIZES,
     GGUF_DEFAULT_ALIGNMENT,
     GGUF_MAGIC,
     GGUF_VERSION,
@@ -195,7 +196,7 @@ class GGUFWriter:
         return ((x + n - 1) // n) * n
 
     def add_tensor_info(
-        self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32],
+        self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype,
         tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None,
     ) -> None:
         if self.state is not WriterState.EMPTY:
@@ -208,10 +209,6 @@ class GGUFWriter:
         encoded_name = name.encode("utf-8")
         self.ti_data += self._pack("Q", len(encoded_name))
         self.ti_data += encoded_name
-        n_dims = len(tensor_shape)
-        self.ti_data += self._pack("I", n_dims)
-        for i in range(n_dims):
-            self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
             if tensor_dtype == np.float16:
                 dtype = GGMLQuantizationType.F16
@@ -231,6 +228,15 @@ class GGUFWriter:
                 raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
         else:
             dtype = raw_dtype
+            if tensor_dtype == np.uint8:
+                block_size, type_size = GGML_QUANT_SIZES[raw_dtype]
+                if tensor_shape[-1] % type_size != 0:
+                    raise ValueError(f"Quantized tensor row size ({tensor_shape[-1]}) is not a multiple of {dtype.name} type size ({type_size})")
+                tensor_shape = tuple(tensor_shape[:-1]) + (tensor_shape[-1] // type_size * block_size,)
+        n_dims = len(tensor_shape)
+        self.ti_data += self._pack("I", n_dims)
+        for i in range(n_dims):
+            self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         self.ti_data += self._pack("I", dtype)
         self.ti_data += self._pack("Q", self.offset_tensor)
         self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)

gguf-py/gguf/lazy.py

@@ -6,6 +6,7 @@ from typing import Any, Callable
 from collections import deque
 
 import numpy as np
+from numpy._typing import _Shape
 from numpy.typing import DTypeLike
 
 
@@ -110,7 +111,7 @@ class LazyBase(ABC, metaclass=LazyMeta):
             return o
 
     @classmethod
-    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike = False) -> Callable[[Any], Any]:
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
         def wrapped_fn(*args, **kwargs):
             if kwargs is None:
                 kwargs = {}
@@ -130,9 +131,14 @@ class LazyBase(ABC, metaclass=LazyMeta):
                 res = args[0]
                 assert isinstance(res, cls)
                 res = res._meta
-                # allow operations to override the dtype
+                # allow operations to override the dtype and shape
                 if meta_noop is not True:
-                    res = cls.meta_with_dtype(res, meta_noop)
+                    if isinstance(meta_noop, tuple):
+                        dtype, shape = meta_noop
+                        assert callable(shape)
+                        res = cls.meta_with_dtype_and_shape(dtype, shape(res.shape))
+                    else:
+                        res = cls.meta_with_dtype_and_shape(meta_noop, res.shape)
 
             if isinstance(res, cls._tensor_type):
                 def collect_replace(t: LazyBase):
@@ -168,7 +174,12 @@ class LazyBase(ABC, metaclass=LazyMeta):
             while _t._data is None:
                 lt = _t._lazy.popleft()
                 if lt._data is not None:
-                    raise ValueError(f"{lt} did not belong in the lazy queue")
+                    # Lazy tensor did not belong in the lazy queue.
+                    # Weirdly only happens with Bloom models...
+                    # likely because tensors aren't unique in the queue.
+                    # The final output is still the same as in eager mode,
+                    # so it's safe to ignore this.
+                    continue
                 assert lt._func is not None
                 lt._args = cls._recurse_apply(lt._args, already_eager_to_eager)
                 lt._data = lt._func(lt._args)
@@ -183,12 +194,12 @@ class LazyBase(ABC, metaclass=LazyMeta):
 
     @classmethod
     def eager_to_meta(cls, t: Any) -> Any:
-        return cls.meta_with_dtype(t, t.dtype)
+        return cls.meta_with_dtype_and_shape(t.dtype, t.shape)
 
     # must be overridden, meta tensor init is backend-specific
     @classmethod
     @abstractmethod
-    def meta_with_dtype(cls, m: Any, dtype: Any) -> Any: pass
+    def meta_with_dtype_and_shape(cls, dtype: Any, shape: Any) -> Any: pass
 
     @classmethod
     def from_eager(cls, t: Any) -> Any:
@@ -205,15 +216,15 @@ class LazyNumpyTensor(LazyBase):
     _tensor_type = np.ndarray
 
     @classmethod
-    def meta_with_dtype(cls, m: np.ndarray[Any, Any], dtype: DTypeLike) -> np.ndarray[Any, Any]:
+    def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: _Shape) -> np.ndarray[Any, Any]:
         # The initial idea was to use np.nan as the fill value,
         # but non-float types like np.int16 can't use that.
         # So zero it is.
         cheat = np.zeros(1, dtype)
-        return np.lib.stride_tricks.as_strided(cheat, m.shape, (0 for _ in m.shape))
+        return np.lib.stride_tricks.as_strided(cheat, shape, (0 for _ in shape))
 
     def astype(self, dtype, *args, **kwargs):
-        meta = type(self).meta_with_dtype(self._meta, dtype)
+        meta = type(self).meta_with_dtype_and_shape(dtype, self._meta.shape)
         full_args = (self, dtype,) + args
         # very important to pass the shared _lazy deque, or else there's an infinite loop somewhere.
         return type(self)(meta=meta, args=full_args, lazy=self._lazy, func=(lambda a: a[0].astype(*a[1:], **kwargs)))

gguf-py/gguf/quants.py

@@ -0,0 +1,109 @@
+from __future__ import annotations
+from typing import Callable
+
+from numpy.typing import DTypeLike
+
+from .constants import GGML_QUANT_SIZES, GGMLQuantizationType
+from .lazy import LazyNumpyTensor
+
+import numpy as np
+
+
+# same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
+    n = n.astype(np.float32, copy=False).view(np.int32)
+    # force nan to quiet
+    n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
+    # flush subnormals to zero
+    n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+    # round to nearest even
+    n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
+    return n.astype(np.int16)
+
+
+# This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
+def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
+    rows = arr.reshape((-1, arr.shape[-1]))
+    osize = 1
+    for dim in oshape:
+        osize *= dim
+    out = np.empty(shape=osize, dtype=otype)
+    # compute over groups of 16 rows (arbitrary, but seems good for performance)
+    n_groups = rows.shape[0] // 16
+    np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
+    return out.reshape(oshape)
+
+
+def __quantize_bf16_array(n: np.ndarray) -> np.ndarray:
+    return __apply_over_grouped_rows(__compute_fp32_to_bf16, arr=n, otype=np.int16, oshape=n.shape)
+
+
+__quantize_bf16_lazy = LazyNumpyTensor._wrap_fn(__quantize_bf16_array, meta_noop=np.int16)
+
+
+def quantize_bf16(n: np.ndarray):
+    if type(n) is LazyNumpyTensor:
+        return __quantize_bf16_lazy(n)
+    else:
+        return __quantize_bf16_array(n)
+
+
+__q8_block_size, __q8_type_size = GGML_QUANT_SIZES[GGMLQuantizationType.Q8_0]
+
+
+def can_quantize_to_q8_0(n: np.ndarray) -> bool:
+    return n.shape[-1] % __q8_block_size == 0
+
+
+# round away from zero
+# ref: https://stackoverflow.com/a/59143326/22827863
+def np_roundf(n: np.ndarray) -> np.ndarray:
+    a = abs(n)
+    floored = np.floor(a)
+    b = floored + np.floor(2 * (a - floored))
+    return np.sign(n) * b
+
+
+def __quantize_q8_0_shape_change(s: tuple[int, ...]) -> tuple[int, ...]:
+    return (*s[:-1], s[-1] // __q8_block_size * __q8_type_size)
+
+
+# Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
+def __quantize_q8_0_rows(n: np.ndarray) -> np.ndarray:
+    shape = n.shape
+    assert shape[-1] % __q8_block_size == 0
+
+    n_blocks = n.size // __q8_block_size
+
+    blocks = n.reshape((n_blocks, __q8_block_size)).astype(np.float32, copy=False)
+
+    d = abs(blocks).max(axis=1, keepdims=True) / 127
+    with np.errstate(divide="ignore"):
+        id = np.where(d == 0, 0, 1 / d)
+    qs = np_roundf(blocks * id)
+
+    # (n_blocks, 2)
+    d = d.astype(np.float16).view(np.uint8)
+    # (n_blocks, block_size)
+    qs = qs.astype(np.int8).view(np.uint8)
+
+    assert d.shape[1] + qs.shape[1] == __q8_type_size
+
+    return np.concatenate([d, qs], axis=1).reshape(__quantize_q8_0_shape_change(shape))
+
+
+def __quantize_q8_0_array(n: np.ndarray) -> np.ndarray:
+    return __apply_over_grouped_rows(__quantize_q8_0_rows, arr=n, otype=np.uint8, oshape=__quantize_q8_0_shape_change(n.shape))
+
+
+__quantize_q8_0_lazy = LazyNumpyTensor._wrap_fn(
+    __quantize_q8_0_array,
+    meta_noop=(np.uint8, __quantize_q8_0_shape_change),
+)
+
+
+def quantize_q8_0(data: np.ndarray):
+    if type(data) is LazyNumpyTensor:
+        return __quantize_q8_0_lazy(data)
+    else:
+        return __quantize_q8_0_array(data)

llama.cpp

@@ -7,6 +7,10 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
+#ifdef GGML_USE_RPC
+#  include "ggml-rpc.h"
+#endif
+
 #ifdef GGML_USE_CUDA
 #  include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
@@ -1685,91 +1689,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer
     GGML_UNUSED(host_buffer);
 }
 
-static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
-    ggml_backend_buffer_type_t buft = nullptr;
-
-#ifdef GGML_USE_METAL
-    buft = ggml_backend_metal_buffer_type();
-#elif defined(GGML_USE_CUDA)
-    buft = ggml_backend_cuda_buffer_type(gpu);
-#elif defined(GGML_USE_VULKAN)
-    buft = ggml_backend_vk_buffer_type(gpu);
-#elif defined(GGML_USE_SYCL)
-    buft = ggml_backend_sycl_buffer_type(gpu);
-#elif defined(GGML_USE_CLBLAST)
-    buft = ggml_backend_opencl_buffer_type();
-#elif defined(GGML_USE_KOMPUTE)
-    buft = ggml_backend_kompute_buffer_type(gpu);
-    if (buft == nullptr) {
-        LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
-    }
-#endif
-
-    if (buft == nullptr) {
-        buft = llama_default_buffer_type_cpu(true);
-    }
-    return buft;
-
-    GGML_UNUSED(gpu);
-}
-
-static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_gpu, const float * tensor_split) {
-    ggml_backend_buffer_type_t buft = nullptr;
-
-#ifdef GGML_USE_CUDA
-    if (ggml_backend_cuda_get_device_count() > 1) {
-        buft = ggml_backend_cuda_split_buffer_type(tensor_split);
-    }
-#endif
-
-#ifdef GGML_USE_SYCL
-    if (ggml_backend_sycl_get_device_count() > 1) {
-        buft = ggml_backend_sycl_split_buffer_type(tensor_split);
-    }
-#endif
-
-    if (buft == nullptr) {
-        buft = llama_default_buffer_type_offload(fallback_gpu);
-    }
-    return buft;
-
-    GGML_UNUSED(tensor_split);
-}
-
-static size_t llama_get_device_count() {
-#if defined(GGML_USE_CUDA)
-    return ggml_backend_cuda_get_device_count();
-#elif defined(GGML_USE_SYCL)
-    return ggml_backend_sycl_get_device_count();
-#elif defined(GGML_USE_VULKAN)
-    return ggml_backend_vk_get_device_count();
-#else
-    return 1;
-#endif
-}
-
-static size_t llama_get_device_memory(int device) {
-#if defined(GGML_USE_CUDA)
-    size_t total;
-    size_t free;
-    ggml_backend_cuda_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_SYCL)
-    size_t total;
-    size_t free;
-    ggml_backend_sycl_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_VULKAN)
-    size_t total;
-    size_t free;
-    ggml_backend_vk_get_device_memory(device, &free, &total);
-    return free;
-#else
-    return 1;
-    GGML_UNUSED(device);
-#endif
-}
-
 //
 // globals
 //
@@ -2210,6 +2129,8 @@ struct llama_model {
     int main_gpu;
     int n_gpu_layers;
 
+    std::vector<std::string> rpc_servers;
+
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
@@ -2353,6 +2274,104 @@ struct llama_context {
 #endif
 };
 
+static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
+    ggml_backend_buffer_type_t buft = nullptr;
+
+#ifdef GGML_USE_RPC
+    std::string endpoint = model.rpc_servers[gpu];
+    buft = ggml_backend_rpc_buffer_type(endpoint.c_str());
+#elif defined(GGML_USE_METAL)
+    buft = ggml_backend_metal_buffer_type();
+#elif defined(GGML_USE_CUDA)
+    buft = ggml_backend_cuda_buffer_type(gpu);
+#elif defined(GGML_USE_VULKAN)
+    buft = ggml_backend_vk_buffer_type(gpu);
+#elif defined(GGML_USE_SYCL)
+    buft = ggml_backend_sycl_buffer_type(gpu);
+#elif defined(GGML_USE_CLBLAST)
+    buft = ggml_backend_opencl_buffer_type();
+#elif defined(GGML_USE_KOMPUTE)
+    buft = ggml_backend_kompute_buffer_type(gpu);
+    if (buft == nullptr) {
+        LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
+    }
+#endif
+
+    if (buft == nullptr) {
+        buft = llama_default_buffer_type_cpu(true);
+    }
+    return buft;
+    GGML_UNUSED(model);
+    GGML_UNUSED(gpu);
+}
+
+static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_model & model, int fallback_gpu, const float * tensor_split) {
+    ggml_backend_buffer_type_t buft = nullptr;
+
+#ifdef GGML_USE_CUDA
+    if (ggml_backend_cuda_get_device_count() > 1) {
+        buft = ggml_backend_cuda_split_buffer_type(tensor_split);
+    }
+#endif
+
+#ifdef GGML_USE_SYCL
+    if (ggml_backend_sycl_get_device_count() > 1) {
+        buft = ggml_backend_sycl_split_buffer_type(tensor_split);
+    }
+#endif
+
+    if (buft == nullptr) {
+        buft = llama_default_buffer_type_offload(model, fallback_gpu);
+    }
+    return buft;
+
+    GGML_UNUSED(tensor_split);
+}
+
+static size_t llama_get_device_count(const llama_model & model) {
+#if defined(GGML_USE_RPC)
+    return model.rpc_servers.size();
+#elif defined(GGML_USE_CUDA)
+    return ggml_backend_cuda_get_device_count();
+#elif defined(GGML_USE_SYCL)
+    return ggml_backend_sycl_get_device_count();
+#elif defined(GGML_USE_VULKAN)
+    return ggml_backend_vk_get_device_count();
+#else
+    return 1;
+#endif
+    GGML_UNUSED(model);
+}
+
+static size_t llama_get_device_memory(const llama_model & model, int device) {
+#if defined(GGML_USE_RPC)
+    size_t total;
+    size_t free;
+    std::string endpoint = model.rpc_servers[device];
+    ggml_backend_rpc_get_device_memory(endpoint.c_str(), &free, &total);
+    return free;
+#elif defined(GGML_USE_CUDA)
+    size_t total;
+    size_t free;
+    ggml_backend_cuda_get_device_memory(device, &free, &total);
+    return free;
+#elif defined(GGML_USE_SYCL)
+    size_t total;
+    size_t free;
+    ggml_backend_sycl_get_device_memory(device, &free, &total);
+    return free;
+#elif defined(GGML_USE_VULKAN)
+    size_t total;
+    size_t free;
+    ggml_backend_vk_get_device_memory(device, &free, &total);
+    return free;
+#else
+    return 1;
+#endif
+    GGML_UNUSED(model);
+    GGML_UNUSED(device);
+}
+
 //
 // kv cache helpers
 //
@@ -2805,6 +2824,11 @@ static void llama_kv_cache_defrag(struct llama_kv_cache & cache) {
     cache.do_defrag = true;
 }
 
+static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) {
+    // the FA kernels require padding to avoid extra runtime boundary checks
+    return cparams.flash_attn ? 256u : 32u;
+}
+
 //
 // model loading and saving
 //
@@ -4424,7 +4448,9 @@ static void llm_load_vocab(
             } else if (
                     tokenizer_pre == "gpt-2"   ||
                     tokenizer_pre == "jina-es" ||
-                    tokenizer_pre == "jina-de") {
+                    tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "jina-v2-es" ||
+                    tokenizer_pre == "jina-v2-de") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "refact") {
@@ -4784,13 +4810,13 @@ static bool llm_load_tensors(
 
     if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
         // calculate the split points
-        int device_count = llama_get_device_count();
+        int device_count = llama_get_device_count(model);
         bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + device_count, [](float x) { return x == 0.0f; });
         std::vector<float> splits(device_count);
         if (all_zero) {
             // default split, by free memory
             for (int i = 0; i < device_count; ++i) {
-                splits[i] = llama_get_device_memory(i);
+                splits[i] = llama_get_device_memory(model, i);
             }
         } else {
             std::copy(tensor_split, tensor_split + device_count, splits.begin());
@@ -4810,35 +4836,35 @@ static bool llm_load_tensors(
         int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1);
         for (int64_t i = i_gpu_start; i < n_layer; ++i) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin();
-            model.buft_layer[i] = llama_default_buffer_type_offload(layer_gpu);
+            model.buft_layer[i] = llama_default_buffer_type_offload(model, layer_gpu);
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin();
-            model.buft_output = llama_default_buffer_type_offload(layer_gpu);
+            model.buft_output = llama_default_buffer_type_offload(model, layer_gpu);
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
         }
     } else {
         ggml_backend_buffer_type_t split_buft;
         if (split_mode == LLAMA_SPLIT_MODE_ROW) {
-            split_buft = llama_default_buffer_type_split(main_gpu, tensor_split);
+            split_buft = llama_default_buffer_type_split(model, main_gpu, tensor_split);
         } else {
             // LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_LAYER in backends where it is not supported
-            split_buft = llama_default_buffer_type_offload(main_gpu);
+            split_buft = llama_default_buffer_type_offload(model, main_gpu);
         }
         // assign the repeating layers
         for (int64_t i = i_gpu_start; i < n_layer; ++i) {
             model.buft_layer[i] = {
                 split_buft,
-                llama_default_buffer_type_offload(main_gpu)
+                llama_default_buffer_type_offload(model, main_gpu)
             };
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             model.buft_output = {
                 split_buft,
-                llama_default_buffer_type_offload(main_gpu)
+                llama_default_buffer_type_offload(model, main_gpu)
             };
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
@@ -11508,7 +11534,8 @@ static int llama_decode_internal(
                 // a heuristic, to avoid attending the full cache if it is not yet utilized
                 // after enough generations, the benefit from this heuristic disappears
                 // if we start defragmenting the cache, the benefit from this will be more important
-                kv_self.n = std::min(kv_self.size, std::max(256u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 256)));
+                const uint32_t pad = llama_kv_cache_get_padding(cparams);
+                kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
                 //kv_self.n = llama_kv_cache_cell_max(kv_self);
             }
         }
@@ -13174,6 +13201,58 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
     return rejects;
 }
 
+static bool llama_grammar_detect_left_recursion(
+        const std::vector<std::vector<llama_grammar_element>> & rules,
+        size_t                                                  rule_index,
+        std::vector<bool>                                     * rules_visited,
+        std::vector<bool>                                     * rules_in_progress,
+        std::vector<bool>                                     * rules_may_be_empty) {
+    if ((*rules_in_progress)[rule_index]) {
+        return true;
+    }
+
+    (*rules_in_progress)[rule_index] = true;
+
+    const std::vector<llama_grammar_element> & rule = rules[rule_index];
+
+    // First check if the rule might produce the empty string. This could be done combined with the second
+    // step but it's more readable as two steps.
+    bool at_rule_start = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            if (at_rule_start) {
+                (*rules_may_be_empty)[rule_index] = true;
+                break;
+            }
+            at_rule_start = true;
+        } else {
+            at_rule_start = false;
+        }
+    }
+
+    // Second, recurse into leftmost nonterminals (or next-leftmost as long as the previous nonterminal may
+    // be empty)
+    bool recurse_into_nonterminal = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (rule[i].type == LLAMA_GRETYPE_RULE_REF && recurse_into_nonterminal) {
+            if (llama_grammar_detect_left_recursion(rules, (size_t)rule[i].value, rules_visited, rules_in_progress, rules_may_be_empty)) {
+                return true;
+            }
+            if (!((*rules_may_be_empty)[(size_t)rule[i].value])) {
+                recurse_into_nonterminal = false;
+            }
+        } else if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            recurse_into_nonterminal = true;
+        } else {
+            recurse_into_nonterminal = false;
+        }
+    }
+
+    (*rules_in_progress)[rule_index] = false;
+    (*rules_visited)[rule_index] = true;
+    return false;
+}
+
 //
 // grammar - external
 //
@@ -13193,6 +13272,19 @@ struct llama_grammar * llama_grammar_init(
         vec_rules[i].push_back({LLAMA_GRETYPE_END, 0});
     }
 
+    // Check for left recursion
+    std::vector<bool> rules_visited(n_rules);
+    std::vector<bool> rules_in_progress(n_rules);
+    std::vector<bool> rules_may_be_empty(n_rules);
+    for (size_t i = 0; i < n_rules; i++) {
+        if (rules_visited[i]) {
+            continue;
+        }
+        if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
+            throw std::runtime_error(format("unsupported grammar, left recursion detected for nonterminal at index %zu", i));
+        }
+    }
+
     // loop over alternates of start rule to build initial stacks
     std::vector<std::vector<const llama_grammar_element *>> stacks;
     pos = vec_rules[start_rule_index].data();
@@ -13215,6 +13307,9 @@ struct llama_grammar * llama_grammar_init(
         }
     } while (true);
 
+    // Important: vec_rules has to be moved here, not copied, because stacks contains
+    // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
+    // then the pointers would be invalidated when the local vec_rules goes out of scope.
     return new llama_grammar{ std::move(vec_rules), std::move(stacks), {} };
 }
 
@@ -15314,6 +15409,7 @@ struct llama_model_params llama_model_default_params() {
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
+        /*.rpc_servers                 =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
@@ -15384,7 +15480,9 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
 }
 
 size_t llama_max_devices(void) {
-#if defined(GGML_USE_METAL)
+#if defined(GGML_USE_RPC)
+    return GGML_RPC_MAX_SERVERS;
+#elif defined(GGML_USE_METAL)
     return 1;
 #elif defined(GGML_USE_CUDA)
     return GGML_CUDA_MAX_DEVICES;
@@ -15407,7 +15505,7 @@ bool llama_supports_mlock(void) {
 
 bool llama_supports_gpu_offload(void) {
 #if defined(GGML_USE_CUDA) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
-    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
+    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
     // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
     return true;
 #else
@@ -15470,7 +15568,17 @@ struct llama_model * llama_load_model_from_file(
             return true;
         };
     }
-
+    if (params.rpc_servers != nullptr) {
+        // split the servers set them into model->rpc_servers
+        std::string servers(params.rpc_servers);
+        size_t pos = 0;
+        while ((pos = servers.find(",")) != std::string::npos) {
+            std::string server = servers.substr(0, pos);
+            model->rpc_servers.push_back(server);
+            servers.erase(0, pos + 1);
+        }
+        model->rpc_servers.push_back(servers);
+    }
     int status = llama_model_load(path_model, *model, params);
     GGML_ASSERT(status <= 0);
     if (status < 0) {
@@ -15509,6 +15617,11 @@ struct llama_context * llama_new_context_with_model(
         return nullptr;
     }
 
+    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
     llama_context * ctx = new llama_context(*model);
 
     const auto & hparams = model->hparams;
@@ -15532,7 +15645,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
 
     // this is necessary due to kv_self.n being padded later during inference
-    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, 256);
+    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams));
 
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -15577,11 +15690,6 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-    if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) {
-        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        cparams.flash_attn = false;
-    }
-
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
@@ -15617,7 +15725,17 @@ struct llama_context * llama_new_context_with_model(
 
     if (!hparams.vocab_only) {
         // initialize backends
-#ifdef GGML_USE_METAL
+#if defined(GGML_USE_RPC)
+        for (auto & server : model->rpc_servers) {
+            ggml_backend_t backend = ggml_backend_rpc_init(server.c_str());
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to connect RPC backend to %s\n", __func__, server.c_str());
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_METAL)
         if (model->n_gpu_layers > 0) {
             ctx->backend_metal = ggml_backend_metal_init();
             if (ctx->backend_metal == nullptr) {
@@ -15773,7 +15891,11 @@ struct llama_context * llama_new_context_with_model(
             ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead_custom(LLAMA_MAX_NODES, false));
 
             // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
-            bool pipeline_parallel = llama_get_device_count() > 1 && model->n_gpu_layers > (int)model->hparams.n_layer && model->split_mode == LLAMA_SPLIT_MODE_LAYER;
+            bool pipeline_parallel =
+                llama_get_device_count(*model) > 1 &&
+                model->n_gpu_layers > (int)model->hparams.n_layer &&
+                model->split_mode == LLAMA_SPLIT_MODE_LAYER &&
+                params.offload_kqv;
 #ifndef GGML_USE_CUDA
             // pipeline parallelism requires support for async compute and events
             // currently this is only implemented in the CUDA backend
@@ -16893,13 +17015,13 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam
             }
             else {
                 if (cell_range_begin != kv_self.size) {
-                    cell_ranges.push_back({ cell_range_begin, i });
+                    cell_ranges.emplace_back(cell_range_begin, i);
                     cell_range_begin = kv_self.size;
                 }
             }
         }
         if (cell_range_begin != kv_self.size) {
-            cell_ranges.push_back({ cell_range_begin, kv_self.size });
+            cell_ranges.emplace_back(cell_range_begin, kv_self.size);
         }
 
         // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count

llama.h

@@ -242,6 +242,9 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
+        // comma separated list of RPC servers to use for offloading
+        const char * rpc_servers;
+
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
         // If it returns false, model loading is immediately aborted.

requirements.txt

@@ -9,5 +9,4 @@
 -r ./requirements/requirements-convert-hf-to-gguf.txt
 -r ./requirements/requirements-convert-hf-to-gguf-update.txt
 -r ./requirements/requirements-convert-llama-ggml-to-gguf.txt
--r ./requirements/requirements-convert-lora-to-ggml.txt
 -r ./requirements/requirements-convert-persimmon-to-gguf.txt

requirements/requirements-convert-lora-to-ggml.txt

@@ -1,2 +0,0 @@
--r ./requirements-convert.txt
-torch~=2.1.1

scripts/sync-ggml-am.sh

@@ -112,6 +112,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-opencl.h           -> ggml-opencl.h
     # src/ggml-quants.c           -> ggml-quants.c
     # src/ggml-quants.h           -> ggml-quants.h
+    # src/ggml-rpc.cpp            -> ggml-rpc.cpp
+    # src/ggml-rpc.h              -> ggml-rpc.h
     # src/ggml-sycl.cpp           -> ggml-sycl.cpp
     # src/ggml-sycl.h             -> ggml-sycl.h
     # src/ggml-vulkan.cpp         -> ggml-vulkan.cpp
@@ -149,6 +151,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
         -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \
         -e 's/src\/ggml-quants\.h/ggml-quants.h/g' \
+        -e 's/src\/ggml-rpc\.cpp/ggml-rpc.cpp/g' \
+        -e 's/src\/ggml-rpc\.h/ggml-rpc.h/g' \
         -e 's/src\/ggml-sycl\.cpp/ggml-sycl.cpp/g' \
         -e 's/src\/ggml-sycl\.h/ggml-sycl.h/g' \
         -e 's/src\/ggml-vulkan\.cpp/ggml-vulkan.cpp/g' \

scripts/sync-ggml.last

@@ -1 +1 @@
-30f54cbb3ada3e4c5bc6924de3e5918e5be4ff11
+126d34985705a5a2222723c145cb4e125ac689f3

scripts/sync-ggml.sh

@@ -20,6 +20,8 @@ cp -rpv ../ggml/src/ggml-opencl.cpp         ./ggml-opencl.cpp
 cp -rpv ../ggml/src/ggml-opencl.h           ./ggml-opencl.h
 cp -rpv ../ggml/src/ggml-quants.c           ./ggml-quants.c
 cp -rpv ../ggml/src/ggml-quants.h           ./ggml-quants.h
+cp -rpv ../ggml/src/ggml-rpc.cpp            ./ggml-rpc.cpp
+cp -rpv ../ggml/src/ggml-rpc.h              ./ggml-rpc.h
 cp -rpv ../ggml/src/ggml-sycl.cpp           ./ggml-sycl.cpp
 cp -rpv ../ggml/src/ggml-sycl.h             ./ggml-sycl.h
 cp -rpv ../ggml/src/ggml-vulkan.cpp         ./ggml-vulkan.cpp

tests/test-backend-ops.cpp

@@ -2,6 +2,7 @@
 #include <ggml-alloc.h>
 #include <ggml-backend.h>
 #include <ggml-backend-impl.h>
+
 #include <algorithm>
 #include <array>
 #include <cfloat>
@@ -1328,23 +1329,47 @@ struct test_upscale : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const int32_t scale_factor;
+    const bool transpose;
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, scale_factor);
+        return VARS_TO_STR4(type, ne, scale_factor, transpose);
     }
 
     test_upscale(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {512, 512, 3, 1},
-            int32_t scale_factor = 2)
-        : type(type), ne(ne), scale_factor(scale_factor) {}
+            int32_t scale_factor = 2, bool transpose = false)
+        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        if (transpose) a = ggml_transpose(ctx, a);
         ggml_tensor * out = ggml_upscale(ctx, a, scale_factor);
         return out;
     }
 };
 
+// GGML_OP_UPSCALE (ext)
+struct test_upscale_ext : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const std::array<int64_t, 4> ne_tgt;
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, ne_tgt);
+    }
+
+    test_upscale_ext(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne     = {2, 5,  7, 11},
+            std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13})
+        : type(type), ne(ne), ne_tgt(ne_tgt) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3]);
+        return out;
+    }
+};
+
 // GGML_OP_GROUP_NORM
 struct test_group_norm : public test_case {
     const ggml_type type;
@@ -1486,25 +1511,27 @@ struct test_flash_attn_ext : public test_case {
     const int64_t kv; // kv size
     const int64_t nb; // batch size
 
+    const bool mask; // use mask
+
     const float max_bias; // ALiBi
 
     std::string vars() override {
-        return VARS_TO_STR5(hs, nh, kv, nb, max_bias);
+        return VARS_TO_STR6(hs, nh, kv, nb, mask, max_bias);
     }
 
     double max_nmse_err() override {
         return 5e-4;
     }
 
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, float max_bias = 0.0f)
-        : hs(hs), nh(nh), kv(kv), nb(nb), max_bias(max_bias) {}
+    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, bool mask = true, float max_bias = 0.0f)
+        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
         ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
         ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
-        ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1);
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, mask, 1.0f/sqrtf(hs), max_bias);
+        ggml_tensor * m = mask ? ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1) : nullptr;
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias);
         return out;
     }
 };
@@ -2166,6 +2193,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_upscale());
+    test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
+    test_cases.emplace_back(new test_upscale_ext());
     test_cases.emplace_back(new test_group_norm());
     test_cases.emplace_back(new test_acc());
     test_cases.emplace_back(new test_pad());
@@ -2173,16 +2202,15 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    for (int hs : { 64, 128, }) { // other head sizes not implemented
-#else
     for (int hs : { 64, 80, 128, 256, }) {
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-        for (float max_bias : {0.0f, 8.0f}) {
-            for (int nh : { 32, }) {
-                for (int kv : { 512, 1024, }) {
-                    for (int nb : { 1, 2, 4, 8, }) {
-                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, max_bias));
+        for (bool mask : { true, false } ) {
+            for (float max_bias : { 0.0f, 8.0f }) {
+                if (!mask && max_bias > 0.0f) continue;
+                for (int nh : { 32, }) {
+                    for (int kv : { 512, 1024, }) {
+                        for (int nb : { 1, 2, 4, 8, }) {
+                            test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias));
+                        }
                     }
                 }
             }

tests/test-grammar-integration.cpp

@@ -28,6 +28,19 @@ static llama_grammar* build_grammar(const std::string & grammar_str) {
     return grammar;
 }
 
+static bool test_build_grammar_fails(const std::string & grammar_str) {
+    fprintf(stderr, "⚫ Testing failure for grammar: %s\n", grammar_str.c_str());
+    bool grammar_fails = false;
+    try {
+        build_grammar(grammar_str);
+        fprintf(stderr, "  ❌ Expected build failure, but succeeded\n");
+    } catch (const std::exception & err) {
+        grammar_fails = true;
+        fprintf(stdout, "  ✅︎\n");
+    }
+    return grammar_fails;
+}
+
 static bool match_string(const std::string & input, llama_grammar* grammar) {
     auto decoded = decode_utf8(input, {});
 
@@ -320,6 +333,38 @@ number ::= [0-9]+)""";
     fprintf(stderr, "  ✅︎ Passed\n");
 }
 
+static void test_failure_left_recursion() {
+    fprintf(stderr, "⚫ Testing left recursion detection:\n");
+
+    // Test simple left recursion detection
+    const std::string simple_str = R"""(root ::= "a" | root "a")""";
+    assert(test_build_grammar_fails(simple_str));
+
+    // Test more complicated left recursion detection
+    const std::string medium_str = R"""(
+root ::= asdf
+asdf ::= "a" | asdf "a"
+)""";
+    assert(test_build_grammar_fails(medium_str));
+
+    // Test even more complicated left recursion detection
+    const std::string hard_str = R"""(
+root ::= asdf
+asdf ::= "a" | foo "b"
+foo ::= "c" | asdf "d" | "e")""";
+    assert(test_build_grammar_fails(hard_str));
+
+    // Test yet even more complicated left recursion detection
+    const std::string hardest_str = R"""(
+root ::= asdf
+asdf ::= "a" | foo "b"
+foo ::= "c" | empty asdf "d" | "e"
+empty ::= "blah" | )""";
+    assert(test_build_grammar_fails(hardest_str));
+
+    fprintf(stderr, "  ✅︎ Passed\n");
+}
+
 int main() {
     fprintf(stdout, "Running grammar integration tests...\n");
     test_simple_grammar();
@@ -327,6 +372,7 @@ int main() {
     test_quantifiers();
     test_failure_missing_root();
     test_failure_missing_reference();
+    test_failure_left_recursion();
     fprintf(stdout, "All tests passed.\n");
     return 0;
 }