diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index 2a06f82b7..d93c0be6a 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -3,23 +3,36 @@ ARG UBUNTU_VERSION=22.04 FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ - apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1 - -COPY requirements.txt requirements.txt -COPY requirements requirements - -RUN pip install --upgrade pip setuptools wheel \ - && pip install -r requirements.txt + apt-get install -y build-essential git cmake libcurl4-openssl-dev WORKDIR /app COPY . . -ENV LLAMA_CURL=1 +RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \ + cmake --build build -j $(nproc) && \ + mkdir -p /app/lib && \ + find build -name "*.so" -exec cp {} /app/lib/ \; +FROM ubuntu:$UBUNTU_VERSION as runtime -RUN make -j$(nproc) +WORKDIR /app + +RUN apt-get update && \ + apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1 + +COPY requirements.txt /app/requirements.txt +COPY requirements /app/requirements +COPY .devops/tools.sh /app/tools.sh + +RUN pip install --upgrade pip setuptools wheel && \ + pip install -r /app/requirements.txt + +COPY --from=build /app/build/bin/ /app/ +COPY --from=build /app/lib/ /app/ +COPY --from=build /app/convert_hf_to_gguf.py /app/ +COPY --from=build /app/gguf-py /app/gguf-py ENV LC_ALL=C.utf8 -ENTRYPOINT ["/app/.devops/tools.sh"] +ENTRYPOINT ["/app/tools.sh"] diff --git a/.devops/llama-cli.Dockerfile b/.devops/llama-cli.Dockerfile index 7f741aa46..be234d55d 100644 --- a/.devops/llama-cli.Dockerfile +++ b/.devops/llama-cli.Dockerfile @@ -3,21 +3,27 @@ ARG UBUNTU_VERSION=22.04 FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ - apt-get install -y build-essential git + apt-get install -y build-essential git cmake libcurl4-openssl-dev WORKDIR /app COPY . . -RUN make -j$(nproc) llama-cli +RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \ + cmake --build build -j $(nproc) && \ + mkdir -p /app/lib && \ + find build -name "*.so" -exec cp {} /app/lib/ \; FROM ubuntu:$UBUNTU_VERSION AS runtime -RUN apt-get update && \ - apt-get install -y libgomp1 +WORKDIR /app -COPY --from=build /app/llama-cli /llama-cli +RUN apt-get update && \ + apt-get install -y libcurl4-openssl-dev libgomp1 curl + +COPY --from=build /app/build/bin/llama-cli /app/ +COPY --from=build /app/lib/ /app/ ENV LC_ALL=C.utf8 -ENTRYPOINT [ "/llama-cli" ] +ENTRYPOINT [ "/app/llama-cli" ] diff --git a/.devops/llama-server.Dockerfile b/.devops/llama-server.Dockerfile index 7110dda9e..72ccde2fe 100644 --- a/.devops/llama-server.Dockerfile +++ b/.devops/llama-server.Dockerfile @@ -9,28 +9,20 @@ WORKDIR /app COPY . . - -RUN \ - # Build multiple versions of the CPU backend - scripts/build-cpu.sh avx -DGGML_AVX=ON -DGGML_AVX2=OFF && \ - scripts/build-cpu.sh avx2 -DGGML_AVX=ON -DGGML_AVX2=ON && \ - scripts/build-cpu.sh avx512 -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON && \ - scripts/build-cpu.sh amx -DGGML_AVX=ON -DGGML_AVX2=ON -DGGML_AVX512=ON -DGGML_AVX_VNNI=ON -DGGML_AVX512_VNNI=ON -DGGML_AMX_TILE=ON -DGGML_AMX_INT8=ON && \ - # Build llama-server - cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \ - cmake --build build --target llama-server -j $(nproc) && \ - # Copy the built libraries to /app/lib +RUN cmake -S . -B build -DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_CURL=ON -DCMAKE_BUILD_TYPE=Release && \ + cmake --build build -j $(nproc) && \ mkdir -p /app/lib && \ - mv libggml-cpu* /app/lib/ && \ find build -name "*.so" -exec cp {} /app/lib/ \; FROM ubuntu:$UBUNTU_VERSION AS runtime +WORKDIR /app + RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev libgomp1 curl -COPY --from=build /app/build/bin/llama-server /llama-server -COPY --from=build /app/lib/ / +COPY --from=build /app/build/bin/llama-server /app/ +COPY --from=build /app/lib/ /app/ ENV LC_ALL=C.utf8 # Must be set to 0.0.0.0 so it can listen to requests from host machine @@ -38,4 +30,4 @@ ENV LLAMA_ARG_HOST=0.0.0.0 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ] -ENTRYPOINT [ "/llama-server" ] +ENTRYPOINT [ "/app/llama-server" ] diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index b88e6ca80..043c4364b 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -31,6 +31,7 @@ # Increases the runtime closure size by ~700M useMpi ? false, useRocm ? config.rocmSupport, + rocmGpuTargets ? builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets, enableCurl ? true, useVulkan ? false, llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake @@ -188,7 +189,7 @@ effectiveStdenv.mkDerivation (finalAttrs: { ] ++ optionals useRocm [ (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang") - (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets)) + (cmakeFeature "CMAKE_HIP_ARCHITECTURES" rocmGpuTargets) ] ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") diff --git a/.devops/tools.sh b/.devops/tools.sh index 24dcfd350..9a86e6ea0 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -8,11 +8,11 @@ arg1="$1" shift if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert_hf_to_gguf.py "$@" + exec python3 ./convert_hf_to_gguf.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then - ./llama-quantize "$@" + exec ./llama-quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then - ./llama-cli "$@" + exec ./llama-cli "$@" elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then echo "Converting PTH to GGML..." for i in `ls $1/$2/ggml-model-f16.bin*`; do @@ -20,11 +20,11 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then echo "Skip model quantization, it already exists: ${i/f16/q4_0}" else echo "Converting PTH to GGML: $i into ${i/f16/q4_0}..." - ./llama-quantize "$i" "${i/f16/q4_0}" q4_0 + exec ./llama-quantize "$i" "${i/f16/q4_0}" q4_0 fi done elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then - ./llama-server "$@" + exec ./llama-server "$@" else echo "Unknown command: $arg1" echo "Available commands: " diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index f3326a5fb..a377eff38 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -317,7 +317,7 @@ jobs: wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add - sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list sudo apt-get update -y - sudo apt-get install -y build-essential vulkan-sdk + sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk - name: Build id: cmake_build @@ -327,6 +327,12 @@ jobs: cmake -DGGML_VULKAN=ON .. cmake --build . --config Release -j $(nproc) + - name: Test + id: cmake_test + run: | + cd build + ctest -L main --verbose --timeout 900 + ubuntu-22-cmake-hip: runs-on: ubuntu-22.04 container: rocm/dev-ubuntu-22.04:6.0.2 @@ -552,35 +558,44 @@ jobs: -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO -# TODO: tmp disabled. see for possible re-enable: -# https://github.com/ggerganov/llama.cpp/pull/10525 -# macOS-latest-swift: -# runs-on: macos-latest -# -# strategy: -# matrix: -# destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS'] -# -# steps: -# - name: Clone -# id: checkout -# uses: actions/checkout@v4 -# -# - name: Dependencies -# id: depends -# continue-on-error: true -# run: | -# brew update -# -# - name: xcodebuild for swift package -# id: xcodebuild -# run: | -# xcodebuild -scheme llama -destination "${{ matrix.destination }}" -# -# - name: Build Swift Example -# id: make_build_swift_example -# run: | -# make swift + macOS-latest-swift: + runs-on: macos-latest + + strategy: + matrix: + destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS'] + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build llama.cpp with CMake + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) + sudo cmake --install . --config Release + + - name: xcodebuild for swift package + id: xcodebuild + run: | + xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}" windows-msys2: runs-on: windows-latest @@ -653,6 +668,8 @@ jobs: defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_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 -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON' + - build: 'llvm-arm64-opencl-adreno' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON' steps: - name: Clone @@ -694,6 +711,28 @@ jobs: run: | choco install ninja + - name: Install OpenCL Headers and Libs + id: install_opencl + if: ${{ matrix.build == 'llvm-arm64-opencl-adreno' }} + run: | + git clone https://github.com/KhronosGroup/OpenCL-Headers + cd OpenCL-Headers + mkdir build && cd build + cmake .. ` + -DBUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF ` + -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release" + cmake --build . --target install + git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader + cd OpenCL-ICD-Loader + mkdir build-arm64-release && cd build-arm64-release + cmake .. ` + -A arm64 ` + -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" ` + -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release" + cmake --build . --target install --config release + - name: Build id: cmake_build run: | @@ -723,7 +762,7 @@ jobs: - name: Test id: cmake_test # not all machines have native AVX-512 - if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }} + if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' && 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 @@ -1104,6 +1143,29 @@ jobs: - name: Checkout code uses: actions/checkout@v4 + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \ + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + sudo cmake --install . --config Release + + - name: xcodebuild for swift package + id: xcodebuild + run: | + xcodebuild -scheme llama-Package -destination 'generic/platform=iOS' + - name: Build Xcode project run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build @@ -1131,23 +1193,6 @@ jobs: ./gradlew build --no-daemon -# freeBSD-latest: -# runs-on: macos-12 -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Build -# uses: cross-platform-actions/action@v0.19.0 -# with: -# operating_system: freebsd -# version: '13.2' -# hypervisor: 'qemu' -# run: | -# sudo pkg update -# sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas -# gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu` - release: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 9e66fb68c..671fe595c 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -79,7 +79,7 @@ jobs: # Setup nodejs (to be used for verifying bundled index.html) - uses: actions/setup-node@v4 with: - node-version: 22 + node-version: '22.11.0' - name: Verify bundled index.html id: verify_server_index_html diff --git a/CMakeLists.txt b/CMakeLists.txt index f84fff9e6..a717a508f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -46,11 +46,9 @@ if (WIN32) add_compile_definitions(_CRT_SECURE_NO_WARNINGS) endif() -if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC") - add_compile_options("$<$:/source-charset:utf-8>") - add_compile_options("$<$:/source-charset:utf-8>") - add_compile_options("$<$:/execution-charset:utf-8>") - add_compile_options("$<$:/execution-charset:utf-8>") +if (MSVC) + add_compile_options("$<$:/utf-8>") + add_compile_options("$<$:/utf-8>") endif() # diff --git a/CMakePresets.json b/CMakePresets.json index 436448967..13bdd7907 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -31,6 +31,13 @@ { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } }, { "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } }, + { + "name": "x64-windows-llvm", "hidden": true, + "cacheVariables": { + "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/x64-windows-llvm.cmake" + } + }, + { "name": "arm64-windows-msvc", "hidden": true, "architecture": { "value": "arm64", "strategy": "external" }, @@ -70,6 +77,11 @@ { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] }, { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] }, + { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] }, + { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] }, + { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] }, + { "name": "x64-windows-llvm+static-release", "inherits": [ "base", "x64-windows-llvm", "reldbg", "static" ] }, + { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] }, { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] }, { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] }, diff --git a/CODEOWNERS b/CODEOWNERS index 88ab6de4f..adeba5395 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -1,3 +1,5 @@ # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs -ci/ @ggerganov +/ci/ @ggerganov +/.devops/ @ngxson +/examples/server/ @ngxson diff --git a/Makefile b/Makefile index d76c4ad53..19ae0d5f1 100644 --- a/Makefile +++ b/Makefile @@ -22,6 +22,7 @@ BUILD_TARGETS = \ llama-infill \ llama-llava-cli \ llama-minicpmv-cli\ + llama-qwen2vl-cli\ llama-lookahead \ llama-lookup \ llama-lookup-create \ @@ -445,6 +446,10 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) MK_CFLAGS += -march=native -mtune=native HOST_CXXFLAGS += -march=native -mtune=native + # Usage AMX build test + #MK_CFLAGS += -march=graniterapids -mtune=graniterapids + #HOST_CXXFLAGS += -march=graniterapids -mtune=graniterapids + # Usage AVX-only #MK_CFLAGS += -mfma -mf16c -mavx #MK_CXXFLAGS += -mfma -mf16c -mavx @@ -948,7 +953,6 @@ DIR_COMMON = common OBJ_GGML = \ $(DIR_GGML)/src/ggml.o \ - $(DIR_GGML)/src/ggml-aarch64.o \ $(DIR_GGML)/src/ggml-alloc.o \ $(DIR_GGML)/src/ggml-backend.o \ $(DIR_GGML)/src/ggml-backend-reg.o \ @@ -956,9 +960,11 @@ OBJ_GGML = \ $(DIR_GGML)/src/ggml-quants.o \ $(DIR_GGML)/src/ggml-threading.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \ - $(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \ $(OBJ_GGML_EXT) OBJ_LLAMA = \ @@ -1098,17 +1104,10 @@ DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d) # Default target all: $(BUILD_TARGETS) +# force c++ build for source file that have same name as c file # Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files -# g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml -$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \ - ggml/src/ggml-cpu/ggml-cpu.cpp \ - ggml/include/ggml-backend.h \ - ggml/include/ggml.h \ - ggml/include/ggml-alloc.h \ - ggml/src/ggml-backend-impl.h \ - ggml/include/ggml-cpu.h \ - ggml/src/ggml-impl.h - $(CXX) $(CXXFLAGS) -c $< -o $@ +$(DIR_GGML)/%_cpp.o: $(DIR_GGML)/%.cpp + $(CXX) $(CXXFLAGS) -MMD -c $< -o $@ # Rules for building object files $(DIR_GGML)/%.o: $(DIR_GGML)/%.c @@ -1406,6 +1405,14 @@ llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual +llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \ + examples/llava/llava.cpp \ + examples/llava/llava.h \ + examples/llava/clip.cpp \ + examples/llava/clip.h \ + $(OBJ_ALL) + $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual + ifeq ($(UNAME_S),Darwin) swift: examples/batched.swift (cd examples/batched.swift; make build) diff --git a/Package.swift b/Package.swift index d32b74a63..01c996d24 100644 --- a/Package.swift +++ b/Package.swift @@ -2,59 +2,6 @@ import PackageDescription -var sources = [ - "src/llama.cpp", - "src/llama-vocab.cpp", - "src/llama-grammar.cpp", - "src/llama-sampling.cpp", - "src/unicode.cpp", - "src/unicode-data.cpp", - "ggml/src/ggml.c", - "ggml/src/ggml-aarch64.c", - "ggml/src/ggml-alloc.c", - "ggml/src/ggml-backend.cpp", - "ggml/src/ggml-backend-reg.cpp", - "ggml/src/ggml-cpu/ggml-cpu.c", - "ggml/src/ggml-cpu/ggml-cpu.cpp", - "ggml/src/ggml-cpu/ggml-cpu-aarch64.c", - "ggml/src/ggml-cpu/ggml-cpu-quants.c", - "ggml/src/ggml-threading.cpp", - "ggml/src/ggml-quants.c", -] - -var resources: [Resource] = [] -var linkerSettings: [LinkerSetting] = [] -var cSettings: [CSetting] = [ - .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]), - .unsafeFlags(["-fno-objc-arc"]), - .headerSearchPath("ggml/src"), - .headerSearchPath("ggml/src/ggml-cpu"), - // NOTE: NEW_LAPACK will required iOS version 16.4+ - // We should consider add this in the future when we drop support for iOS 14 - // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc) - // .define("ACCELERATE_NEW_LAPACK"), - // .define("ACCELERATE_LAPACK_ILP64") - .define("GGML_USE_CPU"), -] - - -#if canImport(Darwin) -sources.append("ggml/src/ggml-common.h") -sources.append("ggml/src/ggml-metal/ggml-metal.m") -resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal")) -linkerSettings.append(.linkedFramework("Accelerate")) -cSettings.append( - contentsOf: [ - .define("GGML_USE_ACCELERATE"), - .define("GGML_USE_METAL"), - ] -) -#endif - -#if os(Linux) - cSettings.append(.define("_GNU_SOURCE")) -#endif - let package = Package( name: "llama", platforms: [ @@ -67,26 +14,6 @@ let package = Package( .library(name: "llama", targets: ["llama"]), ], targets: [ - .target( - name: "llama", - path: ".", - exclude: [ - "build", - "cmake", - "examples", - "scripts", - "models", - "tests", - "CMakeLists.txt", - "Makefile", - "ggml/src/ggml-metal-embed.metal" - ], - sources: sources, - resources: resources, - publicHeadersPath: "spm-headers", - cSettings: cSettings, - linkerSettings: linkerSettings - ) - ], - cxxLanguageStandard: .cxx17 + .systemLibrary(name: "llama", pkgConfig: "llama"), + ] ) diff --git a/README.md b/README.md index 7ea1872ec..a94dd84f7 100644 --- a/README.md +++ b/README.md @@ -98,6 +98,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat) - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a) - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM) +- [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct) #### Multimodal @@ -111,6 +112,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2) - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny) - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge) +- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d) @@ -220,7 +222,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo | [SYCL](docs/backend/SYCL.md) | Intel and Nvidia GPU | | [MUSA](docs/build.md#musa) | Moore Threads MTT GPU | | [CUDA](docs/build.md#cuda) | Nvidia GPU | -| [hipBLAS](docs/build.md#hipblas) | AMD GPU | +| [HIP](docs/build.md#hip) | AMD GPU | | [Vulkan](docs/build.md#vulkan) | GPU | | [CANN](docs/build.md#cann) | Ascend NPU | @@ -413,7 +415,7 @@ To learn more about model quantization, [read this documentation](examples/quant [^1]: [examples/perplexity/README.md](examples/perplexity/README.md) [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity) -## [`llama-bench`](example/bench) +## [`llama-bench`](examples/llama-bench) #### Benchmark the performance of the inference for various parameters. @@ -434,6 +436,20 @@ To learn more about model quantization, [read this documentation](examples/quant +## [`llama-run`](examples/run) + +#### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3]. + +-
+ Run a model with a specific prompt (by default it's pulled from Ollama registry) + + ```bash + llama-run granite-code + ``` + +
+ +[^3]: [RamaLama](https://github.com/containers/ramalama) ## [`llama-simple`](examples/simple) diff --git a/Sources/llama/llama.h b/Sources/llama/llama.h new file mode 100644 index 000000000..41725880e --- /dev/null +++ b/Sources/llama/llama.h @@ -0,0 +1,4 @@ +#pragma once + +#include + diff --git a/Sources/llama/module.modulemap b/Sources/llama/module.modulemap new file mode 100644 index 000000000..d010555b1 --- /dev/null +++ b/Sources/llama/module.modulemap @@ -0,0 +1,5 @@ +module llama [system] { + header "llama.h" + link "llama" + export * +} diff --git a/cmake/llama.pc.in b/cmake/llama.pc.in index 326acbb61..0b2b6bcfa 100644 --- a/cmake/llama.pc.in +++ b/cmake/llama.pc.in @@ -6,5 +6,5 @@ includedir=${prefix}/include Name: llama Description: Port of Facebook's LLaMA model in C/C++ Version: @PROJECT_VERSION@ -Libs: -L${libdir} -lllama +Libs: -L${libdir} -lggml -lggml-base -lllama Cflags: -I${includedir} diff --git a/cmake/x64-windows-llvm.cmake b/cmake/x64-windows-llvm.cmake new file mode 100644 index 000000000..0603d738f --- /dev/null +++ b/cmake/x64-windows-llvm.cmake @@ -0,0 +1,11 @@ +set( CMAKE_SYSTEM_NAME Windows ) +set( CMAKE_SYSTEM_PROCESSOR x86_64 ) + +set( CMAKE_C_COMPILER clang ) +set( CMAKE_CXX_COMPILER clang++ ) + +set( arch_c_flags "-march=native" ) + +set( CMAKE_C_FLAGS_INIT "${arch_c_flags}" ) +set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" ) + diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index 89862fe11..df1cdf9a5 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -81,7 +81,7 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info) # Use curl to download model url if (LLAMA_CURL) find_package(CURL REQUIRED) - add_definitions(-DLLAMA_USE_CURL) + target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL) include_directories(${CURL_INCLUDE_DIRS}) find_library(CURL_LIBRARY curl REQUIRED) set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY}) diff --git a/common/arg.cpp b/common/arg.cpp index 078c75384..e5ddd8318 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -119,32 +119,65 @@ std::string common_arg::to_string() { // utils // -static void common_params_handle_model_default(common_params & params) { - if (!params.hf_repo.empty()) { +static void common_params_handle_model_default( + std::string & model, + std::string & model_url, + std::string & hf_repo, + std::string & hf_file) { + if (!hf_repo.empty()) { // short-hand to avoid specifying --hf-file -> default it to --model - if (params.hf_file.empty()) { - if (params.model.empty()) { + if (hf_file.empty()) { + if (model.empty()) { throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n"); } - params.hf_file = params.model; - } else if (params.model.empty()) { + hf_file = model; + } else if (model.empty()) { // this is to avoid different repo having same file name, or same file name in different subdirs - std::string filename = params.hf_repo + "_" + params.hf_file; + std::string filename = hf_repo + "_" + hf_file; // to make sure we don't have any slashes in the filename string_replace_all(filename, "/", "_"); - params.model = fs_get_cache_file(filename); + model = fs_get_cache_file(filename); } - } else if (!params.model_url.empty()) { - if (params.model.empty()) { - auto f = string_split(params.model_url, '#').front(); + } else if (!model_url.empty()) { + if (model.empty()) { + auto f = string_split(model_url, '#').front(); f = string_split(f, '?').front(); - params.model = fs_get_cache_file(string_split(f, '/').back()); + model = fs_get_cache_file(string_split(f, '/').back()); } - } else if (params.model.empty()) { - params.model = DEFAULT_MODEL_PATH; + } else if (model.empty()) { + model = DEFAULT_MODEL_PATH; } } +const std::vector kv_cache_types = { + GGML_TYPE_F32, + GGML_TYPE_F16, + GGML_TYPE_BF16, + GGML_TYPE_Q8_0, + GGML_TYPE_Q4_0, + GGML_TYPE_Q4_1, + GGML_TYPE_IQ4_NL, + GGML_TYPE_Q5_0, + GGML_TYPE_Q5_1, +}; + +static ggml_type kv_cache_type_from_str(const std::string & s) { + for (const auto & type : kv_cache_types) { + if (ggml_type_name(type) == s) { + return type; + } + } + throw std::runtime_error("Unsupported cache type: " + s); +} + +static std::string get_all_kv_cache_types() { + std::ostringstream msg; + for (const auto & type : kv_cache_types) { + msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", "); + } + return msg.str(); +} + // // CLI argument parsing functions // @@ -247,7 +280,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } - common_params_handle_model_default(params); + // TODO: refactor model params in a common struct + common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file); + common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file); if (params.escape) { string_process_escapes(params.prompt); @@ -591,7 +626,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.ctx_shift = false; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT")); + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT")); add_opt(common_arg( {"--chunks"}, "N", string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks), @@ -786,7 +821,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.warmup = false; } - ).set_examples({LLAMA_EXAMPLE_MAIN})); + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"--spm-infill"}, string_format( @@ -813,7 +848,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } ).set_sparam()); add_opt(common_arg( - {"--sampling-seq"}, "SEQUENCE", + {"--sampling-seq", "--sampler-seq"}, "SEQUENCE", string_format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()), [](common_params & params, const std::string & value) { params.sampling.samplers = common_sampler_types_from_chars(value); @@ -826,13 +861,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sampling.ignore_eos = true; } ).set_sparam()); - add_opt(common_arg( - {"--penalize-nl"}, - string_format("penalize newline tokens (default: %s)", params.sampling.penalize_nl ? "true" : "false"), - [](common_params & params) { - params.sampling.penalize_nl = true; - } - ).set_sparam()); add_opt(common_arg( {"--temp"}, "N", string_format("temperature (default: %.1f)", (double)params.sampling.temp), @@ -887,6 +915,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--repeat-last-n"}, "N", string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n), [](common_params & params, int value) { + if (value < -1) { + throw std::runtime_error(string_format("error: invalid repeat-last-n = %d\n", value)); + } params.sampling.penalty_last_n = value; params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n); } @@ -941,6 +972,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--dry-penalty-last-n"}, "N", string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n), [](common_params & params, int value) { + if (value < -1) { + throw std::runtime_error(string_format("error: invalid dry-penalty-last-n = %d\n", value)); + } params.sampling.dry_penalty_last_n = value; } ).set_sparam()); @@ -1174,18 +1208,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_NO_KV_OFFLOAD")); add_opt(common_arg( {"-ctk", "--cache-type-k"}, "TYPE", - string_format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()), + string_format( + "KV cache data type for K\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.cache_type_k) + ), [](common_params & params, const std::string & value) { - // TODO: get the type right here - params.cache_type_k = value; + params.cache_type_k = kv_cache_type_from_str(value); } ).set_env("LLAMA_ARG_CACHE_TYPE_K")); add_opt(common_arg( {"-ctv", "--cache-type-v"}, "TYPE", - string_format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()), + string_format( + "KV cache data type for V\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.cache_type_v) + ), [](common_params & params, const std::string & value) { - // TODO: get the type right here - params.cache_type_v = value; + params.cache_type_v = kv_cache_type_from_str(value); } ).set_env("LLAMA_ARG_CACHE_TYPE_V")); add_opt(common_arg( @@ -1543,6 +1587,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.hf_file = value; } ).set_env("LLAMA_ARG_HF_FILE")); + add_opt(common_arg( + {"-hfrv", "--hf-repo-v"}, "REPO", + "Hugging Face model repository for the vocoder model (default: unused)", + [](common_params & params, const std::string & value) { + params.vocoder.hf_repo = value; + } + ).set_env("LLAMA_ARG_HF_REPO_V")); + add_opt(common_arg( + {"-hffv", "--hf-file-v"}, "FILE", + "Hugging Face model file for the vocoder model (default: unused)", + [](common_params & params, const std::string & value) { + params.vocoder.hf_file = value; + } + ).set_env("LLAMA_ARG_HF_FILE_V")); add_opt(common_arg( {"-hft", "--hf-token"}, "TOKEN", "Hugging Face access token (default: value from HF_TOKEN environment variable)", @@ -1711,6 +1769,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.public_path = value; } ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH")); + add_opt(common_arg( + {"--no-webui"}, + string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"), + [](common_params & params) { + params.webui = false; + } + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_WEBUI")); add_opt(common_arg( {"--embedding", "--embeddings"}, string_format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"), @@ -2076,35 +2141,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, int value) { params.speculative.n_max = value; } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MAX")); add_opt(common_arg( {"--draft-min", "--draft-n-min"}, "N", string_format("minimum number of draft tokens to use for speculative decoding (default: %d)", params.speculative.n_min), [](common_params & params, int value) { params.speculative.n_min = value; } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_MIN")); add_opt(common_arg( {"--draft-p-split"}, "P", string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split), [](common_params & params, const std::string & value) { params.speculative.p_split = std::stof(value); } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT")); add_opt(common_arg( {"--draft-p-min"}, "P", string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min), [](common_params & params, const std::string & value) { params.speculative.p_min = std::stof(value); } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_DRAFT_P_MIN")); add_opt(common_arg( {"-cd", "--ctx-size-draft"}, "N", string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx), [](common_params & params, int value) { params.speculative.n_ctx = value; } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CTX_SIZE_DRAFT")); add_opt(common_arg( {"-devd", "--device-draft"}, "", "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n" @@ -2124,14 +2189,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex fprintf(stderr, "warning: consult docs/build.md for compilation instructions\n"); } } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_GPU_LAYERS_DRAFT")); add_opt(common_arg( {"-md", "--model-draft"}, "FNAME", "draft model for speculative decoding (default: unused)", [](common_params & params, const std::string & value) { params.speculative.model = value; } - ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT")); + + add_opt(common_arg( + {"-mv", "--model-vocoder"}, "FNAME", + "vocoder model for audio generation (default: unused)", + [](common_params & params, const std::string & value) { + params.vocoder.model = value; + } + ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER})); return ctx_arg; } diff --git a/common/common.cpp b/common/common.cpp index 6143516d2..20be92911 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -940,6 +940,25 @@ struct common_init_result common_init_from_params(common_params & params) { params.sampling.ignore_eos = false; } + if (params.sampling.ignore_eos) { + for (llama_token i = 0; i < llama_n_vocab(model); i++) { + if (llama_token_is_eog(model, i)) { + LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY); + params.sampling.logit_bias.push_back({i, -INFINITY}); + } + } + } + + if (params.sampling.penalty_last_n == -1) { + LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx)); + params.sampling.penalty_last_n = llama_n_ctx(lctx); + } + + if (params.sampling.dry_penalty_last_n == -1) { + LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx)); + params.sampling.dry_penalty_last_n = llama_n_ctx(lctx); + } + if (params.warmup) { LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__); @@ -1015,38 +1034,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) { return mparams; } -static ggml_type kv_cache_type_from_str(const std::string & s) { - if (s == "f32") { - return GGML_TYPE_F32; - } - if (s == "f16") { - return GGML_TYPE_F16; - } - if (s == "bf16") { - return GGML_TYPE_BF16; - } - if (s == "q8_0") { - return GGML_TYPE_Q8_0; - } - if (s == "q4_0") { - return GGML_TYPE_Q4_0; - } - if (s == "q4_1") { - return GGML_TYPE_Q4_1; - } - if (s == "iq4_nl") { - return GGML_TYPE_IQ4_NL; - } - if (s == "q5_0") { - return GGML_TYPE_Q5_0; - } - if (s == "q5_1") { - return GGML_TYPE_Q5_1; - } - - throw std::runtime_error("Unsupported cache type: " + s); -} - struct llama_context_params common_context_params_to_llama(const common_params & params) { auto cparams = llama_context_default_params(); @@ -1081,8 +1068,8 @@ struct llama_context_params common_context_params_to_llama(const common_params & cparams.pooling_type = LLAMA_POOLING_TYPE_RANK; } - cparams.type_k = kv_cache_type_from_str(params.cache_type_k); - cparams.type_v = kv_cache_type_from_str(params.cache_type_v); + cparams.type_k = params.cache_type_k; + cparams.type_v = params.cache_type_v; return cparams; } @@ -1108,13 +1095,7 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p #define CURL_MAX_RETRY 3 #define CURL_RETRY_DELAY_SECONDS 2 - -static bool starts_with(const std::string & str, const std::string & prefix) { - // While we wait for C++20's std::string::starts_with... - return str.rfind(prefix, 0) == 0; -} - -static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) { +static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) { int remaining_attempts = max_attempts; while (remaining_attempts > 0) { @@ -1138,7 +1119,6 @@ static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_ } static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) { - // Initialize libcurl std::unique_ptr curl(curl_easy_init(), &curl_easy_cleanup); if (!curl) { @@ -1211,11 +1191,13 @@ static bool common_download_file(const std::string & url, const std::string & pa std::string etag; std::string last_modified; }; + common_load_model_from_url_headers headers; + { typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *); auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { - common_load_model_from_url_headers *headers = (common_load_model_from_url_headers *) userdata; + common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata; static std::regex header_regex("([^:]+): (.*)\r\n"); static std::regex etag_regex("ETag", std::regex_constants::icase); @@ -1799,7 +1781,9 @@ void common_embd_normalize(const float * inp, float * out, int n, int embd_norm) break; case 0: // max absolute for (int i = 0; i < n; i++) { - if (sum < std::abs(inp[i])) sum = std::abs(inp[i]); + if (sum < std::abs(inp[i])) { + sum = std::abs(inp[i]); + } } sum /= 32760.0; // make an int16 range break; diff --git a/common/common.h b/common/common.h index 0373fd3ea..1d2bd932c 100644 --- a/common/common.h +++ b/common/common.h @@ -37,9 +37,9 @@ using llama_tokens = std::vector; // build info extern int LLAMA_BUILD_NUMBER; -extern char const * LLAMA_COMMIT; -extern char const * LLAMA_COMPILER; -extern char const * LLAMA_BUILD_TARGET; +extern const char * LLAMA_COMMIT; +extern const char * LLAMA_COMPILER; +extern const char * LLAMA_BUILD_TARGET; struct common_control_vector_load_info; @@ -80,6 +80,7 @@ enum llama_example { LLAMA_EXAMPLE_LLAVA, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_PARALLEL, + LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_COUNT, }; @@ -95,6 +96,7 @@ enum common_sampler_type { COMMON_SAMPLER_TYPE_TEMPERATURE = 7, COMMON_SAMPLER_TYPE_XTC = 8, COMMON_SAMPLER_TYPE_INFILL = 9, + COMMON_SAMPLER_TYPE_PENALTIES = 10, }; // dimensionality reduction methods, used by cvector-generator @@ -130,7 +132,6 @@ struct common_params_sampling { int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0 float mirostat_tau = 5.00f; // target entropy float mirostat_eta = 0.10f; // learning rate - bool penalize_nl = false; // consider newlines as a repeatable token bool ignore_eos = false; bool no_perf = false; // disable performance metrics bool timing_per_token = false; @@ -139,6 +140,7 @@ struct common_params_sampling { std::vector samplers = { + COMMON_SAMPLER_TYPE_PENALTIES, COMMON_SAMPLER_TYPE_DRY, COMMON_SAMPLER_TYPE_TOP_K, COMMON_SAMPLER_TYPE_TYPICAL_P, @@ -158,6 +160,7 @@ struct common_params_sampling { struct common_params_speculative { std::vector devices; // devices to use for offloading + int32_t n_ctx = 0; // draft context size int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding int32_t n_min = 5; // minimum number of draft tokens to use for speculative decoding @@ -171,6 +174,14 @@ struct common_params_speculative { std::string model = ""; // draft model for speculative decoding // NOLINT }; +struct common_params_vocoder { + std::string hf_repo = ""; // HF repo // NOLINT + std::string hf_file = ""; // HF file // NOLINT + + std::string model = ""; // model path // NOLINT + std::string model_url = ""; // model url to download // NOLINT +}; + struct common_params { int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 4096; // context size @@ -193,11 +204,13 @@ struct common_params { float defrag_thold = 0.1f; // KV cache defragmentation threshold // offload params - std::vector devices; // devices to use for offloading - int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default) - int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors - float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs - enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs + std::vector devices; // devices to use for offloading + + int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default) + int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors + float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs + + enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs struct cpu_params cpuparams; struct cpu_params cpuparams_batch; @@ -211,11 +224,12 @@ struct common_params { enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings - struct common_params_sampling sampling; + struct common_params_sampling sampling; struct common_params_speculative speculative; + struct common_params_vocoder vocoder; std::string model = ""; // model path // NOLINT - std::string model_alias = "unknown"; // model alias // NOLINT + std::string model_alias = ""; // model alias // NOLINT std::string model_url = ""; // model url to download // NOLINT std::string hf_token = ""; // HF token // NOLINT std::string hf_repo = ""; // HF repo // NOLINT @@ -286,8 +300,8 @@ struct common_params { bool warmup = true; // warmup run bool check_tensors = false; // validate tensor data - std::string cache_type_k = "f16"; // KV cache data type for the K - std::string cache_type_v = "f16"; // KV cache data type for the V + ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K + ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V // multimodal models (see examples/llava) std::string mmproj = ""; // path to multimodal projector // NOLINT @@ -437,6 +451,11 @@ std::vector string_split(const std::string & input, ch return parts; } +static bool string_starts_with(const std::string & str, + const std::string & prefix) { // While we wait for C++20's std::string::starts_with... + return str.rfind(prefix, 0) == 0; +} + bool string_parse_kv_override(const char * data, std::vector & overrides); void string_process_escapes(std::string & input); @@ -588,7 +607,8 @@ void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_si // Embedding utils // -void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2); +// TODO: repace embd_norm with an enum +void common_embd_normalize(const float * inp, float * out, int n, int embd_norm); float common_embd_similarity_cos(const float * embd1, const float * embd2, int n); diff --git a/common/sampling.cpp b/common/sampling.cpp index 0c4699a89..e83a971c7 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -161,32 +161,20 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co params.logit_bias.size(), params.logit_bias.data())); - llama_sampler_chain_add(result->chain, - llama_sampler_init_penalties( - llama_n_vocab (model), - llama_token_eos(model), - llama_token_nl (model), - params.penalty_last_n, - params.penalty_repeat, - params.penalty_freq, - params.penalty_present, - params.penalize_nl, - params.ignore_eos)); - if (params.mirostat == 0) { for (const auto & cnstr : params.samplers) { switch (cnstr) { - case COMMON_SAMPLER_TYPE_DRY: + case COMMON_SAMPLER_TYPE_DRY: { - std::vector c_breakers; + std::vector c_breakers; c_breakers.reserve(params.dry_sequence_breakers.size()); - for (const auto& str : params.dry_sequence_breakers) { + for (const auto & str : params.dry_sequence_breakers) { c_breakers.push_back(str.c_str()); } llama_sampler_chain_add(result->chain, llama_sampler_init_dry (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size())); } - break; + break; case COMMON_SAMPLER_TYPE_TOP_K: llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); break; @@ -208,6 +196,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co case COMMON_SAMPLER_TYPE_INFILL: llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model)); break; + case COMMON_SAMPLER_TYPE_PENALTIES: + llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present)); + break; default: GGML_ASSERT(false && "unknown sampler type"); } @@ -415,6 +406,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) { case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't'; case COMMON_SAMPLER_TYPE_XTC: return 'x'; case COMMON_SAMPLER_TYPE_INFILL: return 'i'; + case COMMON_SAMPLER_TYPE_PENALTIES: return 'e'; default : return '?'; } } @@ -429,6 +421,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) { case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature"; case COMMON_SAMPLER_TYPE_XTC: return "xtc"; case COMMON_SAMPLER_TYPE_INFILL: return "infill"; + case COMMON_SAMPLER_TYPE_PENALTIES: return "penalties"; default : return ""; } } @@ -443,6 +436,7 @@ std::vector common_sampler_types_from_names(const std::vect { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE }, { "xtc", COMMON_SAMPLER_TYPE_XTC }, { "infill", COMMON_SAMPLER_TYPE_INFILL }, + { "penalties", COMMON_SAMPLER_TYPE_PENALTIES }, }; // since samplers names are written multiple ways @@ -489,6 +483,7 @@ std::vector common_sampler_types_from_chars(const std::stri { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC), COMMON_SAMPLER_TYPE_XTC }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL), COMMON_SAMPLER_TYPE_INFILL }, + { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_PENALTIES), COMMON_SAMPLER_TYPE_PENALTIES }, }; std::vector samplers; diff --git a/common/speculative.cpp b/common/speculative.cpp index e559675c4..3fcbb0020 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -62,6 +62,10 @@ struct common_speculative * common_speculative_init( } void common_speculative_free(struct common_speculative * spec) { + if (spec == nullptr) { + return; + } + common_sampler_free(spec->smpl); llama_batch_free(spec->batch); diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index c3931aa2f..7c99d7a68 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -221,17 +221,17 @@ class Model: self.gguf_writer.add_context_length(n_ctx) logger.info(f"gguf: context length = {n_ctx}") - n_embd = self.find_hparam(["hidden_size", "n_embd"]) - self.gguf_writer.add_embedding_length(n_embd) - logger.info(f"gguf: embedding length = {n_embd}") + if (n_embd := self.find_hparam(["hidden_size", "n_embd"], optional=True)) is not None: + self.gguf_writer.add_embedding_length(n_embd) + logger.info(f"gguf: embedding length = {n_embd}") if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None: self.gguf_writer.add_feed_forward_length(n_ff) logger.info(f"gguf: feed forward length = {n_ff}") - n_head = self.find_hparam(["num_attention_heads", "n_head"]) - self.gguf_writer.add_head_count(n_head) - logger.info(f"gguf: head count = {n_head}") + if (n_head := self.find_hparam(["num_attention_heads", "n_head"], optional=True)) is not None: + self.gguf_writer.add_head_count(n_head) + logger.info(f"gguf: head count = {n_head}") if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None: self.gguf_writer.add_head_count_kv(n_head_kv) @@ -296,7 +296,9 @@ class Model: break for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)): - data = data_torch.squeeze().numpy() + # TODO: why do we squeeze here? + # data = data_torch.squeeze().numpy() + data = data_torch.numpy() # if data ends up empty, it means data_torch was a scalar tensor -> restore if len(data.shape) == 0: @@ -324,6 +326,8 @@ class Model: gguf.MODEL_TENSOR.TIME_MIX_W2, gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1, gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2, + gguf.MODEL_TENSOR.POSNET_NORM1, + gguf.MODEL_TENSOR.POSNET_NORM2, ) ) or not new_name.endswith(".weight") @@ -658,6 +662,15 @@ class Model: if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450": # ref: https://huggingface.co/facebook/chameleon-7b res = "chameleon" + if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35": + # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0 + res = "minerva-7b" + if chkhsh == "8b5a93ed704057481f240da0be7e7dca721d7f8f4755263b6807227a2cbeae65": + # ref: https://huggingface.co/sentence-transformers/stsb-roberta-base + res = "roberta-bpe" + if chkhsh == "ad851be1dba641f2e3711822f816db2c265f788b37c63b4e1aeacb9ee92de8eb": + # ref: https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct + res = "gigachat" if res is None: logger.warning("\n") @@ -680,6 +693,9 @@ class Model: return res # Marker: End get_vocab_base_pre + def _set_vocab_none(self) -> None: + self.gguf_writer.add_tokenizer_model("none") + def _set_vocab_gpt2(self) -> None: tokens, toktypes, tokpre = self.get_vocab_base() self.gguf_writer.add_tokenizer_model("gpt2") @@ -1831,29 +1847,40 @@ class MiniCPMModel(Model): model_arch = gguf.MODEL_ARCH.MINICPM def set_gguf_parameters(self): - block_count = self.hparams["num_hidden_layers"] - self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) - self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) - self.gguf_writer.add_block_count(block_count) - self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) - 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_head_count_kv(self.hparams["num_key_value_heads"]) - self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) - self.gguf_writer.add_file_type(self.ftype) + super().set_gguf_parameters() + embedding_scale = float(self.hparams["scale_emb"]) + self.gguf_writer.add_embedding_scale(embedding_scale) + logger.info(f"gguf: (minicpm) embedding_scale = {embedding_scale}") + residual_scale = self.hparams["scale_depth"] / self.hparams["num_hidden_layers"] ** 0.5 + self.gguf_writer.add_residual_scale(residual_scale) + logger.info(f"gguf: (minicpm) residual_scale = {residual_scale}") + logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"] + self.gguf_writer.add_logit_scale(logit_scale) + logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}") + if self.hparams.get("rope_scaling") is not None: + if self.hparams["rope_scaling"].get("type") == "longrope": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE) + logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}") + + def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: + rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] + + rope_scaling = self.find_hparam(['rope_scaling'], True) + if rope_scaling is not None: + long_factors = rope_scaling.get('long_factor', None) + short_factors = rope_scaling.get('short_factor', None) + + if long_factors is None or short_factors is None: + raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') + + if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: + raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') + + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) + yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) def set_vocab(self): - self._set_vocab_llama_hf() - - def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: - if n_kv_head is not None and n_head != n_kv_head: - n_head //= n_kv_head - - return ( - weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape) - ) + self._set_vocab_sentencepiece() def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused @@ -1863,9 +1890,9 @@ class MiniCPMModel(Model): # HF models permute some of the tensors, so we need to undo that if name.endswith(("q_proj.weight")): - data_torch = self._reverse_hf_permute(data_torch, n_head, n_head) + data_torch = LlamaModel.permute(data_torch, n_head, n_head) if name.endswith(("k_proj.weight")): - data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head) + data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) return [(self.map_tensor_name(name), data_torch)] @@ -1975,6 +2002,75 @@ class Qwen2Model(Model): except FileNotFoundError: self._set_vocab_gpt2() + def set_gguf_parameters(self): + super().set_gguf_parameters() + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "yarn": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + + +@Model.register("Qwen2VLForConditionalGeneration") +class Qwen2VLModel(Model): + model_arch = gguf.MODEL_ARCH.QWEN2VL + + def set_gguf_parameters(self): + super().set_gguf_parameters() + mrope_section = self.hparams["rope_scaling"]["mrope_section"] + mrope_section += [0] * max(0, 4 - len(mrope_section)) + self.gguf_writer.add_rope_dimension_sections(mrope_section) + + def set_vocab(self): + try: + self._set_vocab_sentencepiece() + except FileNotFoundError: + self._set_vocab_gpt2() + + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: + for name, data in super().get_tensors(): + if name.startswith("visual."): + continue + yield name, data + + +@Model.register("WavTokenizerDec") +class WavTokenizerDecModel(Model): + model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + if \ + name.endswith("codebook.cluster_size") or \ + name.endswith("codebook.embed_avg") or \ + name.endswith("codebook.inited"): + logger.debug(f"Skipping {name!r}") + return [] + + logger.info(f"{self.map_tensor_name(name)} -> {data_torch.shape}") + + return [(self.map_tensor_name(name), data_torch)] + + def set_vocab(self): + self._set_vocab_none() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_vocab_size (self.hparams["vocab_size"]) + self.gguf_writer.add_features_length (self.hparams["n_embd_features"]) + self.gguf_writer.add_feed_forward_length(self.hparams["n_ff"]) + self.gguf_writer.add_group_norm_eps (self.hparams["group_norm_epsilon"]) + self.gguf_writer.add_group_norm_groups (self.hparams["group_norm_groups"]) + + self.gguf_writer.add_posnet_embedding_length(self.hparams["posnet"]["n_embd"]) + self.gguf_writer.add_posnet_block_count (self.hparams["posnet"]["n_layer"]) + + self.gguf_writer.add_convnext_embedding_length(self.hparams["convnext"]["n_embd"]) + self.gguf_writer.add_convnext_block_count (self.hparams["convnext"]["n_layer"]) + + self.gguf_writer.add_causal_attention(False) + @Model.register("Qwen2MoeForCausalLM") class Qwen2MoeModel(Model): @@ -2519,7 +2615,7 @@ class InternLM2Model(Model): return [(self.map_tensor_name(name), data_torch)] -@Model.register("BertModel", "CamembertModel") +@Model.register("BertModel", "CamembertModel", "RobertaModel") class BertModel(Model): model_arch = gguf.MODEL_ARCH.BERT @@ -2560,7 +2656,8 @@ class BertModel(Model): # we need this to validate the size of the token_type embeddings # though currently we are passing all zeros to the token_type embeddings - self.gguf_writer.add_token_type_count(2) # "Sequence A" or "Sequence B" + # "Sequence A" or "Sequence B" + self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1)) # convert to phantom space vocab def phantom(tok): @@ -3378,6 +3475,97 @@ class ArcticModel(Model): raise ValueError(f"Unprocessed experts: {experts}") +@Model.register("DeepseekForCausalLM") +class DeepseekModel(Model): + model_arch = gguf.MODEL_ARCH.DEEPSEEK + + def set_vocab(self): + try: + self._set_vocab_sentencepiece() + except FileNotFoundError: + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + if "head_dim" in hparams: + rope_dim = hparams["head_dim"] + else: + rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] + + self.gguf_writer.add_rope_dimension_count(rope_dim) + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) + self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"]) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_weights_scale(1.0) + self.gguf_writer.add_expert_count(hparams["n_routed_experts"]) + self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"]) + + _experts: list[dict[str, Tensor]] | None = None + + @staticmethod + def permute(weights: Tensor, n_head: int, n_head_kv: int | None): + if n_head_kv is not None and n_head != n_head_kv: + n_head = n_head_kv + return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape)) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + n_head = self.hparams["num_attention_heads"] + n_kv_head = self.hparams.get("num_key_value_heads") + + if name.endswith(("q_proj.weight", "q_proj.bias")): + data_torch = DeepseekModel.permute(data_torch, n_head, n_head) + if name.endswith(("k_proj.weight", "k_proj.bias")): + data_torch = DeepseekModel.permute(data_torch, n_head, n_kv_head) + + # process the experts separately + if name.find("mlp.experts") != -1: + n_experts = self.hparams["n_routed_experts"] + assert bid is not None + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + tensors: list[tuple[str, Tensor]] = [] + + # merge the experts into a single 3d tensor + for w_name in ["down_proj", "gate_proj", "up_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + return tensors + else: + return [] + + return [(self.map_tensor_name(name), data_torch)] + + def prepare_tensors(self): + super().prepare_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + @Model.register("DeepseekV2ForCausalLM") class DeepseekV2Model(Model): model_arch = gguf.MODEL_ARCH.DEEPSEEK2 diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 28cd02e5a..88058442f 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -17,7 +17,7 @@ # # python3 convert_hf_to_gguf_update.py # -# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py +# - The convert_hf_to_gguf.py script will have had its get_vocab_base_pre() function updated # - Update llama.cpp with the new pre-tokenizer if necessary # # TODO: generate tokenizer tests for llama.cpp @@ -102,6 +102,9 @@ models = [ {"name": "exaone", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", }, {"name": "phi-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", }, {"name": "chameleon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", }, + {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", }, + {"name": "roberta-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"}, + {"name": "gigachat", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct"}, ] diff --git a/docs/build.md b/docs/build.md index a4964cbd1..84019b204 100644 --- a/docs/build.md +++ b/docs/build.md @@ -55,7 +55,14 @@ cmake --build build --config Release cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF cmake --build build-arm64-windows-llvm-release ``` - Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_4_8 CPU kernels. + Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels. + + For building with ninja generator and clang compiler as default: + -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64 + ```bash + cmake --preset x64-windows-llvm-release + cmake --build build-x64-windows-llvm-release + ``` ## BLAS Build diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 9210e9fea..66cfab2c3 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -20,7 +20,12 @@ else() add_subdirectory(batched) add_subdirectory(embedding) add_subdirectory(eval-callback) - add_subdirectory(gbnf-validator) + + if (NOT WIN32) + # disabled on Windows because it uses internal functions not exported with LLAMA_API + add_subdirectory(gbnf-validator) + endif() + add_subdirectory(gguf-hash) add_subdirectory(gguf-split) add_subdirectory(gguf) @@ -46,12 +51,17 @@ else() add_subdirectory(speculative) add_subdirectory(speculative-simple) add_subdirectory(tokenize) + add_subdirectory(tts) + add_subdirectory(gen-docs) if (NOT GGML_BACKEND_DL) # these examples use the backends directly and cannot be built with dynamic loading add_subdirectory(convert-llama2c-to-ggml) add_subdirectory(cvector-generator) add_subdirectory(export-lora) - add_subdirectory(quantize-stats) + if (NOT WIN32) + # disabled on Windows because it uses internal functions not exported with LLAMA_API + add_subdirectory(quantize-stats) + endif() add_subdirectory(llava) if (GGML_RPC) add_subdirectory(rpc) diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index ba219cd4b..e2e01f2d5 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -65,6 +65,7 @@ int main(int argc, char ** argv) { llama_context * ctx = llama_new_context_with_model(model, ctx_params); auto sparams = llama_sampler_chain_default_params(); + sparams.no_perf = false; llama_sampler * smpl = llama_sampler_chain_init(sparams); diff --git a/examples/deprecation-warning/deprecation-warning.cpp b/examples/deprecation-warning/deprecation-warning.cpp index 11b35d2c2..c2958ea12 100644 --- a/examples/deprecation-warning/deprecation-warning.cpp +++ b/examples/deprecation-warning/deprecation-warning.cpp @@ -12,7 +12,7 @@ int main(int argc, char** argv) { } // Get only the program name from the full path - auto pos = filename.find_last_of('/'); + auto pos = filename.find_last_of("/\\"); if (pos != std::string::npos) { filename = filename.substr(pos+1); } diff --git a/examples/gguf-split/gguf-split.cpp b/examples/gguf-split/gguf-split.cpp index 7e62657e1..75f63f938 100644 --- a/examples/gguf-split/gguf-split.cpp +++ b/examples/gguf-split/gguf-split.cpp @@ -287,7 +287,7 @@ struct split_strategy { } void print_info() { - printf("n_split: %ld\n", ctx_outs.size()); + printf("n_split: %zu\n", ctx_outs.size()); int i_split = 0; for (auto & ctx_out : ctx_outs) { // re-calculate the real gguf size for each split (= metadata size + total size of all tensors) @@ -297,7 +297,7 @@ struct split_strategy { total_size += ggml_nbytes(t); } total_size = total_size / 1000 / 1000; // convert to megabytes - printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size); + printf("split %05d: n_tensors = %d, total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size); i_split++; } } diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index 6e42fa073..18a945b33 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -75,7 +75,7 @@ static std::vector> encode(llama_context * ctx, const std::ve } std::vector emb_norm(emb_unorm.size()); - common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd); + common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2); result.push_back(emb_norm); #ifdef GRIT_DEBUG diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index bac606f47..2338ad106 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -1521,7 +1521,7 @@ int main(int argc, char ** argv) { for (const auto & inst : params_instances) { params_idx++; if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count); + fprintf(stderr, "llama-bench: benchmark %d/%zu: starting\n", params_idx, params_count); } // keep the same model between tests when possible if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) { @@ -1573,14 +1573,14 @@ int main(int argc, char ** argv) { // warmup run if (t.n_prompt > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count); + fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count); } //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads); test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads); } if (t.n_gen > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count); + fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count); } test_gen(ctx, 1, t.n_threads); } @@ -1592,14 +1592,14 @@ int main(int argc, char ** argv) { if (t.n_prompt > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, + fprintf(stderr, "llama-bench: benchmark %d/%zu: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps); } test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads); } if (t.n_gen > 0) { if (params.progress) { - fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, + fprintf(stderr, "llama-bench: benchmark %d/%zu: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps); } test_gen(ctx, t.n_gen, t.n_threads); diff --git a/examples/llama.android/llama/build.gradle.kts b/examples/llama.android/llama/build.gradle.kts index 2d1dfba20..28dbc1904 100644 --- a/examples/llama.android/llama/build.gradle.kts +++ b/examples/llama.android/llama/build.gradle.kts @@ -19,6 +19,7 @@ android { externalNativeBuild { cmake { arguments += "-DLLAMA_BUILD_COMMON=ON" + arguments += "-DGGML_LLAMAFILE=OFF" arguments += "-DCMAKE_BUILD_TYPE=Release" cppFlags += listOf() arguments += listOf() diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 65cd4eb51..998c673d5 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -210,20 +210,20 @@ actor LlamaContext { llama_kv_cache_clear(context) - let t_pp_start = ggml_time_us() + let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000; if llama_decode(context, batch) != 0 { print("llama_decode() failed during prompt") } llama_synchronize(context) - let t_pp_end = ggml_time_us() + let t_pp_end = DispatchTime.now().uptimeNanoseconds / 1000; // bench text generation llama_kv_cache_clear(context) - let t_tg_start = ggml_time_us() + let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000; for i in 0.. PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_LDPV2, "ldpv2"}, { PROJECTOR_TYPE_RESAMPLER, "resampler"}, { PROJECTOR_TYPE_ADAPTER, "adapter"} + { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"}, }; @@ -475,7 +480,8 @@ struct clip_vision_model { // embeddings struct ggml_tensor * class_embedding; - struct ggml_tensor * patch_embeddings; + struct ggml_tensor * patch_embeddings_0; + struct ggml_tensor * patch_embeddings_1; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL) struct ggml_tensor * patch_bias; struct ggml_tensor * position_embeddings; @@ -572,6 +578,7 @@ struct clip_ctx { bool has_llava_projector = false; bool has_minicpmv_projector = false; bool has_glm_projector = false; + bool has_qwen2vl_merger = false; int minicpmv_version = 2; struct clip_vision_model vision_model; @@ -580,6 +587,7 @@ struct clip_ctx { float image_mean[3]; float image_std[3]; bool use_gelu = false; + bool use_silu = false; int32_t ftype = 1; bool has_class_embedding = true; @@ -625,14 +633,26 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 image_size_height = imgs->data->ny; } } + else if (ctx->has_qwen2vl_merger) { + // use the image's native resolution when image is avaible + if (is_inf) { + // if (imgs->data->nx && imgs->data->ny) { + image_size_width = imgs->data->nx; + image_size_height = imgs->data->ny; + } + } const int patch_size = hparams.patch_size; const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); + const int patches_w = image_size_width / patch_size; + const int patches_h = image_size_height / patch_size; const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); + const int num_position_ids = ctx->has_qwen2vl_merger ? num_positions * 4 : num_positions; const int hidden_size = hparams.hidden_size; const int n_head = hparams.n_head; const int d_head = hidden_size / n_head; int n_layer = hparams.n_layer; const float eps = hparams.eps; + int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; const int batch_size = imgs->size; @@ -653,10 +673,30 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 ggml_set_name(inp_raw, "inp_raw"); ggml_set_input(inp_raw); - struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); - inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size); - inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); + if (ctx->has_qwen2vl_merger) { + GGML_ASSERT(image_size_width % (patch_size * 2) == 0); + GGML_ASSERT(image_size_height % (patch_size * 2) == 0); + + auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + inp = ggml_add(ctx0, inp, inp_1); + inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3)); // [w, h, c, b] -> [c, w, h, b] + inp = ggml_reshape_4d( + ctx0, inp, + hidden_size * 2, patches_w / 2, patches_h, batch_size); + inp = ggml_reshape_4d( + ctx0, inp, + hidden_size * 2, patches_w / 2, 2, batch_size * (patches_h / 2)); + inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3)); + inp = ggml_reshape_3d( + ctx0, inp, + hidden_size, patches_w * patches_h, batch_size); + } + else { + inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size); + inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); + } if (ctx->has_patch_bias) { // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp)); @@ -678,12 +718,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } } - struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions); + struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids); ggml_set_name(positions, "positions"); ggml_set_input(positions); - embeddings = - ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions)); + if (!ctx->has_qwen2vl_merger) { // qwen2vl use rope position embedding + embeddings = + ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions)); + } if (ctx->has_minicpmv_projector) { int pos_w = image_size_width/patch_size; @@ -707,7 +749,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // loop over layers - if (ctx->has_minicpmv_projector || ctx->has_glm_projector) { + if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) { n_layer += 1; } for (int il = 0; il < n_layer - 1; il++) { @@ -729,8 +771,13 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b); - Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size); + if (ctx->has_qwen2vl_merger) { + Q = ggml_rope_multi( + ctx0, Q, positions, nullptr, + d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); + } + Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size); @@ -738,6 +785,11 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b); K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size); + if (ctx->has_qwen2vl_merger) { + K = ggml_rope_multi( + ctx0, K, positions, nullptr, + d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); + } K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size); @@ -777,6 +829,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 if (ctx->use_gelu) { cur = ggml_gelu_inplace(ctx0, cur); + } else if (ctx->use_silu) { + cur = ggml_silu_inplace(ctx0, cur); } else { cur = ggml_gelu_quick_inplace(ctx0, cur); } @@ -788,6 +842,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 cur = ggml_add(ctx0, embeddings, cur); embeddings = cur; + } // post-layernorm @@ -859,7 +914,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); // stride = 1, padding = 1, bias is nullptr - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); + block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); // layer norm // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] @@ -907,7 +962,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // block_2 { // stride = 2 - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); + block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] // layer norm @@ -968,7 +1023,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // mlp_2 ne [24, 24, 2048, 1] mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0); // weight ne = [3, 3, 2048, 1] - struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1); + struct ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1); peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3)); peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b); mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3)); @@ -1075,6 +1130,18 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 }else{ GGML_ABORT("fatel error"); } + else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size); + + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + + // GELU activation + embeddings = ggml_gelu(ctx0, embeddings); + + // Second linear layer + embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_1_b); } // build the graph @@ -1257,6 +1324,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->has_glm_projector = gguf_get_val_bool(ctx, idx); } + idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER); + if (idx != -1) { + new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx); + } // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search GGML_ASSERT(new_clip->has_vision_encoder); @@ -1265,6 +1336,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { idx = get_key_idx(ctx, KEY_USE_GELU); new_clip->use_gelu = gguf_get_val_bool(ctx, idx); + try { + idx = get_key_idx(ctx, KEY_USE_SILU); + new_clip->use_silu = gguf_get_val_bool(ctx, idx); + } catch (std::runtime_error & /*e*/) { + new_clip->use_silu = false; + } + if (verbosity >= 1) { LOG_INF("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder); LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder); @@ -1441,11 +1519,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } try { - vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); + vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); } catch(const std::exception& /*e*/) { LOG_ERR("%s: failed to load vision model tensors\n", __func__); } + try { + vision_model.patch_embeddings_1 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1); + } catch(const std::exception& /*e*/) { + new_clip->has_qwen2vl_merger = false; + } // LLaVA projection if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) { @@ -1544,6 +1627,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_4H_2_H,"weight")); vision_model.boi_w = get_tensor(new_clip->ctx_data, TN_GLM_BOI_W); vision_model.eoi_w = get_tensor(new_clip->ctx_data, TN_GLM_EOI_W); + else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) { + vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); + vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); + vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); } else { std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; @@ -1583,6 +1671,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend)); clip_image_f32_batch batch; batch.size = 1; + batch.data = nullptr; ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false); ggml_gallocr_reserve(new_clip->compute_alloc, gf); size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0); @@ -1596,6 +1685,10 @@ void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size ctx_clip->load_image_size = load_image_size; } +struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) { + return ctx_clip->load_image_size; +} + struct clip_image_size * clip_image_size_init() { struct clip_image_size * load_image_size = new struct clip_image_size(); load_image_size->width = 448; @@ -2048,6 +2141,23 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli } return true; } + else if (ctx->has_qwen2vl_merger) { + clip_image_u8 * resized = clip_image_u8_init(); + auto patch_size = clip_patch_size(ctx) * 2; + int nx = ceil((float)img->nx / patch_size) * patch_size; + int ny = ceil((float)img->ny / patch_size) * patch_size; + bicubic_resize(*img, *resized, nx, ny); + + res_imgs->data = new clip_image_f32[1]; + // clip_image_f32 * res = clip_image_f32_init(); + normalize_image_u8_to_f32(resized, res_imgs->data, ctx->image_mean, ctx->image_std); + // res_imgs->data[0] = *res; + res_imgs->size = 1; + + // clip_image_f32_free(res); + clip_image_u8_free(resized); + return true; + } if(ctx->has_glm_projector){ res_imgs->size = 1; @@ -2253,6 +2363,13 @@ size_t clip_embd_nbytes(const struct clip_ctx * ctx) { return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float); } +size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) { + clip_image_f32 img; + img.nx = img_w; + img.ny = img_h; + return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float); +} + int32_t clip_image_size(const struct clip_ctx * ctx) { return ctx->vision_model.hparams.image_size; } @@ -2274,6 +2391,13 @@ const int32_t * clip_image_grid(const struct clip_ctx * ctx) { } int clip_n_patches(const struct clip_ctx * ctx) { + clip_image_f32 img; + img.nx = ctx->vision_model.hparams.image_size; + img.ny = ctx->vision_model.hparams.image_size; + return clip_n_patches_by_img(ctx, &img); +} + +int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) { const auto & params = ctx->vision_model.hparams; int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); @@ -2287,6 +2411,11 @@ int clip_n_patches(const struct clip_ctx * ctx) { else if (ctx->minicpmv_version == 3) { n_patches = 64; } + } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + int patch_size = params.patch_size * 2; + int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0); + int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0); + n_patches = x_patch * y_patch; } return n_patches; @@ -2421,7 +2550,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima const int image_size = hparams.image_size; int image_size_width = image_size; int image_size_height = image_size; - if (ctx->has_minicpmv_projector) { + if (ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger) { image_size_width = imgs->data[0].nx; image_size_height = imgs->data[0].ny; } @@ -2441,7 +2570,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima for (size_t i = 0; i < imgs->size; i++) { const int nx = imgs->data[i].nx; const int ny = imgs->data[i].ny; - if (!ctx->has_minicpmv_projector) { + if (!(ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger)) { GGML_ASSERT(nx == image_size && ny == image_size); } @@ -2499,9 +2628,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h)); float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed)); - for(int i=0;ihas_qwen2vl_merger) { + struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + + const int pw = image_size_width / patch_size; + const int ph = image_size_height / patch_size; + int* positions_data = (int*)malloc(ggml_nbytes(positions)); + + int ptr = 0; + for (int y = 0; y < ph; y+=2) + { + for (int x = 0; x < pw; x+=2) + { + for (int dy = 0; dy < 2; dy++) { + for (int dx = 0; dx < 2; dx++) { + positions_data[ptr] = y + dy; + positions_data[num_patches + ptr] = x + dx; + positions_data[num_patches * 2 + ptr] = y + dy; + positions_data[num_patches * 3 + ptr] = x + dx; + ptr++; + } + } + } + } + + ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); + free(positions_data); + } + else { struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); int* positions_data = (int*)malloc(ggml_nbytes(positions)); @@ -2530,16 +2686,16 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); free(positions_data); - } - if (!ctx->has_glm_projector){ - struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); - int* patches_data = (int*)malloc(ggml_nbytes(patches)); - for (int i = 0; i < num_patches; i++) { - patches_data[i] = i + 1; + if (!ctx->has_glm_projector){ + struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); + int* patches_data = (int*)malloc(ggml_nbytes(patches)); + for (int i = 0; i < num_patches; i++) { + patches_data[i] = i + 1; + } + ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); + free(patches_data); } - ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); - free(patches_data); } } @@ -2722,6 +2878,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){ return ctx->vision_model.mm_model_mlp_3_w->ne[1]; } + if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + return ctx->vision_model.mm_1_b->ne[0]; + } std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); @@ -2737,3 +2896,20 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) { bool clip_is_glm(const struct clip_ctx * ctx) { return ctx->has_glm_projector; } +bool clip_is_qwen2vl(const struct clip_ctx * ctx) { + return ctx->has_qwen2vl_merger; +} + + +bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) { + clip_image_f32 clip_img; + clip_img.buf.resize(h * w * 3); + for (int i = 0; i < h*w*3; i++) + { + clip_img.buf[i] = img[i]; + } + clip_img.nx = w; + clip_img.ny = h; + clip_image_encode(ctx, n_threads, &clip_img, vec); + return true; +} diff --git a/examples/llava/clip.h b/examples/llava/clip.h index c92cb5dee..841b4f6f9 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -45,6 +45,7 @@ CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity CLIP_API void clip_free(struct clip_ctx * ctx); CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx); +CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w); CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx); CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx); @@ -55,11 +56,13 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx); CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); -CLIP_API int clip_n_patches (const struct clip_ctx * ctx); -CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); +CLIP_API int clip_n_patches (const struct clip_ctx * ctx); +CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img); +CLIP_API int clip_n_mmproj_embd (const struct clip_ctx * ctx); CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip); CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size); +CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip); CLIP_API struct clip_image_size * clip_image_size_init(); CLIP_API struct clip_image_u8 * clip_image_u8_init (); @@ -86,6 +89,9 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype); CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx); +CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx); + +CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec); CLIP_API bool clip_is_glm(const struct clip_ctx * ctx); diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index 784291c02..624154fc4 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -259,25 +259,33 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); - if (clip_is_minicpmv(ctx_clip)) { + if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) { std::vector image_embd_v; image_embd_v.resize(img_res_v.size); struct clip_image_size * load_image_size = clip_image_size_init(); + for (size_t i = 0; i < img_res_v.size; i++) { const int64_t t_img_enc_step_start_us = ggml_time_us(); - image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); + image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny)); int patch_size=14; load_image_size->width = img_res_v.data[i].nx; load_image_size->height = img_res_v.data[i].ny; clip_add_load_image_size(ctx_clip, load_image_size); + bool encoded = false; - int has_minicpmv_projector = clip_is_minicpmv(ctx_clip); - if (has_minicpmv_projector == 2) { - encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); - } - else if (has_minicpmv_projector == 3) { + if (clip_is_qwen2vl(ctx_clip)) { encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); } + else { + int has_minicpmv_projector = clip_is_minicpmv(ctx_clip); + if (has_minicpmv_projector == 2) { + encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); + } + else if (has_minicpmv_projector == 3) { + encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); + } + } + if (!encoded) { LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); return false; @@ -290,8 +298,11 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli int n_img_pos_out = 0; for (size_t i = 0; i < image_embd_v.size(); i++) { - std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip)); - n_img_pos_out += clip_n_patches(ctx_clip); + std::memcpy( + image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), + image_embd_v[i], + clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny)); + n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]); } *n_img_pos = n_img_pos_out; for (size_t i = 0; i < image_embd_v.size(); i++) { @@ -405,6 +416,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co num_max_patches = 1; } float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model + float * image_embd; + if (clip_is_qwen2vl(ctx_clip)) { + // qwen2vl don't split image into chunks, so `num_max_patches` is not needed. + image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny)); + } else { + image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model + } if (!image_embd) { LOG_ERR("Unable to allocate memory for image embeddings\n"); return false; diff --git a/examples/llava/qwen2_vl_surgery.py b/examples/llava/qwen2_vl_surgery.py new file mode 100644 index 000000000..c87606b4f --- /dev/null +++ b/examples/llava/qwen2_vl_surgery.py @@ -0,0 +1,165 @@ +import argparse +from typing import Dict + +import torch +import numpy as np +from gguf import * +from transformers import ( + Qwen2VLForConditionalGeneration, + Qwen2VLProcessor, + AutoProcessor, + Qwen2VLConfig +) + + +VISION = "clip.vision" + + +def k(raw_key: str, arch: str) -> str: + return raw_key.format(arch=arch) + + +def to_gguf_name(name: str) -> str: + og = name + name = name.replace("text_model", "t").replace("vision_model", "v") + name = name.replace("blocks", "blk").replace("embeddings.", "") + name = name.replace("attn.", "attn_") + name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.") + # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln") + name = name.replace("norm1", "ln1").replace("norm2", "ln2") + name = name.replace("merger.mlp", 'mm') + print(f"[to_gguf_name] {og} --> {name}") + return name + + +def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]: + vision_model = qwen2vl.visual + tensor_map = {} + for name, ten in vision_model.state_dict().items(): + ten = ten.numpy() + if 'qkv' in name: + if ten.ndim == 2: # weight + c3, _ = ten.shape + else: # bias + c3 = ten.shape[0] + assert c3 % 3 == 0 + c = c3 // 3 + wq = ten[:c] + wk = ten[c: c * 2] + wv = ten[c * 2:] + tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq + tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk + tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv + elif 'merger' in name: + if name.endswith("ln_q.weight"): + tensor_map['v.post_ln.weight'] = ten + elif name.endswith("ln_q.bias"): + tensor_map['v.post_ln.bias'] = ten + else: + # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias" + tensor_map[to_gguf_name(name)] = ten + elif 'patch_embed.proj.weight' in name: + # NOTE: split Conv3D into Conv2Ds + c1, c2, kt, kh, kw = ten.shape + assert kt == 2, "Current implmentation only support temporal_patch_size of 2" + tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...] + tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...] + else: + tensor_map[to_gguf_name(f"vision_model.{name}")] = ten + + for new_name, ten in tensor_map.items(): + if ten.ndim <= 1 or new_name.endswith("_norm.weight"): + tensor_map[new_name] = ten.astype(np.float32) + else: + tensor_map[new_name] = ten.astype(dtype) + tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32) # dummy tensor, just here as a placeholder + return tensor_map + + +def main(args): + if args.data_type == 'fp32': + dtype = torch.float32 + np_dtype = np.float32 + ftype = 0 + elif args.data_type == 'fp16': + dtype = torch.float32 + np_dtype = np.float16 + ftype = 1 + else: + raise ValueError() + + local_model = False + model_path = "" + model_name = args.model_name + print("model_name: ", model_name) + qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained( + model_name, torch_dtype=dtype, device_map="cpu" + ) + cfg: Qwen2VLConfig = qwen2vl.config # type: ignore[reportAssignmentType] + vcfg = cfg.vision_config + + if os.path.isdir(model_name): + local_model = True + if model_name.endswith(os.sep): + model_name = model_name[:-1] + model_path = model_name + model_name = os.path.basename(model_name) + fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf" + + fout = GGUFWriter(path=fname_out, arch="clip") + fout.add_description("image encoder for Qwen2VL") + + fout.add_file_type(ftype) + fout.add_bool("clip.has_text_encoder", False) + fout.add_bool("clip.has_vision_encoder", True) + fout.add_bool("clip.has_qwen2vl_merger", True) + fout.add_string("clip.projector_type", "qwen2vl_merger") + + print(cfg.vision_config) + if 'silu' in cfg.vision_config.hidden_act.lower(): + fout.add_bool("clip.use_silu", True) + fout.add_bool("clip.use_gelu", False) + elif 'gelu' in cfg.vision_config.hidden_act.lower(): + fout.add_bool("clip.use_silu", False) + fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower()) + else: + raise ValueError() + + tensor_map = find_vision_tensors(qwen2vl, np_dtype) + for name, data in tensor_map.items(): + fout.add_tensor(name, data) + + fout.add_uint32("clip.vision.patch_size", vcfg.patch_size) + fout.add_uint32("clip.vision.image_size", 14 * 40) # some reasonable size that is divable by (14*2) + fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim) + fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size) + fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads) + fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6) + fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth) + fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0) # not sure what this does, put 0 here as a placeholder + fout.add_name(model_name) + """ + HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig, + it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`. + """ + + if local_model: + processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path) + else: + processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name) + fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue] + fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue] + + fout.write_header_to_file() + fout.write_kv_data_to_file() + fout.write_tensors_to_file() + fout.close() + print("save model as: ", fname_out) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct") + parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32") + args = parser.parse_args() + main(args) diff --git a/examples/llava/qwen2vl-cli.cpp b/examples/llava/qwen2vl-cli.cpp new file mode 100644 index 000000000..e86a60280 --- /dev/null +++ b/examples/llava/qwen2vl-cli.cpp @@ -0,0 +1,581 @@ +#include "arg.h" +#include "base64.hpp" +#include "log.h" +#include "common.h" +#include "sampling.h" +#include "clip.h" +#include "llava.h" +#include "llama.h" +#include "ggml.h" + +#ifdef GGML_USE_CUDA +#include "ggml-cuda.h" +#endif +#ifdef NDEBUG +#include "ggml-alloc.h" +#include "ggml-backend.h" +#endif + +#include +#include +#include +#include +#include +#include +#include + + +static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, + int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) { + int n_embd = llama_n_embd(llama_get_model(ctx_llama)); + const int patch_size = 14 * 2; + const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0); + const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0); + auto img_tokens = image_embed->n_image_pos; + // llama_pos mrope_pos[img_tokens * 4]; + std::vector mrope_pos; + mrope_pos.resize(img_tokens * 4); + + for (int y = 0; y < ph; y++) + { + for (int x = 0; x < pw; x++) + { + int i = y * pw + x; + mrope_pos[i] = *st_pos_id; + mrope_pos[i + img_tokens] = *st_pos_id + y; + mrope_pos[i + img_tokens * 2] = *st_pos_id + x; + mrope_pos[i + img_tokens * 3] = 0; + } + } + *st_pos_id += std::max(pw, ph); + + int processed = 0; + std::vector batch_mrope_pos; + batch_mrope_pos.resize(img_tokens * 4); + + for (int i = 0; i < img_tokens; i += n_batch) { + int n_eval = img_tokens - i; + if (n_eval > n_batch) { + n_eval = n_batch; + } + + // llama_pos batch_mrope_pos[n_eval * 4]; + std::fill(batch_mrope_pos.begin(), batch_mrope_pos.end(), 0); + memcpy(batch_mrope_pos.data(), &mrope_pos[processed], n_eval * sizeof(llama_pos)); + memcpy(&batch_mrope_pos[n_eval * 1], &mrope_pos[img_tokens * 1 + processed], n_eval * sizeof(llama_pos)); + memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos)); + memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos)); + + llama_batch batch = { + int32_t(n_eval), // n_tokens + nullptr, // token + (image_embed->embed+i*n_embd), // embed + batch_mrope_pos.data(), // pos + nullptr, // n_seq_id + nullptr, // seq_id + nullptr, // logits + }; + + if (llama_decode(ctx_llama, batch)) { + LOG_ERR("%s : failed to eval\n", __func__); + return false; + } + *n_past += n_eval; + processed += n_eval; + } + return true; +} + + +static bool eval_tokens(struct llama_context * ctx_llama, std::vector tokens, int n_batch, int * n_past, int * st_pos_id) { + int N = (int) tokens.size(); + std::vector pos; + for (int i = 0; i < N; i += n_batch) { + int n_eval = (int) tokens.size() - i; + if (n_eval > n_batch) { + n_eval = n_batch; + } + auto batch = llama_batch_get_one(&tokens[i], n_eval); + // TODO: add mrope pos ids somewhere else + pos.resize(batch.n_tokens * 4); + std::fill(pos.begin(), pos.end(), 0); + for (int j = 0; j < batch.n_tokens * 3; j ++) { + pos[j] = *st_pos_id + (j % batch.n_tokens); + } + batch.pos = pos.data(); + + if (llama_decode(ctx_llama, batch)) { + LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past); + return false; + } + *n_past += n_eval; + *st_pos_id += n_eval; + } + return true; +} + +static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past, int * st_pos_id) { + std::vector tokens; + tokens.push_back(id); + return eval_tokens(ctx_llama, tokens, 1, n_past, st_pos_id); +} + +static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, int * st_pos_id, bool add_bos){ + std::string str2 = str; + std::vector embd_inp = common_tokenize(ctx_llama, str2, add_bos, true); + eval_tokens(ctx_llama, embd_inp, n_batch, n_past, st_pos_id); + return true; +} + +static const char * sample(struct common_sampler * smpl, + struct llama_context * ctx_llama, + int * n_past, int * st_pos_id) { + const llama_token id = common_sampler_sample(smpl, ctx_llama, -1); + common_sampler_accept(smpl, id, true); + static std::string ret; + if (llama_token_is_eog(llama_get_model(ctx_llama), id)) { + ret = ""; + } else { + ret = common_token_to_piece(ctx_llama, id); + } + eval_id(ctx_llama, id, n_past, st_pos_id); + return ret.c_str(); +} + +static const char* IMG_BASE64_TAG_BEGIN = ""; + +static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) { + begin_out = prompt.find(IMG_BASE64_TAG_BEGIN); + end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out); +} + +static bool prompt_contains_image(const std::string& prompt) { + size_t begin, end; + find_image_tag_in_prompt(prompt, begin, end); + return (begin != std::string::npos); +} + +// replaces the base64 image tag in the prompt with `replacement` +static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) { + size_t img_base64_str_start, img_base64_str_end; + find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end); + if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) { + LOG_ERR("%s: invalid base64 image tag. must be %s%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END); + return NULL; + } + + auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN); + auto base64_bytes_count = img_base64_str_end - base64_bytes_start; + auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count ); + + auto required_bytes = base64::required_encode_size(base64_str.size()); + auto img_bytes = std::vector(required_bytes); + base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin()); + + auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size()); + if (!embed) { + LOG_ERR("%s: could not load image from base64 string.\n", __func__); + return NULL; + } + + return embed; +} + +static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") { + size_t begin, end; + find_image_tag_in_prompt(prompt, begin, end); + if (begin == std::string::npos || end == std::string::npos) { + return prompt; + } + auto pre = prompt.substr(0, begin); + auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END)); + return pre + replacement + post; +} + +struct llava_context { + struct clip_ctx * ctx_clip = NULL; + struct llama_context * ctx_llama = NULL; + struct llama_model * model = NULL; +}; + +static void print_usage(int, char ** argv) { + LOG("\n example usage:\n"); + LOG("\n %s -m --mmproj --image --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); + LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n"); +} + +static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) { + + // load and preprocess the image + llava_image_embed * embed = NULL; + auto prompt = params->prompt; + if (prompt_contains_image(prompt)) { + if (!params->image.empty()) { + LOG_INF("using base64 encoded image instead of command line image path\n"); + } + embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt); + if (!embed) { + LOG_ERR("%s: can't load image from prompt\n", __func__); + return NULL; + } + params->prompt = remove_image_from_prompt(prompt); + } else { + embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str()); + if (!embed) { + fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str()); + return NULL; + } + } + + return embed; +} + +static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) { + int n_past = 0; + int cur_pos_id = 0; + + const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict; + + std::string system_prompt, user_prompt; + size_t image_pos = prompt.find("<|vision_start|>"); + if (image_pos != std::string::npos) { + // new templating mode: Provide the full prompt including system message and use as a placeholder for the image + system_prompt = prompt.substr(0, image_pos); + user_prompt = prompt.substr(image_pos + std::string("<|vision_pad|>").length()); + LOG_INF("system_prompt: %s\n", system_prompt.c_str()); + if (params->verbose_prompt) { + auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } + LOG_INF("user_prompt: %s\n", user_prompt.c_str()); + if (params->verbose_prompt) { + auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } + } else { + // llava-1.5 native mode + system_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|>"; + user_prompt = "<|vision_end|>" + prompt + "<|im_end|>\n<|im_start|>assistant\n"; + if (params->verbose_prompt) { + auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true); + for (int i = 0; i < (int) tmp.size(); i++) { + LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str()); + } + } + } + + eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, true); + if (image_embed != nullptr) { + auto image_size = clip_get_load_image_size(ctx_llava->ctx_clip); + qwen2vl_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past, &cur_pos_id, image_size); + } + eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, false); + + // generate the response + + LOG("\n"); + + struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling); + if (!smpl) { + LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__); + exit(1); + } + + std::string response = ""; + for (int i = 0; i < max_tgt_len; i++) { + const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past, &cur_pos_id); + response += tmp; + if (strcmp(tmp, "") == 0) break; + if (strstr(tmp, "###")) break; // Yi-VL behavior + LOG("%s", tmp); + if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works) + if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6 + if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6 + + fflush(stdout); + } + + common_sampler_free(smpl); + LOG("\n"); +} + +static struct llama_model * llava_init(common_params * params) { + llama_backend_init(); + llama_numa_init(params->numa); + + llama_model_params model_params = common_model_params_to_llama(*params); + + llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params); + if (model == NULL) { + LOG_ERR("%s: unable to load model\n" , __func__); + return NULL; + } + return model; +} + +static struct llava_context * llava_init_context(common_params * params, llama_model * model) { + const char * clip_path = params->mmproj.c_str(); + + auto prompt = params->prompt; + if (prompt.empty()) { + prompt = "describe the image in detail."; + } + + auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + + + llama_context_params ctx_params = common_context_params_to_llama(*params); + ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings + + llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params); + + if (ctx_llama == NULL) { + LOG_ERR("%s: failed to create the llama_context\n" , __func__); + return NULL; + } + + auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context)); + + ctx_llava->ctx_llama = ctx_llama; + ctx_llava->ctx_clip = ctx_clip; + ctx_llava->model = model; + return ctx_llava; +} + +static void llava_free(struct llava_context * ctx_llava) { + if (ctx_llava->ctx_clip) { + clip_free(ctx_llava->ctx_clip); + ctx_llava->ctx_clip = NULL; + } + + llama_free(ctx_llava->ctx_llama); + llama_free_model(ctx_llava->model); + llama_backend_free(); +} + +#ifndef NDEBUG + +static void debug_test_mrope_2d() { + // 1. Initialize backend + ggml_backend_t backend = NULL; + std::string backend_name = ""; +#ifdef GGML_USE_CUDA + fprintf(stderr, "%s: using CUDA backend\n", __func__); + backend = ggml_backend_cuda_init(0); // init device 0 + backend_name = "cuda"; + if (!backend) { + fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__); + } +#endif + // if there aren't GPU Backends fallback to CPU backend + if (!backend) { + backend = ggml_backend_cpu_init(); + backend_name = "cpu"; + } + + // Calculate the size needed to allocate + size_t ctx_size = 0; + ctx_size += 2 * ggml_tensor_overhead(); // tensors + // no need to allocate anything else! + + // 2. Allocate `ggml_context` to store tensor data + struct ggml_init_params params = { + /*.mem_size =*/ ctx_size, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, // the tensors will be allocated later by ggml_backend_alloc_ctx_tensors() + }; + struct ggml_context * ctx = ggml_init(params); + + struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 128, 12, 30); + ggml_set_name(inp_raw, "inp_raw"); + ggml_set_input(inp_raw); + + struct ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 30 * 4); + ggml_set_name(pos, "pos"); + ggml_set_input(pos); + + std::vector dummy_q; + dummy_q.resize(128 * 12 * 30); + std::fill(dummy_q.begin(), dummy_q.end(), 0.1); + // memcpy(inp_raw->data, dummy_q.data(), 128 * 12 * 30 * ggml_element_size(inp_raw)); + + std::vector pos_id; + pos_id.resize(30 * 4); + for (int i = 0; i < 30; i ++) { + pos_id[i] = i; + pos_id[i + 30] = i + 10; + pos_id[i + 60] = i + 20; + pos_id[i + 90] = i + 30; + } + int sections[4] = {32, 32, 0, 0}; + + // 4. Allocate a `ggml_backend_buffer` to store all tensors + ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend); + + // 5. Copy tensor data from main memory (RAM) to backend buffer + ggml_backend_tensor_set(inp_raw, dummy_q.data(), 0, ggml_nbytes(inp_raw)); + ggml_backend_tensor_set(pos, pos_id.data(), 0, ggml_nbytes(pos)); + + // 6. Create a `ggml_cgraph` for mul_mat operation + struct ggml_cgraph * gf = NULL; + struct ggml_context * ctx_cgraph = NULL; + + // create a temporally context to build the graph + struct ggml_init_params params0 = { + /*.mem_size =*/ ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph() + }; + ctx_cgraph = ggml_init(params0); + gf = ggml_new_graph(ctx_cgraph); + + struct ggml_tensor * result0 = ggml_rope_multi( + ctx_cgraph, inp_raw, pos, nullptr, + 128/2, sections, LLAMA_ROPE_TYPE_VISION, 32768, 1000000, 1, + 0, 1, 32, 1); + + // Add "result" tensor and all of its dependencies to the cgraph + ggml_build_forward_expand(gf, result0); + + // 7. Create a `ggml_gallocr` for cgraph computation + ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend)); + ggml_gallocr_alloc_graph(allocr, gf); + + // 9. Run the computation + int n_threads = 1; // Optional: number of threads to perform some operations with multi-threading + if (ggml_backend_is_cpu(backend)) { + ggml_backend_cpu_set_n_threads(backend, n_threads); + } + ggml_backend_graph_compute(backend, gf); + + // 10. Retrieve results (output tensors) + // in this example, output tensor is always the last tensor in the graph + struct ggml_tensor * result = result0; + // struct ggml_tensor * result = gf->nodes[gf->n_nodes - 1]; + float * result_data = (float *)malloc(ggml_nbytes(result)); + // because the tensor data is stored in device buffer, we need to copy it back to RAM + ggml_backend_tensor_get(result, result_data, 0, ggml_nbytes(result)); + const std::string bin_file = "mrope_2d_" + backend_name +".bin"; + std::ofstream outFile(bin_file, std::ios::binary); + + if (outFile.is_open()) { + outFile.write(reinterpret_cast(result_data), ggml_nbytes(result)); + outFile.close(); + std::cout << "Data successfully written to " + bin_file << std::endl; + } else { + std::cerr << "Error opening file!" << std::endl; + } + + free(result_data); + // 11. Free memory and exit + ggml_free(ctx_cgraph); + ggml_gallocr_free(allocr); + ggml_free(ctx); + ggml_backend_buffer_free(buffer); + ggml_backend_free(backend); +} + +static void debug_dump_img_embed(struct llava_context * ctx_llava) { + int n_embd = llama_n_embd(llama_get_model(ctx_llava->ctx_llama)); + int ne = n_embd * 4; + float vals[56 * 56 * 3]; + // float embd[ne]; + std::vector embd; + embd.resize(ne); + + for (int i = 0; i < 56*56; i++) + { + for (int c = 0; c < 3; c++) + vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56); + } + + clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data()); + + std::ofstream outFile("img_embed.bin", std::ios::binary); + if (outFile.is_open()) { + outFile.write(reinterpret_cast(embd.data()), ne * sizeof(float)); + + outFile.close(); + std::cout << "Data successfully written to mrope.bin" << std::endl; + } else { + std::cerr << "Error opening file!" << std::endl; + } +} + +#endif + + +int main(int argc, char ** argv) { + ggml_time_init(); + + common_params params; + + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) { + return 1; + } + + common_init(); + + if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { + print_usage(argc, argv); + return 1; + } + + auto * model = llava_init(¶ms); + if (model == NULL) { + fprintf(stderr, "%s: error: failed to init llava model\n", __func__); + return 1; + } + + if (prompt_contains_image(params.prompt)) { + auto * ctx_llava = llava_init_context(¶ms, model); + + auto * image_embed = load_image(ctx_llava, ¶ms, ""); + + // process the prompt + process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); + + llama_perf_context_print(ctx_llava->ctx_llama); + llava_image_embed_free(image_embed); + ctx_llava->model = NULL; + llava_free(ctx_llava); +#ifndef NDEBUG + } else if (params.image[0].empty()) { + auto ctx_llava = llava_init_context(¶ms, model); + + debug_test_mrope_2d(); + debug_dump_img_embed(ctx_llava); + + llama_perf_context_print(ctx_llava->ctx_llama); + ctx_llava->model = NULL; + llava_free(ctx_llava); +#endif + } else { + for (auto & image : params.image) { + auto * ctx_llava = llava_init_context(¶ms, model); + + auto * image_embed = load_image(ctx_llava, ¶ms, image); + if (!image_embed) { + LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str()); + return 1; + } + + // process the prompt + process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); + + llama_perf_context_print(ctx_llava->ctx_llama); + llava_image_embed_free(image_embed); + ctx_llava->model = NULL; + llava_free(ctx_llava); + } + } + + llama_free_model(model); + + return 0; +} diff --git a/examples/main/README.md b/examples/main/README.md index 7787f7b11..17d80a622 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -177,16 +177,11 @@ Example usage: `--temp 0` - `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled). - `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size). -- `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty. The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1. The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`). -Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases. - -Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl` - ### DRY Repetition Penalty DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)). diff --git a/examples/quantize/README.md b/examples/quantize/README.md index 704f0d56b..f9cce7b21 100644 --- a/examples/quantize/README.md +++ b/examples/quantize/README.md @@ -54,8 +54,6 @@ As the models are currently fully loaded into memory, you will need adequate dis Several quantization methods are supported. They differ in the resulting model disk size and inference speed. -The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format. - *(outdated)* | Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | @@ -83,7 +81,7 @@ The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interle - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930) - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957) - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969) - - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996) + - [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996) - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060) - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196) - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index b98993210..8d47b17b6 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -48,9 +48,6 @@ static const std::vector QUANT_OPTIONS = { { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, - { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp index e78a8596d..a5c6fe7e5 100644 --- a/examples/retrieval/retrieval.cpp +++ b/examples/retrieval/retrieval.cpp @@ -107,7 +107,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu } float * out = output + batch.seq_id[i][0] * n_embd; - common_embd_normalize(embd, out, n_embd); + common_embd_normalize(embd, out, n_embd, 2); } } @@ -143,7 +143,7 @@ int main(int argc, char ** argv) { std::vector file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator); chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end()); } - LOG_INF("Number of chunks: %ld\n", chunks.size()); + LOG_INF("Number of chunks: %zu\n", chunks.size()); llama_backend_init(); llama_numa_init(params.numa); diff --git a/examples/run/CMakeLists.txt b/examples/run/CMakeLists.txt index 52add51ef..0686d6305 100644 --- a/examples/run/CMakeLists.txt +++ b/examples/run/CMakeLists.txt @@ -1,5 +1,5 @@ set(TARGET llama-run) add_executable(${TARGET} run.cpp) install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_17) diff --git a/examples/run/README.md b/examples/run/README.md index 6e926811f..874293516 100644 --- a/examples/run/README.md +++ b/examples/run/README.md @@ -3,5 +3,47 @@ The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models. ```bash -./llama-run Meta-Llama-3.1-8B-Instruct.gguf -... +llama-run granite-code +``` + +```bash +llama-run -h +Description: + Runs a llm + +Usage: + llama-run [options] model [prompt] + +Options: + -c, --context-size + Context size (default: 2048) + -n, --ngl + Number of GPU layers (default: 0) + -v, --verbose, --log-verbose + Set verbosity level to infinity (i.e. log all messages, useful for debugging) + -h, --help + Show help message + +Commands: + model + Model is a string with an optional prefix of + huggingface:// (hf://), ollama://, https:// or file://. + If no protocol is specified and a file exists in the specified + path, file:// is assumed, otherwise if a file does not exist in + the specified path, ollama:// is assumed. Models that are being + pulled are downloaded with .partial extension while being + downloaded and then renamed as the file without the .partial + extension when complete. + +Examples: + llama-run llama3 + llama-run ollama://granite-code + llama-run ollama://smollm:135m + llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf + llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf + llama-run https://example.com/some-file1.gguf + llama-run some-file2.gguf + llama-run file://some-file3.gguf + llama-run --ngl 999 some-file4.gguf + llama-run --ngl 999 some-file5.gguf Hello World +``` diff --git a/examples/run/run.cpp b/examples/run/run.cpp index cac2faefc..03da54ca3 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -1,128 +1,501 @@ #if defined(_WIN32) -#include +# include #else -#include +# include +# include +# include +#endif + +#if defined(LLAMA_USE_CURL) +# include #endif #include +#include #include #include +#include #include #include #include -#include #include +#include "common.h" +#include "json.hpp" #include "llama-cpp.h" -typedef std::unique_ptr char_array_ptr; +GGML_ATTRIBUTE_FORMAT(1, 2) +static std::string fmt(const char * fmt, ...) { + va_list ap; + va_list ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + const int size = vsnprintf(NULL, 0, fmt, ap); + GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT + std::string buf; + buf.resize(size); + const int size2 = vsnprintf(const_cast(buf.data()), buf.size() + 1, fmt, ap2); + GGML_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); -struct Argument { - std::string flag; - std::string help_text; -}; + return buf; +} -struct Options { - std::string model_path, prompt_non_interactive; - int ngl = 99; - int n_ctx = 2048; -}; +GGML_ATTRIBUTE_FORMAT(1, 2) +static int printe(const char * fmt, ...) { + va_list args; + va_start(args, fmt); + const int ret = vfprintf(stderr, fmt, args); + va_end(args); -class ArgumentParser { - public: - ArgumentParser(const char * program_name) : program_name(program_name) {} + return ret; +} - void add_argument(const std::string & flag, std::string & var, const std::string & help_text = "") { - string_args[flag] = &var; - arguments.push_back({flag, help_text}); +class Opt { + public: + int init(int argc, const char ** argv) { + // Parse arguments + if (parse(argc, argv)) { + printe("Error: Failed to parse arguments.\n"); + help(); + return 1; + } + + // If help is requested, show help and exit + if (help_) { + help(); + return 2; + } + + return 0; // Success } - void add_argument(const std::string & flag, int & var, const std::string & help_text = "") { - int_args[flag] = &var; - arguments.push_back({flag, help_text}); + std::string model_; + std::string user_; + int context_size_ = -1, ngl_ = -1; + bool verbose_ = false; + + private: + bool help_ = false; + + bool parse_flag(const char ** argv, int i, const char * short_opt, const char * long_opt) { + return strcmp(argv[i], short_opt) == 0 || strcmp(argv[i], long_opt) == 0; + } + + int handle_option_with_value(int argc, const char ** argv, int & i, int & option_value) { + if (i + 1 >= argc) { + return 1; + } + + option_value = std::atoi(argv[++i]); + return 0; } int parse(int argc, const char ** argv) { - for (int i = 1; i < argc; ++i) { - std::string arg = argv[i]; - if (string_args.count(arg)) { - if (i + 1 < argc) { - *string_args[arg] = argv[++i]; - } else { - fprintf(stderr, "error: missing value for %s\n", arg.c_str()); - print_usage(); + bool options_parsing = true; + for (int i = 1, positional_args_i = 0; i < argc; ++i) { + if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) { + if (handle_option_with_value(argc, argv, i, context_size_) == 1) { return 1; } - } else if (int_args.count(arg)) { - if (i + 1 < argc) { - if (parse_int_arg(argv[++i], *int_args[arg]) != 0) { - fprintf(stderr, "error: invalid value for %s: %s\n", arg.c_str(), argv[i]); - print_usage(); - return 1; - } - } else { - fprintf(stderr, "error: missing value for %s\n", arg.c_str()); - print_usage(); + } else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) { + if (handle_option_with_value(argc, argv, i, ngl_) == 1) { + return 1; + } + } else if (options_parsing && + (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) { + verbose_ = true; + } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) { + help_ = true; + return 0; + } else if (options_parsing && strcmp(argv[i], "--") == 0) { + options_parsing = false; + } else if (positional_args_i == 0) { + if (!argv[i][0] || argv[i][0] == '-') { return 1; } - } else { - fprintf(stderr, "error: unrecognized argument %s\n", arg.c_str()); - print_usage(); - return 1; - } - } - if (string_args["-m"]->empty()) { - fprintf(stderr, "error: -m is required\n"); - print_usage(); - return 1; + ++positional_args_i; + model_ = argv[i]; + } else if (positional_args_i == 1) { + ++positional_args_i; + user_ = argv[i]; + } else { + user_ += " " + std::string(argv[i]); + } } return 0; } - private: - const char * program_name; - std::unordered_map string_args; - std::unordered_map int_args; - std::vector arguments; - - int parse_int_arg(const char * arg, int & value) { - char * end; - const long val = std::strtol(arg, &end, 10); - if (*end == '\0' && val >= INT_MIN && val <= INT_MAX) { - value = static_cast(val); - return 0; - } - return 1; - } - - void print_usage() const { - printf("\nUsage:\n"); - printf(" %s [OPTIONS]\n\n", program_name); - printf("Options:\n"); - for (const auto & arg : arguments) { - printf(" %-10s %s\n", arg.flag.c_str(), arg.help_text.c_str()); - } - - printf("\n"); + void help() const { + printf( + "Description:\n" + " Runs a llm\n" + "\n" + "Usage:\n" + " llama-run [options] model [prompt]\n" + "\n" + "Options:\n" + " -c, --context-size \n" + " Context size (default: %d)\n" + " -n, --ngl \n" + " Number of GPU layers (default: %d)\n" + " -v, --verbose, --log-verbose\n" + " Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n" + " -h, --help\n" + " Show help message\n" + "\n" + "Commands:\n" + " model\n" + " Model is a string with an optional prefix of \n" + " huggingface:// (hf://), ollama://, https:// or file://.\n" + " If no protocol is specified and a file exists in the specified\n" + " path, file:// is assumed, otherwise if a file does not exist in\n" + " the specified path, ollama:// is assumed. Models that are being\n" + " pulled are downloaded with .partial extension while being\n" + " downloaded and then renamed as the file without the .partial\n" + " extension when complete.\n" + "\n" + "Examples:\n" + " llama-run llama3\n" + " llama-run ollama://granite-code\n" + " llama-run ollama://smollm:135m\n" + " llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n" + " llama-run " + "huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n" + " llama-run https://example.com/some-file1.gguf\n" + " llama-run some-file2.gguf\n" + " llama-run file://some-file3.gguf\n" + " llama-run --ngl 999 some-file4.gguf\n" + " llama-run --ngl 999 some-file5.gguf Hello World\n", + llama_context_default_params().n_batch, llama_model_default_params().n_gpu_layers); } }; -class LlamaData { - public: - llama_model_ptr model; - llama_sampler_ptr sampler; - llama_context_ptr context; - std::vector messages; +struct progress_data { + size_t file_size = 0; + std::chrono::steady_clock::time_point start_time = std::chrono::steady_clock::now(); + bool printed = false; +}; - int init(const Options & opt) { - model = initialize_model(opt.model_path, opt.ngl); +static int get_terminal_width() { +#if defined(_WIN32) + CONSOLE_SCREEN_BUFFER_INFO csbi; + GetConsoleScreenBufferInfo(GetStdHandle(STD_OUTPUT_HANDLE), &csbi); + return csbi.srWindow.Right - csbi.srWindow.Left + 1; +#else + struct winsize w; + ioctl(STDOUT_FILENO, TIOCGWINSZ, &w); + return w.ws_col; +#endif +} + +#ifdef LLAMA_USE_CURL +class File { + public: + FILE * file = nullptr; + + FILE * open(const std::string & filename, const char * mode) { + file = fopen(filename.c_str(), mode); + + return file; + } + + int lock() { + if (file) { +# ifdef _WIN32 + fd = _fileno(file); + hFile = (HANDLE) _get_osfhandle(fd); + if (hFile == INVALID_HANDLE_VALUE) { + fd = -1; + + return 1; + } + + OVERLAPPED overlapped = { 0 }; + if (!LockFileEx(hFile, LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, MAXDWORD, MAXDWORD, + &overlapped)) { + fd = -1; + + return 1; + } +# else + fd = fileno(file); + if (flock(fd, LOCK_EX | LOCK_NB) != 0) { + fd = -1; + + return 1; + } +# endif + } + + return 0; + } + + ~File() { + if (fd >= 0) { +# ifdef _WIN32 + if (hFile != INVALID_HANDLE_VALUE) { + OVERLAPPED overlapped = { 0 }; + UnlockFileEx(hFile, 0, MAXDWORD, MAXDWORD, &overlapped); + } +# else + flock(fd, LOCK_UN); +# endif + } + + if (file) { + fclose(file); + } + } + + private: + int fd = -1; +# ifdef _WIN32 + HANDLE hFile; +# endif +}; + +class HttpClient { + public: + int init(const std::string & url, const std::vector & headers, const std::string & output_file, + const bool progress, std::string * response_str = nullptr) { + std::string output_file_partial; + curl = curl_easy_init(); + if (!curl) { + return 1; + } + + progress_data data; + File out; + if (!output_file.empty()) { + output_file_partial = output_file + ".partial"; + if (!out.open(output_file_partial, "ab")) { + printe("Failed to open file\n"); + + return 1; + } + + if (out.lock()) { + printe("Failed to exclusively lock file\n"); + + return 1; + } + } + + set_write_options(response_str, out); + data.file_size = set_resume_point(output_file_partial); + set_progress_options(progress, data); + set_headers(headers); + perform(url); + if (!output_file.empty()) { + std::filesystem::rename(output_file_partial, output_file); + } + + return 0; + } + + ~HttpClient() { + if (chunk) { + curl_slist_free_all(chunk); + } + + if (curl) { + curl_easy_cleanup(curl); + } + } + + private: + CURL * curl = nullptr; + struct curl_slist * chunk = nullptr; + + void set_write_options(std::string * response_str, const File & out) { + if (response_str) { + curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, capture_data); + curl_easy_setopt(curl, CURLOPT_WRITEDATA, response_str); + } else { + curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, write_data); + curl_easy_setopt(curl, CURLOPT_WRITEDATA, out.file); + } + } + + size_t set_resume_point(const std::string & output_file) { + size_t file_size = 0; + if (std::filesystem::exists(output_file)) { + file_size = std::filesystem::file_size(output_file); + curl_easy_setopt(curl, CURLOPT_RESUME_FROM_LARGE, static_cast(file_size)); + } + + return file_size; + } + + void set_progress_options(bool progress, progress_data & data) { + if (progress) { + curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L); + curl_easy_setopt(curl, CURLOPT_XFERINFODATA, &data); + curl_easy_setopt(curl, CURLOPT_XFERINFOFUNCTION, update_progress); + } + } + + void set_headers(const std::vector & headers) { + if (!headers.empty()) { + if (chunk) { + curl_slist_free_all(chunk); + chunk = 0; + } + + for (const auto & header : headers) { + chunk = curl_slist_append(chunk, header.c_str()); + } + + curl_easy_setopt(curl, CURLOPT_HTTPHEADER, chunk); + } + } + + void perform(const std::string & url) { + CURLcode res; + curl_easy_setopt(curl, CURLOPT_URL, url.c_str()); + curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L); + curl_easy_setopt(curl, CURLOPT_DEFAULT_PROTOCOL, "https"); + curl_easy_setopt(curl, CURLOPT_FAILONERROR, 1L); + res = curl_easy_perform(curl); + if (res != CURLE_OK) { + printe("curl_easy_perform() failed: %s\n", curl_easy_strerror(res)); + } + } + + static std::string human_readable_time(double seconds) { + int hrs = static_cast(seconds) / 3600; + int mins = (static_cast(seconds) % 3600) / 60; + int secs = static_cast(seconds) % 60; + + if (hrs > 0) { + return fmt("%dh %02dm %02ds", hrs, mins, secs); + } else if (mins > 0) { + return fmt("%dm %02ds", mins, secs); + } else { + return fmt("%ds", secs); + } + } + + static std::string human_readable_size(curl_off_t size) { + static const char * suffix[] = { "B", "KB", "MB", "GB", "TB" }; + char length = sizeof(suffix) / sizeof(suffix[0]); + int i = 0; + double dbl_size = size; + if (size > 1024) { + for (i = 0; (size / 1024) > 0 && i < length - 1; i++, size /= 1024) { + dbl_size = size / 1024.0; + } + } + + return fmt("%.2f %s", dbl_size, suffix[i]); + } + + static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t, + curl_off_t) { + progress_data * data = static_cast(ptr); + if (total_to_download <= 0) { + return 0; + } + + total_to_download += data->file_size; + const curl_off_t now_downloaded_plus_file_size = now_downloaded + data->file_size; + const curl_off_t percentage = calculate_percentage(now_downloaded_plus_file_size, total_to_download); + std::string progress_prefix = generate_progress_prefix(percentage); + + const double speed = calculate_speed(now_downloaded, data->start_time); + const double tim = (total_to_download - now_downloaded) / speed; + std::string progress_suffix = + generate_progress_suffix(now_downloaded_plus_file_size, total_to_download, speed, tim); + + int progress_bar_width = calculate_progress_bar_width(progress_prefix, progress_suffix); + std::string progress_bar; + generate_progress_bar(progress_bar_width, percentage, progress_bar); + + print_progress(progress_prefix, progress_bar, progress_suffix); + data->printed = true; + + return 0; + } + + static curl_off_t calculate_percentage(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download) { + return (now_downloaded_plus_file_size * 100) / total_to_download; + } + + static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", percentage); } + + static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) { + const auto now = std::chrono::steady_clock::now(); + const std::chrono::duration elapsed_seconds = now - start_time; + return now_downloaded / elapsed_seconds.count(); + } + + static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download, + double speed, double estimated_time) { + const int width = 10; + return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width, + human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width, + human_readable_time(estimated_time).c_str()); + } + + static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) { + int progress_bar_width = get_terminal_width() - progress_prefix.size() - progress_suffix.size() - 3; + if (progress_bar_width < 1) { + progress_bar_width = 1; + } + + return progress_bar_width; + } + + static std::string generate_progress_bar(int progress_bar_width, curl_off_t percentage, + std::string & progress_bar) { + const curl_off_t pos = (percentage * progress_bar_width) / 100; + for (int i = 0; i < progress_bar_width; ++i) { + progress_bar.append((i < pos) ? "█" : " "); + } + + return progress_bar; + } + + static void print_progress(const std::string & progress_prefix, const std::string & progress_bar, + const std::string & progress_suffix) { + printe("\r%*s\r%s%s| %s", get_terminal_width(), " ", progress_prefix.c_str(), progress_bar.c_str(), + progress_suffix.c_str()); + } + // Function to write data to a file + static size_t write_data(void * ptr, size_t size, size_t nmemb, void * stream) { + FILE * out = static_cast(stream); + return fwrite(ptr, size, nmemb, out); + } + + // Function to capture data into a string + static size_t capture_data(void * ptr, size_t size, size_t nmemb, void * stream) { + std::string * str = static_cast(stream); + str->append(static_cast(ptr), size * nmemb); + return size * nmemb; + } +}; +#endif + +class LlamaData { + public: + llama_model_ptr model; + llama_sampler_ptr sampler; + llama_context_ptr context; + std::vector messages; + std::vector msg_strs; + std::vector fmtted; + + int init(Opt & opt) { + model = initialize_model(opt); if (!model) { return 1; } - context = initialize_context(model, opt.n_ctx); + context = initialize_context(model, opt.context_size_); if (!context) { return 1; } @@ -131,29 +504,144 @@ class LlamaData { return 0; } - private: - // Initializes the model and returns a unique pointer to it - llama_model_ptr initialize_model(const std::string & model_path, const int ngl) { - llama_model_params model_params = llama_model_default_params(); - model_params.n_gpu_layers = ngl; - - llama_model_ptr model(llama_load_model_from_file(model_path.c_str(), model_params)); - if (!model) { - fprintf(stderr, "%s: error: unable to load model\n", __func__); + private: +#ifdef LLAMA_USE_CURL + int download(const std::string & url, const std::vector & headers, const std::string & output_file, + const bool progress, std::string * response_str = nullptr) { + HttpClient http; + if (http.init(url, headers, output_file, progress, response_str)) { + return 1; } + return 0; + } +#else + int download(const std::string &, const std::vector &, const std::string &, const bool, + std::string * = nullptr) { + printe("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__); + return 1; + } +#endif + + int huggingface_dl(const std::string & model, const std::vector headers, const std::string & bn) { + // Find the second occurrence of '/' after protocol string + size_t pos = model.find('/'); + pos = model.find('/', pos + 1); + if (pos == std::string::npos) { + return 1; + } + + const std::string hfr = model.substr(0, pos); + const std::string hff = model.substr(pos + 1); + const std::string url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff; + return download(url, headers, bn, true); + } + + int ollama_dl(std::string & model, const std::vector headers, const std::string & bn) { + if (model.find('/') == std::string::npos) { + model = "library/" + model; + } + + std::string model_tag = "latest"; + size_t colon_pos = model.find(':'); + if (colon_pos != std::string::npos) { + model_tag = model.substr(colon_pos + 1); + model = model.substr(0, colon_pos); + } + + std::string manifest_url = "https://registry.ollama.ai/v2/" + model + "/manifests/" + model_tag; + std::string manifest_str; + const int ret = download(manifest_url, headers, "", false, &manifest_str); + if (ret) { + return ret; + } + + nlohmann::json manifest = nlohmann::json::parse(manifest_str); + std::string layer; + for (const auto & l : manifest["layers"]) { + if (l["mediaType"] == "application/vnd.ollama.image.model") { + layer = l["digest"]; + break; + } + } + + std::string blob_url = "https://registry.ollama.ai/v2/" + model + "/blobs/" + layer; + return download(blob_url, headers, bn, true); + } + + std::string basename(const std::string & path) { + const size_t pos = path.find_last_of("/\\"); + if (pos == std::string::npos) { + return path; + } + + return path.substr(pos + 1); + } + + int remove_proto(std::string & model_) { + const std::string::size_type pos = model_.find("://"); + if (pos == std::string::npos) { + return 1; + } + + model_ = model_.substr(pos + 3); // Skip past "://" + return 0; + } + + int resolve_model(std::string & model_) { + int ret = 0; + if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) { + remove_proto(model_); + + return ret; + } + + const std::string bn = basename(model_); + const std::vector headers = { "--header", + "Accept: application/vnd.docker.distribution.manifest.v2+json" }; + if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) { + remove_proto(model_); + ret = huggingface_dl(model_, headers, bn); + } else if (string_starts_with(model_, "ollama://")) { + remove_proto(model_); + ret = ollama_dl(model_, headers, bn); + } else if (string_starts_with(model_, "https://")) { + download(model_, headers, bn, true); + } else { + ret = ollama_dl(model_, headers, bn); + } + + model_ = bn; + + return ret; + } + + // Initializes the model and returns a unique pointer to it + llama_model_ptr initialize_model(Opt & opt) { + ggml_backend_load_all(); + llama_model_params model_params = llama_model_default_params(); + model_params.n_gpu_layers = opt.ngl_ >= 0 ? opt.ngl_ : model_params.n_gpu_layers; + resolve_model(opt.model_); + printe( + "\r%*s" + "\rLoading model", + get_terminal_width(), " "); + llama_model_ptr model(llama_load_model_from_file(opt.model_.c_str(), model_params)); + if (!model) { + printe("%s: error: unable to load model from file: %s\n", __func__, opt.model_.c_str()); + } + + printe("\r%*s\r", static_cast(sizeof("Loading model")), " "); return model; } // Initializes the context with the specified parameters llama_context_ptr initialize_context(const llama_model_ptr & model, const int n_ctx) { llama_context_params ctx_params = llama_context_default_params(); - ctx_params.n_ctx = n_ctx; - ctx_params.n_batch = n_ctx; - + ctx_params.n_ctx = ctx_params.n_batch = n_ctx >= 0 ? n_ctx : ctx_params.n_batch; llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params)); if (!context) { - fprintf(stderr, "%s: error: failed to create the llama_context\n", __func__); + printe("%s: error: failed to create the llama_context\n", __func__); } return context; @@ -170,23 +658,22 @@ class LlamaData { } }; -// Add a message to `messages` and store its content in `owned_content` -static void add_message(const char * role, const std::string & text, LlamaData & llama_data, - std::vector & owned_content) { - char_array_ptr content(new char[text.size() + 1]); - std::strcpy(content.get(), text.c_str()); - llama_data.messages.push_back({role, content.get()}); - owned_content.push_back(std::move(content)); +// Add a message to `messages` and store its content in `msg_strs` +static void add_message(const char * role, const std::string & text, LlamaData & llama_data) { + llama_data.msg_strs.push_back(std::move(text)); + llama_data.messages.push_back({ role, llama_data.msg_strs.back().c_str() }); } // Function to apply the chat template and resize `formatted` if needed -static int apply_chat_template(const LlamaData & llama_data, std::vector & formatted, const bool append) { - int result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(), - llama_data.messages.size(), append, formatted.data(), formatted.size()); - if (result > static_cast(formatted.size())) { - formatted.resize(result); +static int apply_chat_template(LlamaData & llama_data, const bool append) { + int result = llama_chat_apply_template( + llama_data.model.get(), nullptr, llama_data.messages.data(), llama_data.messages.size(), append, + append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0); + if (append && result > static_cast(llama_data.fmtted.size())) { + llama_data.fmtted.resize(result); result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(), - llama_data.messages.size(), append, formatted.data(), formatted.size()); + llama_data.messages.size(), append, llama_data.fmtted.data(), + llama_data.fmtted.size()); } return result; @@ -199,7 +686,8 @@ static int tokenize_prompt(const llama_model_ptr & model, const std::string & pr prompt_tokens.resize(n_prompt_tokens); if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) { - GGML_ABORT("failed to tokenize the prompt\n"); + printe("failed to tokenize the prompt\n"); + return -1; } return n_prompt_tokens; @@ -207,11 +695,11 @@ static int tokenize_prompt(const llama_model_ptr & model, const std::string & pr // Check if we have enough space in the context to evaluate this batch static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) { - const int n_ctx = llama_n_ctx(ctx.get()); + const int n_ctx = llama_n_ctx(ctx.get()); const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get()); if (n_ctx_used + batch.n_tokens > n_ctx) { printf("\033[0m\n"); - fprintf(stderr, "context size exceeded\n"); + printe("context size exceeded\n"); return 1; } @@ -221,9 +709,10 @@ static int check_context_size(const llama_context_ptr & ctx, const llama_batch & // convert the token to a string static int convert_token_to_string(const llama_model_ptr & model, const llama_token token_id, std::string & piece) { char buf[256]; - int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true); + int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true); if (n < 0) { - GGML_ABORT("failed to convert token to piece\n"); + printe("failed to convert token to piece\n"); + return 1; } piece = std::string(buf, n); @@ -238,19 +727,19 @@ static void print_word_and_concatenate_to_response(const std::string & piece, st // helper function to evaluate a prompt and generate a response static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) { - std::vector prompt_tokens; - const int n_prompt_tokens = tokenize_prompt(llama_data.model, prompt, prompt_tokens); - if (n_prompt_tokens < 0) { + std::vector tokens; + if (tokenize_prompt(llama_data.model, prompt, tokens) < 0) { return 1; } // prepare a batch for the prompt - llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size()); + llama_batch batch = llama_batch_get_one(tokens.data(), tokens.size()); llama_token new_token_id; while (true) { check_context_size(llama_data.context, batch); if (llama_decode(llama_data.context.get(), batch)) { - GGML_ABORT("failed to decode\n"); + printe("failed to decode\n"); + return 1; } // sample the next token, check is it an end of generation? @@ -273,44 +762,34 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str return 0; } -static int parse_arguments(const int argc, const char ** argv, Options & opt) { - ArgumentParser parser(argv[0]); - parser.add_argument("-m", opt.model_path, "model"); - parser.add_argument("-p", opt.prompt_non_interactive, "prompt"); - parser.add_argument("-c", opt.n_ctx, "context_size"); - parser.add_argument("-ngl", opt.ngl, "n_gpu_layers"); - if (parser.parse(argc, argv)) { - return 1; - } - - return 0; -} - static int read_user_input(std::string & user) { std::getline(std::cin, user); - return user.empty(); // Indicate an error or empty input + return user.empty(); // Should have data in happy path } // Function to generate a response based on the prompt -static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response) { +static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response, + const bool stdout_a_terminal) { // Set response color - printf("\033[33m"); + if (stdout_a_terminal) { + printf("\033[33m"); + } + if (generate(llama_data, prompt, response)) { - fprintf(stderr, "failed to generate response\n"); + printe("failed to generate response\n"); return 1; } // End response with color reset and newline - printf("\n\033[0m"); + printf("\n%s", stdout_a_terminal ? "\033[0m" : ""); return 0; } // Helper function to apply the chat template and handle errors -static int apply_chat_template_with_error_handling(const LlamaData & llama_data, std::vector & formatted, - const bool is_user_input, int & output_length) { - const int new_len = apply_chat_template(llama_data, formatted, is_user_input); +static int apply_chat_template_with_error_handling(LlamaData & llama_data, const bool append, int & output_length) { + const int new_len = apply_chat_template(llama_data, append); if (new_len < 0) { - fprintf(stderr, "failed to apply the chat template\n"); + printe("failed to apply the chat template\n"); return -1; } @@ -319,62 +798,82 @@ static int apply_chat_template_with_error_handling(const LlamaData & llama_data, } // Helper function to handle user input -static bool handle_user_input(std::string & user_input, const std::string & prompt_non_interactive) { - if (!prompt_non_interactive.empty()) { - user_input = prompt_non_interactive; - return true; // No need for interactive input +static int handle_user_input(std::string & user_input, const std::string & user_) { + if (!user_.empty()) { + user_input = user_; + return 0; // No need for interactive input } - printf("\033[32m> \033[0m"); - return !read_user_input(user_input); // Returns false if input ends the loop -} - -// Function to tokenize the prompt -static int chat_loop(LlamaData & llama_data, std::string & prompt_non_interactive) { - std::vector owned_content; - std::vector fmtted(llama_n_ctx(llama_data.context.get())); - int prev_len = 0; - - while (true) { - // Get user input - std::string user_input; - if (!handle_user_input(user_input, prompt_non_interactive)) { - break; - } - - add_message("user", prompt_non_interactive.empty() ? user_input : prompt_non_interactive, llama_data, - owned_content); - - int new_len; - if (apply_chat_template_with_error_handling(llama_data, fmtted, true, new_len) < 0) { - return 1; - } - - std::string prompt(fmtted.begin() + prev_len, fmtted.begin() + new_len); - std::string response; - if (generate_response(llama_data, prompt, response)) { - return 1; - } - } - return 0; -} - -static void log_callback(const enum ggml_log_level level, const char * text, void *) { - if (level == GGML_LOG_LEVEL_ERROR) { - fprintf(stderr, "%s", text); - } + printf( + "\r%*s" + "\r\033[32m> \033[0m", + get_terminal_width(), " "); + return read_user_input(user_input); // Returns true if input ends the loop } static bool is_stdin_a_terminal() { #if defined(_WIN32) HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE); - DWORD mode; + DWORD mode; return GetConsoleMode(hStdin, &mode); #else return isatty(STDIN_FILENO); #endif } +static bool is_stdout_a_terminal() { +#if defined(_WIN32) + HANDLE hStdout = GetStdHandle(STD_OUTPUT_HANDLE); + DWORD mode; + return GetConsoleMode(hStdout, &mode); +#else + return isatty(STDOUT_FILENO); +#endif +} + +// Function to tokenize the prompt +static int chat_loop(LlamaData & llama_data, const std::string & user_) { + int prev_len = 0; + llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get())); + static const bool stdout_a_terminal = is_stdout_a_terminal(); + while (true) { + // Get user input + std::string user_input; + while (handle_user_input(user_input, user_)) { + } + + add_message("user", user_.empty() ? user_input : user_, llama_data); + int new_len; + if (apply_chat_template_with_error_handling(llama_data, true, new_len) < 0) { + return 1; + } + + std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len); + std::string response; + if (generate_response(llama_data, prompt, response, stdout_a_terminal)) { + return 1; + } + + if (!user_.empty()) { + break; + } + + add_message("assistant", response, llama_data); + if (apply_chat_template_with_error_handling(llama_data, false, prev_len) < 0) { + return 1; + } + } + + return 0; +} + +static void log_callback(const enum ggml_log_level level, const char * text, void * p) { + const Opt * opt = static_cast(p); + if (opt->verbose_ || level == GGML_LOG_LEVEL_ERROR) { + printe("%s", text); + } +} + static std::string read_pipe_data() { std::ostringstream result; result << std::cin.rdbuf(); // Read all data from std::cin @@ -382,26 +881,29 @@ static std::string read_pipe_data() { } int main(int argc, const char ** argv) { - Options opt; - if (parse_arguments(argc, argv, opt)) { + Opt opt; + const int ret = opt.init(argc, argv); + if (ret == 2) { + return 0; + } else if (ret) { return 1; } if (!is_stdin_a_terminal()) { - if (!opt.prompt_non_interactive.empty()) { - opt.prompt_non_interactive += "\n\n"; + if (!opt.user_.empty()) { + opt.user_ += "\n\n"; } - opt.prompt_non_interactive += read_pipe_data(); + opt.user_ += read_pipe_data(); } - llama_log_set(log_callback, nullptr); + llama_log_set(log_callback, &opt); LlamaData llama_data; if (llama_data.init(opt)) { return 1; } - if (chat_loop(llama_data, opt.prompt_non_interactive)) { + if (chat_loop(llama_data, opt.user_)) { return 1; } diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 0035859a6..a27597cbc 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -15,7 +15,7 @@ set(TARGET_SRCS httplib.h ) set(PUBLIC_ASSETS - index.html + index.html.gz loading.html ) @@ -34,14 +34,6 @@ endforeach() add_executable(${TARGET} ${TARGET_SRCS}) install(TARGETS ${TARGET} RUNTIME) -# clean up generated files in pre-build step -foreach(asset ${PUBLIC_ASSETS}) - set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp") - add_custom_command(TARGET ${TARGET} PRE_BUILD - COMMAND "${CMAKE_COMMAND}" -E remove -f "${output}" - ) -endforeach() - target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) if (LLAMA_SERVER_SSL) diff --git a/examples/server/README.md b/examples/server/README.md index b2dd7b65a..d006a8d37 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -62,8 +62,8 @@ The project is under active development, and we are [looking for feedback and co | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)
(env: LLAMA_ARG_YARN_BETA_FAST) | | `-dkvc, --dump-kv-cache` | verbose print of the KV cache | | `-nkvo, --no-kv-offload` | disable KV offload
(env: LLAMA_ARG_NO_KV_OFFLOAD) | -| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)
(env: LLAMA_ARG_CACHE_TYPE_K) | -| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)
(env: LLAMA_ARG_CACHE_TYPE_V) | +| `-ctk, --cache-type-k TYPE` | KV cache data type for K
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_K) | +| `-ctv, --cache-type-v TYPE` | KV cache data type for V
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_V) | | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)
(env: LLAMA_ARG_DEFRAG_THOLD) | | `-np, --parallel N` | number of parallel sequences to decode (default: 1)
(env: LLAMA_ARG_N_PARALLEL) | | `--mlock` | force system to keep model in RAM rather than swapping or compressing
(env: LLAMA_ARG_MLOCK) | @@ -104,7 +104,6 @@ The project is under active development, and we are [looking for feedback and co | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | -| `--penalize-nl` | penalize newline tokens (default: false) | | `--temp N` | temperature (default: 0.8) | | `--top-k N` | top-k sampling (default: 40, 0 = disabled) | | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) | @@ -138,6 +137,7 @@ The project is under active development, and we are [looking for feedback and co | -------- | ----------- | | `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)
(env: LLAMA_ARG_NO_CONTEXT_SHIFT) | | `-sp, --special` | special tokens output enabled (default: false) | +| `--no-warmup` | skip warming up the model with an empty run | | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) | | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified
(env: LLAMA_ARG_POOLING) | | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)
(env: LLAMA_ARG_CONT_BATCHING) | @@ -146,6 +146,7 @@ The project is under active development, and we are [looking for feedback and co | `--host HOST` | ip address to listen (default: 127.0.0.1)
(env: LLAMA_ARG_HOST) | | `--port PORT` | port to listen (default: 8080)
(env: LLAMA_ARG_PORT) | | `--path PATH` | path to serve static files from (default: )
(env: LLAMA_ARG_STATIC_PATH) | +| `--no-webui` | Disable the Web UI (default: enabled)
(env: LLAMA_ARG_NO_WEBUI) | | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
(env: LLAMA_ARG_EMBEDDINGS) | | `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)
(env: LLAMA_ARG_RERANKING) | | `--api-key KEY` | API key to use for authentication (default: none)
(env: LLAMA_API_KEY) | @@ -163,13 +164,13 @@ The project is under active development, and we are [looking for feedback and co | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
list of built-in templates:
chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
| | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) | -| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16) | -| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5) | -| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9) | -| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model) | +| `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)
(env: LLAMA_ARG_DRAFT_MAX) | +| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5)
(env: LLAMA_ARG_DRAFT_MIN) | +| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9)
(env: LLAMA_ARG_DRAFT_P_MIN) | +| `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)
(env: LLAMA_ARG_CTX_SIZE_DRAFT) | | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | -| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model | -| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused) | +| `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | +| `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)
(env: LLAMA_ARG_MODEL_DRAFT) | Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var. @@ -302,23 +303,23 @@ mkdir llama-client cd llama-client ``` -Create a index.js file and put this inside: +Create an index.js file and put this inside: ```javascript -const prompt = `Building a website can be done in 10 simple steps:`; +const prompt = "Building a website can be done in 10 simple steps:" -async function Test() { +async function test() { let response = await fetch("http://127.0.0.1:8080/completion", { - method: 'POST', + method: "POST", body: JSON.stringify({ prompt, - n_predict: 512, + n_predict: 64, }) }) console.log((await response.json()).content) } -Test() +test() ``` And run it: @@ -380,7 +381,7 @@ Multiple prompts are also supported. In this case, the completion result will be `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token. By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt. -`stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`. +`stream`: Allows receiving each predicted token in real-time instead of waiting for the completion to finish (uses a different response format). To enable this, set to `true`. `stop`: Specify a JSON array of stopping strings. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]` @@ -391,8 +392,6 @@ These words will not be included in the completion, so make sure to add them to `repeat_last_n`: Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size. -`penalize_nl`: Penalize newline tokens when applying the repeat penalty. Default: `true` - `presence_penalty`: Repeat alpha presence penalty. Default: `0.0`, which is disabled. `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled. @@ -439,19 +438,22 @@ These words will not be included in the completion, so make sure to add them to `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true` +`return_tokens`: Return the raw generated token ids in the `tokens` field. Otherwise `tokens` remains empty. Default: `false` + `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values. - `timings_per_token`: Include prompt processing and text generation speed information in each response. Default: `false` +`timings_per_token`: Include prompt processing and text generation speed information in each response. Default: `false` **Response format** -- Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion. +- Note: In streaming mode (`stream`), only `content`, `tokens` and `stop` will be returned until end of completion. Responses are sent using the [Server-sent events](https://html.spec.whatwg.org/multipage/server-sent-events.html) standard. Note: the browser's `EventSource` interface cannot be used due to its lack of `POST` request support. - `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure: ```json { - "content": "", + "content": "", + "tokens": [ generated token ids if requested ], "probs": [ { "prob": float, @@ -469,13 +471,16 @@ These words will not be included in the completion, so make sure to add them to Notice that each `probs` is an array of length `n_probs`. - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string. +- `tokens`: Same as `content` but represented as raw token ids. Only populated if `"return_tokens": true` or `"stream": true` in the request. - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options) - `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.). - `model`: The path to the model loaded with `-m` - `prompt`: The provided `prompt` -- `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token -- `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered -- `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided +- `stop_type`: Indicating whether the completion has stopped. Possible values are: + - `none`: Generating (not stopped) + - `eos`: Stopped because it encountered the EOS token + - `limit`: Stopped because `n_predict` tokens were generated before stop words or EOS was encountered + - `word`: Stopped due to encountering a stopping word from `stop` JSON array provided - `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word) - `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second` - `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`) @@ -616,14 +621,82 @@ This endpoint is public (no API key check). By default, it is read-only. To make ```json { - "default_generation_settings": { ... }, + "default_generation_settings": { + "id": 0, + "id_task": -1, + "n_ctx": 1024, + "speculative": false, + "is_processing": false, + "params": { + "n_predict": -1, + "seed": 4294967295, + "temperature": 0.800000011920929, + "dynatemp_range": 0.0, + "dynatemp_exponent": 1.0, + "top_k": 40, + "top_p": 0.949999988079071, + "min_p": 0.05000000074505806, + "xtc_probability": 0.0, + "xtc_threshold": 0.10000000149011612, + "typical_p": 1.0, + "repeat_last_n": 64, + "repeat_penalty": 1.0, + "presence_penalty": 0.0, + "frequency_penalty": 0.0, + "dry_multiplier": 0.0, + "dry_base": 1.75, + "dry_allowed_length": 2, + "dry_penalty_last_n": -1, + "dry_sequence_breakers": [ + "\n", + ":", + "\"", + "*" + ], + "mirostat": 0, + "mirostat_tau": 5.0, + "mirostat_eta": 0.10000000149011612, + "stop": [], + "max_tokens": -1, + "n_keep": 0, + "n_discard": 0, + "ignore_eos": false, + "stream": true, + "n_probs": 0, + "min_keep": 0, + "grammar": "", + "samplers": [ + "dry", + "top_k", + "typ_p", + "top_p", + "min_p", + "xtc", + "temperature" + ], + "speculative.n_max": 16, + "speculative.n_min": 5, + "speculative.p_min": 0.8999999761581421, + "timings_per_token": false + }, + "prompt": "", + "next_token": { + "has_next_token": true, + "has_new_line": false, + "n_remain": -1, + "n_decoded": 0, + "stopping_word": "" + } + }, "total_slots": 1, - "chat_template": "" + "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf", + "chat_template": "..." } ``` - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint. - `total_slots` - the total number of slots for process requests (defined by `--parallel` option) +- `model_path` - the path to model file (same with `-m` argument) - `chat_template` - the model's original Jinja2 prompt template ### POST `/props`: Change server global properties. @@ -690,6 +763,8 @@ curl http://localhost:8080/v1/chat/completions \ ### POST `/v1/embeddings`: OpenAI-compatible embeddings API +This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm. + *Options:* See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings). @@ -722,6 +797,46 @@ See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-r }' ``` +### POST `/embeddings`: non-OpenAI-compatible embeddings API + +This endpoint supports all poolings, including `--pooling none`. When the pooling is `none`, the responses will contain the *unnormalized* embeddings for *all* input tokens. For all other pooling types, only the pooled embeddings are returned, normalized using Euclidian norm. + +Note that the response format of this endpoint is different from `/v1/embeddings`. + +*Options:* + +Same as the `/v1/embeddings` endpoint. + +*Examples:* + +Same as the `/v1/embeddings` endpoint. + +**Response format** + +```json +[ + { + "index": 0, + "embedding": [ + [ ... embeddings for token 0 ... ], + [ ... embeddings for token 1 ... ], + [ ... ] + [ ... embeddings for token N-1 ... ], + ] + }, + ... + { + "index": P, + "embedding": [ + [ ... embeddings for token 0 ... ], + [ ... embeddings for token 1 ... ], + [ ... ] + [ ... embeddings for token N-1 ... ], + ] + } +] +``` + ### GET `/slots`: Returns the current slots processing state > [!WARNING] @@ -737,56 +852,73 @@ Example: ```json [ - { - "dynatemp_exponent": 1.0, - "dynatemp_range": 0.0, - "frequency_penalty": 0.0, - "grammar": "", - "id": 0, - "ignore_eos": false, - "is_processing": false, - "logit_bias": [], - "min_p": 0.05000000074505806, - "mirostat": 0, - "mirostat_eta": 0.10000000149011612, - "mirostat_tau": 5.0, - "model": "llama-2-7b-32k-instruct.Q2_K.gguf", - "n_ctx": 2048, - "n_keep": 0, - "n_predict": 100000, - "n_probs": 0, - "next_token": { - "has_next_token": true, - "n_remain": -1, - "n_decoded": 0, - "stopped_eos": false, - "stopped_limit": false, - "stopped_word": false, - "stopping_word": "" - }, - "penalize_nl": true, - "presence_penalty": 0.0, - "prompt": "Say hello to llama.cpp", - "repeat_last_n": 64, - "repeat_penalty": 1.100000023841858, - "samplers": [ - "top_k", - "typical_p", - "top_p", - "min_p", - "temperature" - ], - "seed": 42, - "stop": [ - "\n" - ], - "stream": false, - "task_id": 0, - "temperature": 0.0, - "top_k": 40, - "top_p": 0.949999988079071, - "typical_p": 1.0 + { + "id": 0, + "id_task": -1, + "n_ctx": 1024, + "speculative": false, + "is_processing": false, + "params": { + "n_predict": -1, + "seed": 4294967295, + "temperature": 0.800000011920929, + "dynatemp_range": 0.0, + "dynatemp_exponent": 1.0, + "top_k": 40, + "top_p": 0.949999988079071, + "min_p": 0.05000000074505806, + "xtc_probability": 0.0, + "xtc_threshold": 0.10000000149011612, + "typical_p": 1.0, + "repeat_last_n": 64, + "repeat_penalty": 1.0, + "presence_penalty": 0.0, + "frequency_penalty": 0.0, + "dry_multiplier": 0.0, + "dry_base": 1.75, + "dry_allowed_length": 2, + "dry_penalty_last_n": -1, + "dry_sequence_breakers": [ + "\n", + ":", + "\"", + "*" + ], + "mirostat": 0, + "mirostat_tau": 5.0, + "mirostat_eta": 0.10000000149011612, + "stop": [], + "max_tokens": -1, + "n_keep": 0, + "n_discard": 0, + "ignore_eos": false, + "stream": true, + "n_probs": 0, + "min_keep": 0, + "grammar": "", + "samplers": [ + "dry", + "top_k", + "typ_p", + "top_p", + "min_p", + "xtc", + "temperature" + ], + "speculative.n_max": 16, + "speculative.n_min": 5, + "speculative.p_min": 0.8999999761581421, + "timings_per_token": false + }, + "prompt": "", + "next_token": { + "has_next_token": true, + "has_new_line": false, + "n_remain": -1, + "n_decoded": 0, + "stopping_word": "" } + } ] ``` diff --git a/examples/server/public/index.html b/examples/server/public/index.html deleted file mode 100644 index 250729a44..000000000 --- a/examples/server/public/index.html +++ /dev/null @@ -1,351 +0,0 @@ - - - - - - - - 🦙 llama.cpp - chat - - - - - -
-
- - - -
- -
-
-

Conversations

- - - -
- - -
- + New conversation -
-
- {{ conv.messages[0].content }} -
-
- Conversations are saved to browser's localStorage -
-
-
- - -
- -
- - - -
llama.cpp
- - -
-
- - - - -
- - - -
-
- - - -
-
    -
  • - -
    • -
  • - -
    • -
-
-
-
- - -
-
- - {{ messages.length === 0 ? 'Send a message to start' : '' }} -
-
-
-
- - - - -
-
- - -
- - - - - -
-
- - -
-
- - -
-
-
- - -
- - - -
-
- -
- - - - -
-

Settings

-
-

Settings below are saved in browser's localStorage

- - - - - -
- Other sampler settings -
- - - - -
-
- -
- Penalties settings -
- -
-
- -
- Advanced config -
- -
-
-
- - -
- - - -
-
- - -
- - - - - - - diff --git a/examples/server/public/index.html.gz b/examples/server/public/index.html.gz new file mode 100644 index 000000000..6a6491557 Binary files /dev/null and b/examples/server/public/index.html.gz differ diff --git a/examples/server/public_legacy/index-new.html b/examples/server/public_legacy/index-new.html index 8bfa380e5..cbfbbdf28 100644 --- a/examples/server/public_legacy/index-new.html +++ b/examples/server/public_legacy/index-new.html @@ -39,7 +39,6 @@ temperature: 0.8, // adapt all following parameters to optimized min-p requierements. If for non-english, set to 0.6 or lower repeat_last_n: 0, // 0 = disable penalty, -1 = context size repeat_penalty: 1.0, // 1.0 = disabled - penalize_nl: false, // true only useful for infinite completion dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well dry_base: 1.75, // 0.0 = disabled dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well diff --git a/examples/server/public_legacy/index.html b/examples/server/public_legacy/index.html index a95f5c6df..75f39330a 100644 --- a/examples/server/public_legacy/index.html +++ b/examples/server/public_legacy/index.html @@ -303,7 +303,6 @@ temperature: 0.7, repeat_last_n: 256, // 0 = disable penalty, -1 = context size repeat_penalty: 1.18, // 1.0 = disabled - penalize_nl: false, dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well dry_base: 1.75, // 0.0 = disabled dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well @@ -1006,7 +1005,6 @@ ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })} ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })} ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })} - ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })} ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })} ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })} ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })} diff --git a/examples/server/public_simplechat/simplechat.js b/examples/server/public_simplechat/simplechat.js index 8e0df3b61..2fcd24a86 100644 --- a/examples/server/public_simplechat/simplechat.js +++ b/examples/server/public_simplechat/simplechat.js @@ -407,6 +407,9 @@ class SimpleChat { if (curLine.startsWith("data:")) { curLine = curLine.substring(5); } + if (curLine.trim() === "[DONE]") { + break; + } let curJson = JSON.parse(curLine); console.debug("DBUG:SC:PART:Json:", curJson); this.append_response(this.response_extract_stream(curJson, apiEP)); diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 9bca3f30e..5ed4e8d27 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -15,7 +15,7 @@ #define MIMETYPE_JSON "application/json; charset=utf-8" // auto generated files (update with ./deps.sh) -#include "index.html.hpp" +#include "index.html.gz.hpp" #include "loading.html.hpp" #include @@ -33,8 +33,10 @@ using json = nlohmann::ordered_json; enum stop_type { - STOP_TYPE_FULL, - STOP_TYPE_PARTIAL, + STOP_TYPE_NONE, + STOP_TYPE_EOS, + STOP_TYPE_WORD, + STOP_TYPE_LIMIT, }; // state diagram: https://github.com/ggerganov/llama.cpp/pull/9283 @@ -52,7 +54,10 @@ enum server_state { }; enum server_task_type { - SERVER_TASK_TYPE_INFERENCE, + SERVER_TASK_TYPE_COMPLETION, + SERVER_TASK_TYPE_EMBEDDING, + SERVER_TASK_TYPE_RERANK, + SERVER_TASK_TYPE_INFILL, SERVER_TASK_TYPE_CANCEL, SERVER_TASK_TYPE_NEXT_RESPONSE, SERVER_TASK_TYPE_METRICS, @@ -62,22 +67,306 @@ enum server_task_type { SERVER_TASK_TYPE_SET_LORA, }; -enum server_task_inf_type { - SERVER_TASK_INF_TYPE_COMPLETION, - SERVER_TASK_INF_TYPE_EMBEDDING, - SERVER_TASK_INF_TYPE_RERANK, - SERVER_TASK_INF_TYPE_INFILL, +// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11 +enum error_type { + ERROR_TYPE_INVALID_REQUEST, + ERROR_TYPE_AUTHENTICATION, + ERROR_TYPE_SERVER, + ERROR_TYPE_NOT_FOUND, + ERROR_TYPE_PERMISSION, + ERROR_TYPE_UNAVAILABLE, // custom error + ERROR_TYPE_NOT_SUPPORTED, // custom error +}; + +struct slot_params { + bool stream = true; + bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt + bool return_tokens = false; + + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_discard = 0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half + int32_t n_predict = -1; // new tokens to predict + int32_t n_indent = 0; // mininum line indentation for the generated text in number of whitespace characters + + int64_t t_max_prompt_ms = -1; // TODO: implement + int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit + + std::vector antiprompt; + bool timings_per_token = false; + bool ignore_eos = false; + + struct common_params_sampling sampling; + struct common_params_speculative speculative; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + + json to_json() const { + std::vector samplers; + samplers.reserve(sampling.samplers.size()); + for (const auto & sampler : sampling.samplers) { + samplers.emplace_back(common_sampler_type_to_str(sampler)); + } + + return json { + {"n_predict", n_predict}, // Server configured n_predict + {"seed", sampling.seed}, + {"temperature", sampling.temp}, + {"dynatemp_range", sampling.dynatemp_range}, + {"dynatemp_exponent", sampling.dynatemp_exponent}, + {"top_k", sampling.top_k}, + {"top_p", sampling.top_p}, + {"min_p", sampling.min_p}, + {"xtc_probability", sampling.xtc_probability}, + {"xtc_threshold", sampling.xtc_threshold}, + {"typical_p", sampling.typ_p}, + {"repeat_last_n", sampling.penalty_last_n}, + {"repeat_penalty", sampling.penalty_repeat}, + {"presence_penalty", sampling.penalty_present}, + {"frequency_penalty", sampling.penalty_freq}, + {"dry_multiplier", sampling.dry_multiplier}, + {"dry_base", sampling.dry_base}, + {"dry_allowed_length", sampling.dry_allowed_length}, + {"dry_penalty_last_n", sampling.dry_penalty_last_n}, + {"dry_sequence_breakers", sampling.dry_sequence_breakers}, + {"mirostat", sampling.mirostat}, + {"mirostat_tau", sampling.mirostat_tau}, + {"mirostat_eta", sampling.mirostat_eta}, + {"stop", antiprompt}, + {"max_tokens", n_predict}, // User configured n_predict + {"n_keep", n_keep}, + {"n_discard", n_discard}, + {"ignore_eos", sampling.ignore_eos}, + {"stream", stream}, + {"logit_bias", format_logit_bias(sampling.logit_bias)}, + {"n_probs", sampling.n_probs}, + {"min_keep", sampling.min_keep}, + {"grammar", sampling.grammar}, + {"samplers", samplers}, + {"speculative.n_max", speculative.n_max}, + {"speculative.n_min", speculative.n_min}, + {"speculative.p_min", speculative.p_min}, + {"timings_per_token", timings_per_token}, + }; + } }; struct server_task { - int id = -1; // to be filled by server_queue - int id_target = -1; // used by SERVER_TASK_TYPE_CANCEL + int id = -1; // to be filled by server_queue + int index = -1; // used when there are multiple prompts (batch request) - llama_tokens prompt_tokens; server_task_type type; - json data; - server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION; + // used by SERVER_TASK_TYPE_CANCEL + int id_target = -1; + + // used by SERVER_TASK_TYPE_INFERENCE + slot_params params; + llama_tokens prompt_tokens; + int id_selected_slot = -1; + + // used by SERVER_TASK_TYPE_SLOT_SAVE, SERVER_TASK_TYPE_SLOT_RESTORE, SERVER_TASK_TYPE_SLOT_ERASE + struct slot_action { + int slot_id; + std::string filename; + std::string filepath; + }; + slot_action slot_action; + + // used by SERVER_TASK_TYPE_METRICS + bool metrics_reset_bucket = false; + + server_task(server_task_type type) : type(type) {} + + static slot_params params_from_json_cmpl( + const llama_model * model, + const llama_context * ctx, + const common_params & params_base, + const json & data) { + slot_params params; + + // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them) + slot_params defaults; + defaults.sampling = params_base.sampling; + defaults.speculative = params_base.speculative; + + // enabling this will output extra debug information in the HTTP responses from the server + params.verbose = params_base.verbosity > 9; + params.timings_per_token = json_value(data, "timings_per_token", false); + + params.stream = json_value(data, "stream", false); + params.cache_prompt = json_value(data, "cache_prompt", true); + params.return_tokens = json_value(data, "return_tokens", false); + params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", defaults.n_predict)); + params.n_indent = json_value(data, "n_indent", defaults.n_indent); + params.n_keep = json_value(data, "n_keep", defaults.n_keep); + params.n_discard = json_value(data, "n_discard", defaults.n_discard); + //params.t_max_prompt_ms = json_value(data, "t_max_prompt_ms", defaults.t_max_prompt_ms); // TODO: implement + params.t_max_predict_ms = json_value(data, "t_max_predict_ms", defaults.t_max_predict_ms); + + params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k); + params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p); + params.sampling.min_p = json_value(data, "min_p", defaults.sampling.min_p); + params.sampling.xtc_probability = json_value(data, "xtc_probability", defaults.sampling.xtc_probability); + params.sampling.xtc_threshold = json_value(data, "xtc_threshold", defaults.sampling.xtc_threshold); + params.sampling.typ_p = json_value(data, "typical_p", defaults.sampling.typ_p); + params.sampling.temp = json_value(data, "temperature", defaults.sampling.temp); + params.sampling.dynatemp_range = json_value(data, "dynatemp_range", defaults.sampling.dynatemp_range); + params.sampling.dynatemp_exponent = json_value(data, "dynatemp_exponent", defaults.sampling.dynatemp_exponent); + params.sampling.penalty_last_n = json_value(data, "repeat_last_n", defaults.sampling.penalty_last_n); + params.sampling.penalty_repeat = json_value(data, "repeat_penalty", defaults.sampling.penalty_repeat); + params.sampling.penalty_freq = json_value(data, "frequency_penalty", defaults.sampling.penalty_freq); + params.sampling.penalty_present = json_value(data, "presence_penalty", defaults.sampling.penalty_present); + params.sampling.dry_multiplier = json_value(data, "dry_multiplier", defaults.sampling.dry_multiplier); + params.sampling.dry_base = json_value(data, "dry_base", defaults.sampling.dry_base); + params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length); + params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n); + params.sampling.mirostat = json_value(data, "mirostat", defaults.sampling.mirostat); + params.sampling.mirostat_tau = json_value(data, "mirostat_tau", defaults.sampling.mirostat_tau); + params.sampling.mirostat_eta = json_value(data, "mirostat_eta", defaults.sampling.mirostat_eta); + params.sampling.seed = json_value(data, "seed", defaults.sampling.seed); + params.sampling.n_probs = json_value(data, "n_probs", defaults.sampling.n_probs); + params.sampling.min_keep = json_value(data, "min_keep", defaults.sampling.min_keep); + + params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min); + params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max); + params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min); + + params.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min); + params.speculative.n_min = std::max(params.speculative.n_min, 2); + params.speculative.n_max = std::max(params.speculative.n_max, 0); + + // TODO: add more sanity checks for the input parameters + + if (params.sampling.penalty_last_n < -1) { + throw std::runtime_error("Error: repeat_last_n must be >= -1"); + } + + if (params.sampling.dry_penalty_last_n < -1) { + throw std::runtime_error("Error: dry_penalty_last_n must be >= -1"); + } + + if (params.sampling.penalty_last_n == -1) { + // note: should be the slot's context and not the full context, but it's ok + params.sampling.penalty_last_n = llama_n_ctx(ctx); + } + + if (params.sampling.dry_penalty_last_n == -1) { + params.sampling.dry_penalty_last_n = llama_n_ctx(ctx); + } + + if (params.sampling.dry_base < 1.0f) { + params.sampling.dry_base = defaults.sampling.dry_base; + } + + // sequence breakers for DRY + { + // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format + // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39 + + if (data.contains("dry_sequence_breakers")) { + params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector()); + if (params.sampling.dry_sequence_breakers.empty()) { + throw std::runtime_error("Error: dry_sequence_breakers must be a non-empty array of strings"); + } + } + } + + // process "json_schema" and "grammar" + if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) { + throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both"); + } + if (data.contains("json_schema") && !data.contains("grammar")) { + try { + auto schema = json_value(data, "json_schema", json::object()); + params.sampling.grammar = json_schema_to_grammar(schema); + } catch (const std::exception & e) { + throw std::runtime_error(std::string("\"json_schema\": ") + e.what()); + } + } else { + params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar); + } + + { + params.sampling.logit_bias.clear(); + params.ignore_eos = json_value(data, "ignore_eos", false); + + const auto & logit_bias = data.find("logit_bias"); + if (logit_bias != data.end() && logit_bias->is_array()) { + const int n_vocab = llama_n_vocab(model); + for (const auto & el : *logit_bias) { + // TODO: we may want to throw errors here, in case "el" is incorrect + if (el.is_array() && el.size() == 2) { + float bias; + if (el[1].is_number()) { + bias = el[1].get(); + } else if (el[1].is_boolean() && !el[1].get()) { + bias = -INFINITY; + } else { + continue; + } + + if (el[0].is_number_integer()) { + llama_token tok = el[0].get(); + if (tok >= 0 && tok < n_vocab) { + params.sampling.logit_bias.push_back({tok, bias}); + } + } else if (el[0].is_string()) { + auto toks = common_tokenize(model, el[0].get(), false); + for (auto tok : toks) { + params.sampling.logit_bias.push_back({tok, bias}); + } + } + } + } + } + } + + { + params.antiprompt.clear(); + + const auto & stop = data.find("stop"); + if (stop != data.end() && stop->is_array()) { + for (const auto & word : *stop) { + if (!word.empty()) { + params.antiprompt.push_back(word); + } + } + } + } + + { + const auto & samplers = data.find("samplers"); + if (samplers != data.end()) { + if (samplers->is_array()) { + std::vector sampler_names; + for (const auto & name : *samplers) { + if (name.is_string()) { + sampler_names.emplace_back(name); + } + } + params.sampling.samplers = common_sampler_types_from_names(sampler_names, false); + } else if (samplers->is_string()){ + std::string sampler_string; + for (const auto & name : *samplers) { + sampler_string += name; + } + params.sampling.samplers = common_sampler_types_from_chars(sampler_string); + } + } else { + params.sampling.samplers = defaults.sampling.samplers; + } + } + + std::string model_name = params_base.model_alias.empty() ? DEFAULT_OAICOMPAT_MODEL : params_base.model_alias; + params.oaicompat_model = json_value(data, "model", model_name); + + return params; + } // utility function static std::unordered_set get_list_id(const std::vector & tasks) { @@ -89,39 +378,575 @@ struct server_task { } }; -struct server_task_result { - int id = -1; +struct result_timings { + int32_t prompt_n = -1; + double prompt_ms; + double prompt_per_token_ms; + double prompt_per_second; - json data; + int32_t predicted_n = -1; + double predicted_ms; + double predicted_per_token_ms; + double predicted_per_second; - bool stop; - bool error; + json to_json() const { + return { + {"prompt_n", prompt_n}, + {"prompt_ms", prompt_ms}, + {"prompt_per_token_ms", prompt_per_token_ms}, + {"prompt_per_second", prompt_per_second}, + + {"predicted_n", predicted_n}, + {"predicted_ms", predicted_ms}, + {"predicted_per_token_ms", predicted_per_token_ms}, + {"predicted_per_second", predicted_per_second}, + }; + } }; -struct slot_params { - bool stream = true; - bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt +struct server_task_result { + int id = -1; + int id_slot = -1; + virtual bool is_error() { + // only used by server_task_result_error + return false; + } + virtual bool is_stop() { + // only used by server_task_result_cmpl_* + return false; + } + virtual int get_index() { + return -1; + } + virtual json to_json() = 0; + virtual ~server_task_result() = default; +}; - int32_t n_keep = 0; // number of tokens to keep from initial prompt - int32_t n_discard = 0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half - int32_t n_predict = -1; // new tokens to predict - int32_t n_indent = 0; // mininum line indentation for the generated text in number of whitespace characters +// using shared_ptr for polymorphism of server_task_result +using server_task_result_ptr = std::unique_ptr; - int64_t t_max_prompt_ms = -1; // TODO: implement - int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit +inline std::string stop_type_to_str(stop_type type) { + switch (type) { + case STOP_TYPE_EOS: return "eos"; + case STOP_TYPE_WORD: return "word"; + case STOP_TYPE_LIMIT: return "limit"; + default: return "none"; + } +} - std::vector antiprompt; +struct completion_token_output { + llama_token tok; + std::string text_to_send; + struct token_prob { + llama_token tok; + std::string tok_str; + float prob; + }; + std::vector probs; - struct common_params_sampling sampling; - struct common_params_speculative speculative; + json to_json() const { + json probs_for_token = json::array(); + for (const auto & p : probs) { + probs_for_token.push_back(json { + {"tok_str", p.tok_str}, + {"prob", p.prob}, + }); + } + return probs_for_token; + } + + static json probs_vector_to_json(const std::vector & probs) { + json out = json::array(); + for (const auto & prob : probs) { + const std::string tok_str = prob.text_to_send; + out.push_back(json { + {"content", tok_str}, + {"probs", prob.to_json()}, + }); + } + return out; + } +}; + +struct server_task_result_cmpl_final : server_task_result { + int index = 0; + + std::string content; + llama_tokens tokens; + + bool stream; + result_timings timings; + std::string prompt; + + bool truncated; + int32_t n_decoded; + int32_t n_prompt_tokens; + int32_t n_tokens_cached; + bool has_new_line; + std::string stopping_word; + stop_type stop = STOP_TYPE_NONE; + + std::vector probs_output; + + slot_params generation_params; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; // TODO: support oaicompat for non-chat + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + + virtual int get_index() override { + return index; + } + + virtual bool is_stop() override { + return true; // in stream mode, final responses are considered stop + } + + virtual json to_json() override { + return oaicompat + ? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat()) + : to_json_non_oaicompat(); + } + + json to_json_non_oaicompat() { + json res = json { + {"index", index}, + {"content", stream ? "" : content}, // in stream mode, content is already in last partial chunk + {"tokens", stream ? llama_tokens {} : tokens}, + {"id_slot", id_slot}, + {"stop", true}, + {"model", oaicompat_model}, + {"tokens_predicted", n_decoded}, + {"tokens_evaluated", n_prompt_tokens}, + {"generation_settings", generation_params.to_json()}, + {"prompt", prompt}, + {"has_new_line", has_new_line}, + {"truncated", truncated}, + {"stop_type", stop_type_to_str(stop)}, + {"stopping_word", stopping_word}, + {"tokens_cached", n_tokens_cached}, + {"timings", timings.to_json()}, + }; + if (!probs_output.empty()) { + res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output); + } + return res; + } + + json to_json_oaicompat_chat() { + std::string finish_reason = "length"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + finish_reason = "stop"; + } + + json choices = json::array({json{ + {"finish_reason", finish_reason}, + {"index", 0}, + {"message", json { + {"content", content}, + {"role", "assistant"} + } + }}}); + + std::time_t t = std::time(0); + + json res = json { + {"choices", choices}, + {"created", t}, + {"model", oaicompat_model}, + {"object", "chat.completion"}, + {"usage", json { + {"completion_tokens", n_decoded}, + {"prompt_tokens", n_prompt_tokens}, + {"total_tokens", n_decoded + n_prompt_tokens} + }}, + {"id", oaicompat_cmpl_id} + }; + + // extra fields for debugging purposes + if (verbose) { + res["__verbose"] = to_json_non_oaicompat(); + } + if (timings.prompt_n >= 0) { + res.push_back({"timings", timings.to_json()}); + } + + return res; + } + + json to_json_oaicompat_chat_stream() { + std::time_t t = std::time(0); + std::string finish_reason = "length"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + finish_reason = "stop"; + } + + json choices = json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}); + + json ret = json { + {"choices", choices}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}, + {"usage", json { + {"completion_tokens", n_decoded}, + {"prompt_tokens", n_prompt_tokens}, + {"total_tokens", n_decoded + n_prompt_tokens}, + }}, + }; + + if (timings.prompt_n >= 0) { + ret.push_back({"timings", timings.to_json()}); + } + + return ret; + } +}; + +struct server_task_result_cmpl_partial : server_task_result { + int index = 0; + + std::string content; + llama_tokens tokens; + + int32_t n_decoded; + int32_t n_prompt_tokens; + + std::vector probs_output; + result_timings timings; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; // TODO: support oaicompat for non-chat + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + + virtual int get_index() override { + return index; + } + + virtual bool is_stop() override { + return false; // in stream mode, partial responses are not considered stop + } + + virtual json to_json() override { + return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat(); + } + + json to_json_non_oaicompat() { + // non-OAI-compat JSON + json res = json { + {"index", index}, + {"content", content}, + {"tokens", tokens}, + {"stop", false}, + {"id_slot", id_slot}, + {"tokens_predicted", n_decoded}, + {"tokens_evaluated", n_prompt_tokens}, + }; + // populate the timings object when needed (usually for the last response or with timings_per_token enabled) + if (timings.prompt_n > 0) { + res.push_back({"timings", timings.to_json()}); + } + if (!probs_output.empty()) { + res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output); + } + return res; + } + + json to_json_oaicompat() { + bool first = n_decoded == 0; + std::time_t t = std::time(0); + json choices; + + if (first) { + if (content.empty()) { + choices = json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{{"role", "assistant"}}}}}); + } else { + // We have to send this as two updates to conform to openai behavior + json initial_ret = json{{"choices", json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"role", "assistant"} + }}}})}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}}; + + json second_ret = json{ + {"choices", json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json { + {"content", content}}} + }})}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}}; + + return std::vector({initial_ret, second_ret}); + } + } else { + choices = json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", + json { + {"content", content}, + }}, + }}); + } + + json ret = json { + {"choices", choices}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"} + }; + + if (timings.prompt_n >= 0) { + ret.push_back({"timings", timings.to_json()}); + } + + return std::vector({ret}); + } +}; + +struct server_task_result_embd : server_task_result { + int index = 0; + std::vector> embedding; + + int32_t n_tokens; + + // OAI-compat fields + bool oaicompat = false; + + virtual int get_index() override { + return index; + } + + virtual json to_json() override { + return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat(); + } + + json to_json_non_oaicompat() { + return json { + {"index", index}, + {"embedding", embedding}, + }; + } + + json to_json_oaicompat() { + return json { + {"index", index}, + {"embedding", embedding[0]}, + {"tokens_evaluated", n_tokens}, + }; + } +}; + +struct server_task_result_rerank : server_task_result { + int index = 0; + float score = -1e6; + + int32_t n_tokens; + + virtual int get_index() override { + return index; + } + + virtual json to_json() override { + return json { + {"index", index}, + {"score", score}, + {"tokens_evaluated", n_tokens}, + }; + } +}; + +// this function maybe used outside of server_task_result_error +static json format_error_response(const std::string & message, const enum error_type type) { + std::string type_str; + int code = 500; + switch (type) { + case ERROR_TYPE_INVALID_REQUEST: + type_str = "invalid_request_error"; + code = 400; + break; + case ERROR_TYPE_AUTHENTICATION: + type_str = "authentication_error"; + code = 401; + break; + case ERROR_TYPE_NOT_FOUND: + type_str = "not_found_error"; + code = 404; + break; + case ERROR_TYPE_SERVER: + type_str = "server_error"; + code = 500; + break; + case ERROR_TYPE_PERMISSION: + type_str = "permission_error"; + code = 403; + break; + case ERROR_TYPE_NOT_SUPPORTED: + type_str = "not_supported_error"; + code = 501; + break; + case ERROR_TYPE_UNAVAILABLE: + type_str = "unavailable_error"; + code = 503; + break; + } + return json { + {"code", code}, + {"message", message}, + {"type", type_str}, + }; +} + +struct server_task_result_error : server_task_result { + int index = 0; + error_type err_type = ERROR_TYPE_SERVER; + std::string err_msg; + + virtual bool is_error() override { + return true; + } + + virtual json to_json() override { + return format_error_response(err_msg, err_type); + } +}; + +struct server_task_result_metrics : server_task_result { + int n_idle_slots; + int n_processing_slots; + int n_tasks_deferred; + int64_t t_start; + + int32_t kv_cache_tokens_count; + int32_t kv_cache_used_cells; + + // TODO: somehow reuse server_metrics in the future, instead of duplicating the fields + uint64_t n_prompt_tokens_processed_total = 0; + uint64_t t_prompt_processing_total = 0; + uint64_t n_tokens_predicted_total = 0; + uint64_t t_tokens_generation_total = 0; + + uint64_t n_prompt_tokens_processed = 0; + uint64_t t_prompt_processing = 0; + + uint64_t n_tokens_predicted = 0; + uint64_t t_tokens_generation = 0; + + uint64_t n_decode_total = 0; + uint64_t n_busy_slots_total = 0; + + // while we can also use std::vector this requires copying the slot object which can be quite messy + // therefore, we use json to temporarily store the slot.to_json() result + json slots_data = json::array(); + + virtual json to_json() override { + return json { + { "idle", n_idle_slots }, + { "processing", n_processing_slots }, + { "deferred", n_tasks_deferred }, + { "t_start", t_start }, + + { "n_prompt_tokens_processed_total", n_prompt_tokens_processed_total }, + { "t_tokens_generation_total", t_tokens_generation_total }, + { "n_tokens_predicted_total", n_tokens_predicted_total }, + { "t_prompt_processing_total", t_prompt_processing_total }, + + { "n_prompt_tokens_processed", n_prompt_tokens_processed }, + { "t_prompt_processing", t_prompt_processing }, + { "n_tokens_predicted", n_tokens_predicted }, + { "t_tokens_generation", t_tokens_generation }, + + { "n_decode_total", n_decode_total }, + { "n_busy_slots_total", n_busy_slots_total }, + + { "kv_cache_tokens_count", kv_cache_tokens_count }, + { "kv_cache_used_cells", kv_cache_used_cells }, + + { "slots", slots_data }, + }; + } +}; + +struct server_task_result_slot_save_load : server_task_result { + std::string filename; + bool is_save; // true = save, false = load + + size_t n_tokens; + size_t n_bytes; + double t_ms; + + virtual json to_json() override { + if (is_save) { + return json { + { "id_slot", id_slot }, + { "filename", filename }, + { "n_saved", n_tokens }, + { "n_written", n_bytes }, + { "timings", { + { "save_ms", t_ms } + }}, + }; + } else { + return json { + { "id_slot", id_slot }, + { "filename", filename }, + { "n_restored", n_tokens }, + { "n_read", n_bytes }, + { "timings", { + { "restore_ms", t_ms } + }}, + }; + } + } +}; + +struct server_task_result_slot_erase : server_task_result { + size_t n_erased; + + virtual json to_json() override { + return json { + { "id_slot", id_slot }, + { "n_erased", n_erased }, + }; + } +}; + +struct server_task_result_apply_lora : server_task_result { + virtual json to_json() override { + return json {{ "success", true }}; + } }; struct server_slot { int id; int id_task = -1; - llama_batch batch_spec; + // only used for completion/embedding/infill/rerank + server_task_type task_type = SERVER_TASK_TYPE_COMPLETION; + llama_batch batch_spec = {}; + + llama_context * ctx = nullptr; llama_context * ctx_dft = nullptr; common_speculative * spec = nullptr; @@ -153,24 +978,18 @@ struct server_slot { size_t last_nl_pos = 0; - std::string generated_text; - llama_tokens cache_tokens; - std::vector generated_token_probs; + std::string generated_text; + llama_tokens generated_tokens; - server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION; + llama_tokens cache_tokens; + + std::vector generated_token_probs; bool has_next_token = true; bool has_new_line = false; bool truncated = false; - bool stopped_eos = false; - bool stopped_word = false; - bool stopped_limit = false; + stop_type stop; - bool timings_per_token = false; - - bool oaicompat = false; - - std::string oaicompat_model; std::string stopping_word; // sampling @@ -200,18 +1019,21 @@ struct server_slot { generated_text = ""; has_new_line = false; truncated = false; - stopped_eos = false; - stopped_word = false; - stopped_limit = false; + stop = STOP_TYPE_NONE; stopping_word = ""; n_past = 0; n_sent_text = 0; n_sent_token_probs = 0; - inf_type = SERVER_TASK_INF_TYPE_COMPLETION; + task_type = SERVER_TASK_TYPE_COMPLETION; + generated_tokens.clear(); generated_token_probs.clear(); } + bool is_non_causal() const { + return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK; + } + bool has_budget(const common_params & global_params) { if (params.n_predict == -1 && global_params.n_predict == -1) { return true; // limitless @@ -255,38 +1077,40 @@ struct server_slot { } } - json get_formated_timings() const { - return json { - {"prompt_n", n_prompt_tokens_processed}, - {"prompt_ms", t_prompt_processing}, - {"prompt_per_token_ms", t_prompt_processing / n_prompt_tokens_processed}, - {"prompt_per_second", 1e3 / t_prompt_processing * n_prompt_tokens_processed}, + result_timings get_timings() const { + result_timings timings; + timings.prompt_n = n_prompt_tokens_processed; + timings.prompt_ms = t_prompt_processing; + timings.prompt_per_token_ms = t_prompt_processing / n_prompt_tokens_processed; + timings.prompt_per_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed; - {"predicted_n", n_decoded}, - {"predicted_ms", t_token_generation}, - {"predicted_per_token_ms", t_token_generation / n_decoded}, - {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, - }; + timings.predicted_n = n_decoded; + timings.predicted_ms = t_token_generation; + timings.predicted_per_token_ms = t_token_generation / n_decoded; + timings.predicted_per_second = 1e3 / t_token_generation * n_decoded; + + return timings; } - size_t find_stopping_strings(const std::string & text, const size_t last_token_size, const stop_type type) { + size_t find_stopping_strings(const std::string & text, const size_t last_token_size, bool is_full_stop) { size_t stop_pos = std::string::npos; for (const std::string & word : params.antiprompt) { size_t pos; - if (type == STOP_TYPE_FULL) { + if (is_full_stop) { const size_t tmp = word.size() + last_token_size; const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0; pos = text.find(word, from_pos); } else { + // otherwise, partial stop pos = find_partial_stop_string(word, text); } if (pos != std::string::npos && (stop_pos == std::string::npos || pos < stop_pos)) { - if (type == STOP_TYPE_FULL) { - stopped_word = true; + if (is_full_stop) { + stop = STOP_TYPE_WORD; stopping_word = word; has_next_token = false; } @@ -306,13 +1130,35 @@ struct server_slot { SLT_INF(*this, "\n" - "\rprompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n" - "\r eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n" - "\r total time = %10.2f ms / %5d tokens\n", + "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n" + " eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n" + " total time = %10.2f ms / %5d tokens\n", t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second, t_token_generation, n_decoded, t_gen, n_gen_second, t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded); } + + json to_json() const { + return json { + {"id", id}, + {"id_task", id_task}, + {"n_ctx", n_ctx}, + {"speculative", can_speculate()}, + {"is_processing", is_processing()}, + {"non_causal", is_non_causal()}, + {"params", params.to_json()}, + {"prompt", common_detokenize(ctx, prompt_tokens)}, + {"next_token", + { + {"has_next_token", has_next_token}, + {"has_new_line", has_new_line}, + {"n_remain", n_remaining}, + {"n_decoded", n_decoded}, + {"stopping_word", stopping_word}, + } + }, + }; + } }; struct server_metrics { @@ -385,9 +1231,7 @@ struct server_queue { // Add a new task to the end of the queue int post(server_task task, bool front = false) { std::unique_lock lock(mutex_tasks); - if (task.id == -1) { - task.id = id++; - } + GGML_ASSERT(task.id != -1); QUE_DBG("new task, id = %d, front = %d\n", task.id, front); if (front) { queue_tasks.push_front(std::move(task)); @@ -511,8 +1355,8 @@ struct server_response { // for keeping track of all tasks waiting for the result std::unordered_set waiting_task_ids; - // the main result queue - std::vector queue_results; + // the main result queue (using ptr for polymorphism) + std::vector queue_results; std::mutex mutex_results; std::condition_variable condition_results; @@ -552,7 +1396,7 @@ struct server_response { } // This function blocks the thread until there is a response for one of the id_tasks - server_task_result recv(const std::unordered_set & id_tasks) { + server_task_result_ptr recv(const std::unordered_set & id_tasks) { while (true) { std::unique_lock lock(mutex_results); condition_results.wait(lock, [&]{ @@ -560,8 +1404,8 @@ struct server_response { }); for (int i = 0; i < (int) queue_results.size(); i++) { - if (id_tasks.find(queue_results[i].id) != id_tasks.end()) { - server_task_result res = queue_results[i]; + if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) { + server_task_result_ptr res = std::move(queue_results[i]); queue_results.erase(queue_results.begin() + i); return res; } @@ -572,21 +1416,21 @@ struct server_response { } // single-task version of recv() - server_task_result recv(int id_task) { + server_task_result_ptr recv(int id_task) { std::unordered_set id_tasks = {id_task}; return recv(id_tasks); } // Send a new result to a waiting id_task - void send(server_task_result & result) { - SRV_DBG("sending result for task id = %d\n", result.id); + void send(server_task_result_ptr && result) { + SRV_DBG("sending result for task id = %d\n", result->id); std::unique_lock lock(mutex_results); for (const auto & id_task : waiting_task_ids) { - if (result.id == id_task) { - SRV_DBG("task id = %d moved to result queue\n", result.id); + if (result->id == id_task) { + SRV_DBG("task id = %d pushed to result queue\n", result->id); - queue_results.push_back(std::move(result)); + queue_results.emplace_back(std::move(result)); condition_results.notify_all(); return; } @@ -676,7 +1520,7 @@ struct server_context { n_ctx = llama_n_ctx(ctx); add_bos_token = llama_add_bos_token(model); - has_eos_token = !llama_add_eos_token(model); + has_eos_token = llama_token_eos(model) != LLAMA_TOKEN_NULL; if (!params_base.speculative.model.empty()) { SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str()); @@ -745,6 +1589,7 @@ struct server_context { server_slot slot; slot.id = i; + slot.ctx = ctx; slot.n_ctx = n_ctx_slot; slot.n_predict = params_base.n_predict; @@ -777,8 +1622,7 @@ struct server_context { slots.push_back(slot); } - default_generation_settings_for_props = get_formated_generation(slots.front()); - default_generation_settings_for_props["seed"] = -1; + default_generation_settings_for_props = slots[0].to_json(); // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used) @@ -865,97 +1709,14 @@ struct server_context { } bool launch_slot_with_task(server_slot & slot, const server_task & task) { - // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them) - slot_params defaults; - defaults.sampling = params_base.sampling; - defaults.speculative = params_base.speculative; + slot.reset(); + slot.id_task = task.id; + slot.index = task.index; + slot.task_type = task.type; + slot.params = std::move(task.params); + slot.prompt_tokens = std::move(task.prompt_tokens); - const auto & data = task.data; - - if (data.count("__oaicompat") != 0) { - slot.oaicompat = true; - slot.oaicompat_model = json_value(data, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - } else { - slot.oaicompat = false; - slot.oaicompat_model = ""; - } - - slot.timings_per_token = json_value(data, "timings_per_token", false); - - slot.params.stream = json_value(data, "stream", false); - slot.params.cache_prompt = json_value(data, "cache_prompt", true); - slot.params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", defaults.n_predict)); - slot.params.n_indent = json_value(data, "n_indent", defaults.n_indent); - slot.params.n_keep = json_value(data, "n_keep", defaults.n_keep); - slot.params.n_discard = json_value(data, "n_discard", defaults.n_discard); - //slot.params.t_max_prompt_ms = json_value(data, "t_max_prompt_ms", defaults.t_max_prompt_ms); // TODO: implement - slot.params.t_max_predict_ms = json_value(data, "t_max_predict_ms", defaults.t_max_predict_ms); - - slot.params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k); - slot.params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p); - slot.params.sampling.min_p = json_value(data, "min_p", defaults.sampling.min_p); - slot.params.sampling.xtc_probability = json_value(data, "xtc_probability", defaults.sampling.xtc_probability); - slot.params.sampling.xtc_threshold = json_value(data, "xtc_threshold", defaults.sampling.xtc_threshold); - slot.params.sampling.typ_p = json_value(data, "typical_p", defaults.sampling.typ_p); - slot.params.sampling.temp = json_value(data, "temperature", defaults.sampling.temp); - slot.params.sampling.dynatemp_range = json_value(data, "dynatemp_range", defaults.sampling.dynatemp_range); - slot.params.sampling.dynatemp_exponent = json_value(data, "dynatemp_exponent", defaults.sampling.dynatemp_exponent); - slot.params.sampling.penalty_last_n = json_value(data, "repeat_last_n", defaults.sampling.penalty_last_n); - slot.params.sampling.penalty_repeat = json_value(data, "repeat_penalty", defaults.sampling.penalty_repeat); - slot.params.sampling.penalty_freq = json_value(data, "frequency_penalty", defaults.sampling.penalty_freq); - slot.params.sampling.penalty_present = json_value(data, "presence_penalty", defaults.sampling.penalty_present); - slot.params.sampling.dry_multiplier = json_value(data, "dry_multiplier", defaults.sampling.dry_multiplier); - slot.params.sampling.dry_base = json_value(data, "dry_base", defaults.sampling.dry_base); - slot.params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length); - slot.params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n); - slot.params.sampling.mirostat = json_value(data, "mirostat", defaults.sampling.mirostat); - slot.params.sampling.mirostat_tau = json_value(data, "mirostat_tau", defaults.sampling.mirostat_tau); - slot.params.sampling.mirostat_eta = json_value(data, "mirostat_eta", defaults.sampling.mirostat_eta); - slot.params.sampling.penalize_nl = json_value(data, "penalize_nl", defaults.sampling.penalize_nl); - slot.params.sampling.seed = json_value(data, "seed", defaults.sampling.seed); - slot.params.sampling.n_probs = json_value(data, "n_probs", defaults.sampling.n_probs); - slot.params.sampling.min_keep = json_value(data, "min_keep", defaults.sampling.min_keep); - - slot.params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min); - slot.params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max); - slot.params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min); - - slot.params.speculative.n_min = std::min(slot.params.speculative.n_max, slot.params.speculative.n_min); - - if (slot.params.sampling.dry_base < 1.0f) { - slot.params.sampling.dry_base = defaults.sampling.dry_base; - } - - // sequence breakers for DRY - { - // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format - // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39 - - if (data.contains("dry_sequence_breakers")) { - slot.params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector()); - if (slot.params.sampling.dry_sequence_breakers.empty()) { - send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST); - return false; - } - } - } - - // process "json_schema" and "grammar" - if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) { - send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST); - return false; - } - if (data.contains("json_schema") && !data.contains("grammar")) { - try { - auto schema = json_value(data, "json_schema", json::object()); - slot.params.sampling.grammar = json_schema_to_grammar(schema); - } catch (const std::exception & e) { - send_error(task, std::string("\"json_schema\": ") + e.what(), ERROR_TYPE_INVALID_REQUEST); - return false; - } - } else { - slot.params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar); - } + SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str()); if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) { // Might be better to reject the request with a 400 ? @@ -963,78 +1724,8 @@ struct server_context { SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict); } - { - slot.params.sampling.logit_bias.clear(); - - if (json_value(data, "ignore_eos", false) && has_eos_token) { - slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY}); - } - - const auto & logit_bias = data.find("logit_bias"); - if (logit_bias != data.end() && logit_bias->is_array()) { - const int n_vocab = llama_n_vocab(model); - for (const auto & el : *logit_bias) { - // TODO: we may want to throw errors here, in case "el" is incorrect - if (el.is_array() && el.size() == 2) { - float bias; - if (el[1].is_number()) { - bias = el[1].get(); - } else if (el[1].is_boolean() && !el[1].get()) { - bias = -INFINITY; - } else { - continue; - } - - if (el[0].is_number_integer()) { - llama_token tok = el[0].get(); - if (tok >= 0 && tok < n_vocab) { - slot.params.sampling.logit_bias.push_back({tok, bias}); - } - } else if (el[0].is_string()) { - auto toks = common_tokenize(model, el[0].get(), false); - for (auto tok : toks) { - slot.params.sampling.logit_bias.push_back({tok, bias}); - } - } - } - } - } - } - - { - slot.params.antiprompt.clear(); - - const auto & stop = data.find("stop"); - if (stop != data.end() && stop->is_array()) { - for (const auto & word : *stop) { - if (!word.empty()) { - slot.params.antiprompt.push_back(word); - } - } - } - } - - { - const auto & samplers = data.find("samplers"); - if (samplers != data.end()) { - if (samplers->is_array()) { - std::vector sampler_names; - for (const auto & name : *samplers) { - if (name.is_string()) { - sampler_names.emplace_back(name); - } - } - slot.params.sampling.samplers = common_sampler_types_from_names(sampler_names, false); - } else if (samplers->is_string()){ - std::string sampler_string; - for (const auto & name : *samplers) { - sampler_string += name; - } - slot.params.sampling.samplers = common_sampler_types_from_chars(sampler_string); - } - } else { - slot.params.sampling.samplers = defaults.sampling.samplers; - } + if (slot.params.ignore_eos && has_eos_token) { + slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY}); } { @@ -1076,8 +1767,10 @@ struct server_context { const std::string token_str = common_token_to_piece(ctx, result.tok, params_base.special); slot.sampled = result.tok; - // search stop word and delete it slot.generated_text += token_str; + if (slot.params.return_tokens) { + slot.generated_tokens.push_back(result.tok); + } slot.has_next_token = true; // check if there is incomplete UTF-8 character at the end @@ -1102,20 +1795,21 @@ struct server_context { break; } + // search stop word and delete it if (!incomplete) { size_t pos = std::min(slot.n_sent_text, slot.generated_text.size()); const std::string str_test = slot.generated_text.substr(pos); bool send_text = true; - size_t stop_pos = slot.find_stopping_strings(str_test, token_str.size(), STOP_TYPE_FULL); + size_t stop_pos = slot.find_stopping_strings(str_test, token_str.size(), true); if (stop_pos != std::string::npos) { slot.generated_text.erase( slot.generated_text.begin() + pos + stop_pos, slot.generated_text.end()); pos = std::min(slot.n_sent_text, slot.generated_text.size()); } else if (slot.has_next_token) { - stop_pos = slot.find_stopping_strings(str_test, token_str.size(), STOP_TYPE_PARTIAL); + stop_pos = slot.find_stopping_strings(str_test, token_str.size(), false); send_text = stop_pos == std::string::npos; } @@ -1139,7 +1833,7 @@ struct server_context { // check the limits if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params_base)) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped by limit, n_decoded = %d, n_predict = %d\n", slot.n_decoded, slot.params.n_predict); @@ -1148,7 +1842,7 @@ struct server_context { if (slot.has_new_line) { // if we have already seen a new line, we stop after a certain time limit if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); @@ -1168,7 +1862,7 @@ struct server_context { } if (pos < slot.generated_text.size() && n_indent < slot.params.n_indent) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; // cut the last line @@ -1197,7 +1891,7 @@ struct server_context { // if context shift is disabled, we stop when it reaches the context limit if (slot.n_past >= slot.n_ctx) { slot.truncated = true; - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n", @@ -1205,7 +1899,7 @@ struct server_context { } if (llama_token_is_eog(model, result.tok)) { - slot.stopped_eos = true; + slot.stop = STOP_TYPE_EOS; slot.has_next_token = false; SLT_DBG(slot, "%s", "stopped by EOS\n"); @@ -1215,7 +1909,7 @@ struct server_context { if (slot.params.n_predict < 1 && slot.n_predict < 1 && slot.n_prompt_tokens + slot.n_decoded >= n_ctx_train) { slot.truncated = true; - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; // stop prediction SLT_WRN(slot, @@ -1229,60 +1923,6 @@ struct server_context { return slot.has_next_token; // continue } - json get_formated_generation(const server_slot & slot) const { - std::vector samplers; - samplers.reserve(slot.params.sampling.samplers.size()); - for (const auto & sampler : slot.params.sampling.samplers) { - samplers.emplace_back(common_sampler_type_to_str(sampler)); - } - - return json { - {"n_ctx", slot.n_ctx}, - {"n_predict", slot.n_predict}, // Server configured n_predict - {"model", params_base.model_alias}, - {"seed", slot.params.sampling.seed}, - {"seed_cur", slot.smpl ? common_sampler_get_seed(slot.smpl) : 0}, - {"temperature", slot.params.sampling.temp}, - {"dynatemp_range", slot.params.sampling.dynatemp_range}, - {"dynatemp_exponent", slot.params.sampling.dynatemp_exponent}, - {"top_k", slot.params.sampling.top_k}, - {"top_p", slot.params.sampling.top_p}, - {"min_p", slot.params.sampling.min_p}, - {"xtc_probability", slot.params.sampling.xtc_probability}, - {"xtc_threshold", slot.params.sampling.xtc_threshold}, - {"typical_p", slot.params.sampling.typ_p}, - {"repeat_last_n", slot.params.sampling.penalty_last_n}, - {"repeat_penalty", slot.params.sampling.penalty_repeat}, - {"presence_penalty", slot.params.sampling.penalty_present}, - {"frequency_penalty", slot.params.sampling.penalty_freq}, - {"dry_multiplier", slot.params.sampling.dry_multiplier}, - {"dry_base", slot.params.sampling.dry_base}, - {"dry_allowed_length", slot.params.sampling.dry_allowed_length}, - {"dry_penalty_last_n", slot.params.sampling.dry_penalty_last_n}, - {"dry_sequence_breakers", slot.params.sampling.dry_sequence_breakers}, - {"mirostat", slot.params.sampling.mirostat}, - {"mirostat_tau", slot.params.sampling.mirostat_tau}, - {"mirostat_eta", slot.params.sampling.mirostat_eta}, - {"penalize_nl", slot.params.sampling.penalize_nl}, - {"stop", slot.params.antiprompt}, - {"max_tokens", slot.params.n_predict}, // User configured n_predict - {"n_keep", slot.params.n_keep}, - {"n_discard", slot.params.n_discard}, - {"ignore_eos", slot.params.sampling.ignore_eos}, - {"stream", slot.params.stream}, - //{"logit_bias", slot.params.sampling.logit_bias}, - {"n_probs", slot.params.sampling.n_probs}, - {"min_keep", slot.params.sampling.min_keep}, - {"grammar", slot.params.sampling.grammar}, - {"samplers", samplers}, - {"speculative", slot.can_speculate()}, - {"speculative.n_max", slot.params.speculative.n_max}, - {"speculative.n_min", slot.params.speculative.n_min}, - {"speculative.p_min", slot.params.speculative.p_min}, - {"timings_per_token", slot.timings_per_token}, - }; - } - void send_error(const server_task & task, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) { send_error(task.id, error, type); } @@ -1294,112 +1934,107 @@ struct server_context { void send_error(const int id_task, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) { SRV_ERR("task id = %d, error: %s\n", id_task, error.c_str()); - server_task_result res; - res.id = id_task; - res.stop = false; - res.error = true; - res.data = format_error_response(error, type); + auto res = std::make_unique(); + res->id = id_task; + res->err_type = type; + res->err_msg = error; - queue_results.send(res); + queue_results.send(std::move(res)); } - void send_partial_response(server_slot & slot, completion_token_output tkn) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = false; - res.data = json { - {"content", tkn.text_to_send}, - {"stop", false}, - {"id_slot", slot.id}, - {"multimodal", false}, - {"index", slot.index}, - }; + void send_partial_response(server_slot & slot, const completion_token_output & tkn) { + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; + res->content = tkn.text_to_send; + res->tokens = { tkn.tok }; + + res->n_decoded = slot.n_decoded; + res->n_prompt_tokens = slot.n_prompt_tokens; + + res->verbose = slot.params.verbose; + res->oaicompat = slot.params.oaicompat; + res->oaicompat_chat = slot.params.oaicompat_chat; + res->oaicompat_model = slot.params.oaicompat_model; + res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id; + + // populate res.probs_output if (slot.params.sampling.n_probs > 0) { const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false); + const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size()); const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size()); std::vector probs_output; if (probs_pos < probs_stop_pos) { - probs_output = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); } - slot.n_sent_token_probs = probs_stop_pos; - - res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); } - if (slot.oaicompat) { - res.data["oaicompat_token_ctr"] = slot.n_decoded; - res.data["model"] = slot.oaicompat_model; + // populate timings if this is final response or timings_per_token is enabled + if (slot.stop != STOP_TYPE_NONE || slot.params.timings_per_token) { + res->timings = slot.get_timings(); } - if (slot.timings_per_token) { - res.data["timings"] = slot.get_formated_timings(); - } - - queue_results.send(res); + queue_results.send(std::move(res)); } - void send_final_response(const server_slot & slot) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; - res.data = json { - {"content", !slot.params.stream ? slot.generated_text : ""}, - {"id_slot", slot.id}, - {"stop", true}, - {"model", params_base.model_alias}, - {"tokens_predicted", slot.n_decoded}, - {"tokens_evaluated", slot.n_prompt_tokens}, - {"generation_settings", get_formated_generation(slot)}, - {"prompt", common_detokenize(ctx, slot.prompt_tokens)}, - {"has_new_line", slot.has_new_line}, - {"truncated", slot.truncated}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - {"tokens_cached", slot.n_past}, - {"timings", slot.get_formated_timings()}, - {"index", slot.index}, - }; + void send_final_response(server_slot & slot) { + auto res = std::make_unique(); + res->id = slot.id_task; + res->id_slot = slot.id; + res->index = slot.index; + res->content = slot.generated_text; + res->tokens = slot.generated_tokens; + res->timings = slot.get_timings(); + res->prompt = common_detokenize(ctx, slot.prompt_tokens, true); + + res->truncated = slot.truncated; + res->n_decoded = slot.n_decoded; + res->n_prompt_tokens = slot.n_prompt_tokens; + res->n_tokens_cached = slot.n_past; + res->has_new_line = slot.has_new_line; + res->stopping_word = slot.stopping_word; + res->stop = slot.stop; + + res->verbose = slot.params.verbose; + res->stream = slot.params.stream; + res->oaicompat = slot.params.oaicompat; + res->oaicompat_chat = slot.params.oaicompat_chat; + res->oaicompat_model = slot.params.oaicompat_model; + res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id; + + // populate res.probs_output if (slot.params.sampling.n_probs > 0) { - std::vector probs; - if (!slot.params.stream && slot.stopped_word) { + if (!slot.params.stream && slot.stop == STOP_TYPE_WORD) { const llama_tokens stop_word_toks = common_tokenize(ctx, slot.stopping_word, false); size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size()); - probs = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin(), slot.generated_token_probs.end() - safe_offset); } else { - probs = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin(), slot.generated_token_probs.end()); } - - res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs); } - if (slot.oaicompat) { - res.data["oaicompat_token_ctr"] = slot.n_decoded; - res.data["model"] = slot.oaicompat_model; - } + res->generation_params = slot.params; // copy the parameters - queue_results.send(res); + queue_results.send(std::move(res)); } void send_embedding(const server_slot & slot, const llama_batch & batch) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; + res->n_tokens = slot.n_prompt_tokens; + res->oaicompat = slot.params.oaicompat; const int n_embd = llama_n_embd(model); @@ -1418,32 +2053,30 @@ struct server_context { if (embd == NULL) { SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]); - res.data = json { - {"embedding", std::vector(n_embd, 0.0f)}, - {"index", slot.index}, - }; - + res->embedding.push_back(std::vector(n_embd, 0.0f)); continue; } - common_embd_normalize(embd, embd_res.data(), n_embd); - - res.data = json { - {"embedding", embd_res}, - {"index", slot.index}, - }; + // normalize only when there is pooling + // TODO: configurable + if (llama_pooling_type(slot.ctx) != LLAMA_POOLING_TYPE_NONE) { + common_embd_normalize(embd, embd_res.data(), n_embd, 2); + res->embedding.push_back(embd_res); + } else { + res->embedding.push_back({ embd, embd + n_embd }); + } } SLT_DBG(slot, "%s", "sending embeddings\n"); - queue_results.send(res); + queue_results.send(std::move(res)); } void send_rerank(const server_slot & slot, const llama_batch & batch) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; + res->n_tokens = slot.n_prompt_tokens; for (int i = 0; i < batch.n_tokens; ++i) { if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) { @@ -1458,104 +2091,29 @@ struct server_context { if (embd == NULL) { SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]); - res.data = json { - {"index", slot.index}, - {"score", -1e6}, - }; - + res->score = -1e6; continue; } - res.data = json { - {"index", slot.index}, - {"score", embd[0]}, - }; + res->score = embd[0]; } - SLT_DBG(slot, "sending rerank result, res = '%s'\n", res.data.dump().c_str()); + SLT_DBG(slot, "sending rerank result, res.score = %f\n", res->score); - queue_results.send(res); + queue_results.send(std::move(res)); } // // Functions to create new task(s) and receive result(s) // - // break the input "prompt" into multiple tasks if needed, then format and tokenize the input prompt(s) - std::vector create_tasks_inference(json data, server_task_inf_type inf_type) { - std::vector tasks; - auto create_task = [&](json & task_data, llama_tokens & prompt_tokens) { - SRV_DBG("create task, n_tokens = %d\n", (int) prompt_tokens.size()); - server_task task; - task.id = queue_tasks.get_new_id(); - task.inf_type = inf_type; - task.type = SERVER_TASK_TYPE_INFERENCE; - task.data = task_data; - task.prompt_tokens = std::move(prompt_tokens); - tasks.push_back(std::move(task)); - }; - - static constexpr const char * error_msg = "\"prompt\" must be a string, an array of token ids or an array of prompts"; - if (!data.contains("prompt")) { - throw std::runtime_error(error_msg); - } - - // because llama_tokenize api is thread-safe, we can tokenize the prompt from HTTP thread - bool add_special = inf_type != SERVER_TASK_INF_TYPE_RERANK && inf_type != SERVER_TASK_INF_TYPE_INFILL; - std::vector tokenized_prompts = tokenize_input_prompts(ctx, data.at("prompt"), add_special, true); - switch (inf_type) { - case SERVER_TASK_INF_TYPE_RERANK: - { - // prompts[0] is the question - // the rest are the answers/documents - GGML_ASSERT(tokenized_prompts.size() > 1); - SRV_DBG("creating rerank tasks, n_prompts = %d\n", (int) tokenized_prompts.size() - 1); - for (size_t i = 1; i < tokenized_prompts.size(); i++) { - data["index"] = i - 1; - auto tokens = format_rerank(model, tokenized_prompts[0], tokenized_prompts[i]); - create_task(data, tokens); - } - } break; - case SERVER_TASK_INF_TYPE_INFILL: - { - SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); - for (size_t i = 0; i < tokenized_prompts.size(); i++) { - data["index"] = i; - auto tokens = format_infill( - ctx, - data.at("input_prefix"), - data.at("input_suffix"), - data.at("input_extra"), - params_base.n_batch, - params_base.n_predict, - slots[0].n_ctx, // TODO: there should be a better way - params_base.spm_infill, - tokenized_prompts[i] - ); - create_task(data, tokens); - } - } break; - default: - { - SRV_DBG("creating multi-prompt tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); - for (size_t i = 0; i < tokenized_prompts.size(); i++) { - data["index"] = i; - create_task(data, tokenized_prompts[i]); - } - } - } - - return tasks; - } - void cancel_tasks(const std::unordered_set & id_tasks) { std::vector cancel_tasks; cancel_tasks.reserve(id_tasks.size()); for (const auto & id_task : id_tasks) { SRV_WRN("cancel task, id_task = %d\n", id_task); - server_task task; - task.type = SERVER_TASK_TYPE_CANCEL; + server_task task(SERVER_TASK_TYPE_CANCEL); task.id_target = id_task; cancel_tasks.push_back(task); queue_results.remove_waiting_task_id(id_task); @@ -1564,50 +2122,58 @@ struct server_context { queue_tasks.post(cancel_tasks, true); } - // receive the results from task(s) created by create_tasks_inference - void receive_cmpl_results( + // receive the results from task(s) + void receive_multi_results( const std::unordered_set & id_tasks, - const std::function&)> & result_handler, + const std::function&)> & result_handler, const std::function & error_handler) { - // TODO: currently, there is no way to detect the client has cancelled the request - std::vector results(id_tasks.size()); + std::vector results(id_tasks.size()); for (size_t i = 0; i < id_tasks.size(); i++) { - server_task_result result = queue_results.recv(id_tasks); + server_task_result_ptr result = queue_results.recv(id_tasks); - if (result.error) { - error_handler(result.data); + if (result->is_error()) { + error_handler(result->to_json()); cancel_tasks(id_tasks); return; } - const size_t idx = result.data["index"]; + GGML_ASSERT( + dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + ); + const size_t idx = result->get_index(); GGML_ASSERT(idx < results.size() && "index out of range"); - - results[idx] = result; + results[idx] = std::move(result); } result_handler(results); } - // receive the results from task(s) created by create_tasks_inference, in stream mode + // receive the results from task(s), in stream mode void receive_cmpl_results_stream( - const std::unordered_set & id_tasks, const - std::function & result_handler, const - std::function & error_handler) { + const std::unordered_set & id_tasks, + const std::function & result_handler, + const std::function & error_handler) { size_t n_finished = 0; while (true) { - server_task_result result = queue_results.recv(id_tasks); + server_task_result_ptr result = queue_results.recv(id_tasks); + + if (result->is_error()) { + error_handler(result->to_json()); + cancel_tasks(id_tasks); + return; + } + + GGML_ASSERT( + dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + ); if (!result_handler(result)) { cancel_tasks(id_tasks); break; } - if (result.error) { - error_handler(result.data); - cancel_tasks(id_tasks); - break; - } - - if (result.stop) { + if (result->is_stop()) { if (++n_finished == id_tasks.size()) { break; } @@ -1621,9 +2187,12 @@ struct server_context { void process_single_task(server_task task) { switch (task.type) { - case SERVER_TASK_TYPE_INFERENCE: + case SERVER_TASK_TYPE_COMPLETION: + case SERVER_TASK_TYPE_INFILL: + case SERVER_TASK_TYPE_EMBEDDING: + case SERVER_TASK_TYPE_RERANK: { - const int id_slot = json_value(task.data, "id_slot", -1); + const int id_slot = task.id_selected_slot; server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task); @@ -1640,13 +2209,6 @@ struct server_context { break; } - slot->reset(); - - slot->id_task = task.id; - slot->inf_type = task.inf_type; - slot->index = json_value(task.data, "index", 0); - slot->prompt_tokens = std::move(task.prompt_tokens); - if (!launch_slot_with_task(*slot, task)) { SRV_ERR("failed to launch slot with task, id_task = %d\n", task.id); break; @@ -1674,21 +2236,7 @@ struct server_context { int n_processing_slots = 0; for (server_slot & slot : slots) { - json slot_data = get_formated_generation(slot); - slot_data["id"] = slot.id; - slot_data["id_task"] = slot.id_task; - slot_data["is_processing"] = slot.is_processing(); - slot_data["prompt"] = common_detokenize(ctx, slot.prompt_tokens); - slot_data["next_token"] = { - {"has_next_token", slot.has_next_token}, - {"has_new_line", slot.has_new_line}, - {"n_remain", slot.n_remaining}, - {"n_decoded", slot.n_decoded}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - }; + json slot_data = slot.to_json(); if (slot.is_processing()) { n_processing_slots++; @@ -1700,43 +2248,38 @@ struct server_context { } SRV_DBG("n_idle_slots = %d, n_processing_slots = %d\n", n_idle_slots, n_processing_slots); - server_task_result res; - res.id = task.id; - res.stop = true; - res.error = false; - res.data = { - { "idle", n_idle_slots }, - { "processing", n_processing_slots }, - { "deferred", queue_tasks.queue_tasks_deferred.size() }, - { "t_start", metrics.t_start}, + auto res = std::make_unique(); + res->id = task.id; + res->slots_data = std::move(slots_data); + res->n_idle_slots = n_idle_slots; + res->n_processing_slots = n_processing_slots; + res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size(); + res->t_start = metrics.t_start; - { "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total}, - { "t_tokens_generation_total", metrics.t_tokens_generation_total}, - { "n_tokens_predicted_total", metrics.n_tokens_predicted_total}, - { "t_prompt_processing_total", metrics.t_prompt_processing_total}, + res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx); + res->kv_cache_used_cells = llama_get_kv_cache_used_cells(ctx); - { "n_prompt_tokens_processed", metrics.n_prompt_tokens_processed}, - { "t_prompt_processing", metrics.t_prompt_processing}, - { "n_tokens_predicted", metrics.n_tokens_predicted}, - { "t_tokens_generation", metrics.t_tokens_generation}, + res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total; + res->t_prompt_processing_total = metrics.t_prompt_processing_total; + res->n_tokens_predicted_total = metrics.n_tokens_predicted_total; + res->t_tokens_generation_total = metrics.t_tokens_generation_total; - { "n_decode_total", metrics.n_decode_total}, - { "n_busy_slots_total", metrics.n_busy_slots_total}, + res->n_prompt_tokens_processed = metrics.n_prompt_tokens_processed; + res->t_prompt_processing = metrics.t_prompt_processing; + res->n_tokens_predicted = metrics.n_tokens_predicted; + res->t_tokens_generation = metrics.t_tokens_generation; - { "kv_cache_tokens_count", llama_get_kv_cache_token_count(ctx)}, - { "kv_cache_used_cells", llama_get_kv_cache_used_cells(ctx)}, + res->n_decode_total = metrics.n_decode_total; + res->n_busy_slots_total = metrics.n_busy_slots_total; - { "slots", slots_data }, - }; - - if (json_value(task.data, "reset_bucket", false)) { + if (task.metrics_reset_bucket) { metrics.reset_bucket(); } - queue_results.send(res); + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_SAVE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1752,32 +2295,27 @@ struct server_context { const size_t token_count = slot->cache_tokens.size(); const int64_t t_start = ggml_time_us(); - std::string filename = task.data.at("filename"); - std::string filepath = task.data.at("filepath"); + std::string filename = task.slot_action.filename; + std::string filepath = task.slot_action.filepath; const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), token_count); const int64_t t_end = ggml_time_us(); const double t_save_ms = (t_end - t_start) / 1000.0; - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "filename", filename }, - { "n_saved", token_count }, // tokens saved - { "n_written", nwrite }, // bytes written - { "timings", { - { "save_ms", t_save_ms } - } } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->filename = filename; + res->is_save = true; + res->n_tokens = token_count; + res->n_bytes = nwrite; + res->t_ms = t_save_ms; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_RESTORE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1792,8 +2330,8 @@ struct server_context { const int64_t t_start = ggml_time_us(); - std::string filename = task.data.at("filename"); - std::string filepath = task.data.at("filepath"); + std::string filename = task.slot_action.filename; + std::string filepath = task.slot_action.filepath; slot->cache_tokens.resize(slot->n_ctx); size_t token_count = 0; @@ -1808,24 +2346,19 @@ struct server_context { const int64_t t_end = ggml_time_us(); const double t_restore_ms = (t_end - t_start) / 1000.0; - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "filename", filename }, - { "n_restored", token_count }, // tokens restored - { "n_read", nread }, // bytes read - { "timings", { - { "restore_ms", t_restore_ms } - } } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->filename = filename; + res->is_save = false; + res->n_tokens = token_count; + res->n_bytes = nread; + res->t_ms = t_restore_ms; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_ERASE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1843,25 +2376,18 @@ struct server_context { llama_kv_cache_seq_rm(ctx, slot->id, -1, -1); slot->cache_tokens.clear(); - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "n_erased", n_erased } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->n_erased = n_erased; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SET_LORA: { common_lora_adapters_apply(ctx, loras); - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json{{ "success", true }}; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + queue_results.send(std::move(res)); } break; } } @@ -1891,10 +2417,8 @@ struct server_context { { SRV_DBG("%s", "posting NEXT_RESPONSE\n"); - server_task task; - task.type = SERVER_TASK_TYPE_NEXT_RESPONSE; - task.id_target = -1; - + server_task task(SERVER_TASK_TYPE_NEXT_RESPONSE); + task.id = queue_tasks.get_new_id(); queue_tasks.post(task); } @@ -2008,7 +2532,7 @@ struct server_context { continue; } - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.is_non_causal()) { if (slot.n_prompt_tokens > n_ubatch) { slot.release(); send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER); @@ -2123,7 +2647,7 @@ struct server_context { } // non-causal tasks require to fit the entire prompt in the physical batch - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.is_non_causal()) { // cannot fit the prompt in the current batch - will try next iter if (batch.n_tokens + slot.n_prompt_tokens > n_batch) { continue; @@ -2131,10 +2655,7 @@ struct server_context { } // check that we are in the right batch_type, if not defer the slot - const bool slot_type = - slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || - slot.inf_type == SERVER_TASK_INF_TYPE_RERANK ? 1 : 0; - + int slot_type = slot.is_non_causal(); if (batch_type == -1) { batch_type = slot_type; } else if (batch_type != slot_type) { @@ -2157,7 +2678,10 @@ struct server_context { // add prompt tokens for processing in the current batch while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) { - common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, false); + // without pooling, we want to output the embeddings for all the tokens in the batch + const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE; + + common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, need_embd); if (slot.params.cache_prompt) { slot.cache_tokens.push_back(prompt_tokens[slot.n_past]); @@ -2251,7 +2775,7 @@ struct server_context { } if (slot.state == SLOT_STATE_DONE_PROMPT) { - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING) { + if (slot.task_type == SERVER_TASK_TYPE_EMBEDDING) { // prompt evaluated for embedding send_embedding(slot, batch_view); slot.release(); @@ -2259,7 +2783,7 @@ struct server_context { continue; // continue loop of slots } - if (slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.task_type == SERVER_TASK_TYPE_RERANK) { send_rerank(slot, batch_view); slot.release(); slot.i_batch = -1; @@ -2296,9 +2820,11 @@ struct server_context { const auto * cur_p = common_sampler_get_candidates(slot.smpl); for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) { + auto tok_id = cur_p->data[i].id; result.probs.push_back({ - cur_p->data[i].id, - i >= cur_p->size ? 0.0f : cur_p->data[i].p, + tok_id, + tokens_to_output_formatted_string(ctx, tok_id), + i >= cur_p->size ? 0.0f : cur_p->data[i].p, }); } @@ -2322,10 +2848,29 @@ struct server_context { continue; } + // determine the max draft that fits the current slot state + int n_draft_max = slot.params.speculative.n_max; + + // note: n_past is not yet increased for the `id` token sampled above + // also, need to leave space for 1 extra token to allow context shifts + n_draft_max = std::min(n_draft_max, slot.n_ctx - slot.n_past - 2); + + if (slot.n_remaining > 0) { + n_draft_max = std::min(n_draft_max, slot.n_remaining - 1); + } + + SLT_DBG(slot, "max possible draft: %d\n", n_draft_max); + + if (n_draft_max < slot.params.speculative.n_min) { + SLT_DBG(slot, "the max possible draft is too small: %d < %d - skipping speculative decoding\n", n_draft_max, slot.params.speculative.n_min); + + continue; + } + llama_token id = slot.sampled; struct common_speculative_params params_spec; - params_spec.n_draft = slot.params.speculative.n_max; + params_spec.n_draft = n_draft_max; params_spec.n_reuse = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max; params_spec.p_min = slot.params.speculative.p_min; @@ -2333,6 +2878,8 @@ struct server_context { // ignore small drafts if (slot.params.speculative.n_min > (int) draft.size()) { + SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min); + continue; } @@ -2344,6 +2891,8 @@ struct server_context { common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true); } + SLT_DBG(slot, "decoding speculative batch, size = %d\n", slot.batch_spec.n_tokens); + llama_decode(ctx, slot.batch_spec); // the accepted tokens from the speculation @@ -2372,7 +2921,7 @@ struct server_context { } } - SRV_DBG("accepted %d/%d draft tokens\n", (int) ids.size() - 1, (int) draft.size()); + SLT_DBG(slot, "accepted %d/%d draft tokens, new n_past = %d\n", (int) ids.size() - 1, (int) draft.size(), slot.n_past); } } @@ -2427,17 +2976,9 @@ int main(int argc, char ** argv) { common_init(); - // enabling this will output extra debug information in the HTTP responses from the server - // see format_final_response_oaicompat() - const bool verbose = params.verbosity > 9; - // struct that contains llama context and inference server_context ctx_server; - if (params.model_alias == "unknown") { - params.model_alias = params.model; - } - llama_backend_init(); llama_numa_init(params.numa); @@ -2472,20 +3013,20 @@ int main(int argc, char ** argv) { auto res_error = [](httplib::Response & res, const json & error_data) { json final_response {{"error", error_data}}; - res.set_content(final_response.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON); + res.set_content(safe_json_to_str(final_response), MIMETYPE_JSON); res.status = json_value(error_data, "code", 500); }; auto res_ok = [](httplib::Response & res, const json & data) { - res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON); + res.set_content(safe_json_to_str(data), MIMETYPE_JSON); res.status = 200; }; - svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { + svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) { std::string message; try { std::rethrow_exception(ep); - } catch (std::exception & e) { + } catch (const std::exception & e) { message = e.what(); } catch (...) { message = "Unknown Exception"; @@ -2614,27 +3155,33 @@ int main(int argc, char ** argv) { } // request slots data using task queue - server_task task; + server_task task(SERVER_TASK_TYPE_METRICS); task.id = ctx_server.queue_tasks.get_new_id(); - task.type = SERVER_TASK_TYPE_METRICS; - ctx_server.queue_results.add_waiting_task_id(task.id); ctx_server.queue_tasks.post(task, true); // high-priority task // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); ctx_server.queue_results.remove_waiting_task_id(task.id); + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } + + // TODO: get rid of this dynamic_cast + auto res_metrics = dynamic_cast(result.get()); + GGML_ASSERT(res_metrics != nullptr); + // optionally return "fail_on_no_slot" error - const int n_idle_slots = result.data.at("idle"); if (req.has_param("fail_on_no_slot")) { - if (n_idle_slots == 0) { + if (res_metrics->n_idle_slots == 0) { res_error(res, format_error_response("no slot available", ERROR_TYPE_UNAVAILABLE)); return; } } - res_ok(res, result.data.at("slots")); + res_ok(res, res_metrics->slots_data); }; const auto handle_metrics = [&](const httplib::Request &, httplib::Response & res) { @@ -2644,83 +3191,77 @@ int main(int argc, char ** argv) { } // request slots data using task queue - server_task task; + server_task task(SERVER_TASK_TYPE_METRICS); task.id = ctx_server.queue_tasks.get_new_id(); - task.id_target = -1; - task.type = SERVER_TASK_TYPE_METRICS; - task.data.push_back({{"reset_bucket", true}}); + task.metrics_reset_bucket = true; ctx_server.queue_results.add_waiting_task_id(task.id); ctx_server.queue_tasks.post(task, true); // high-priority task // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); ctx_server.queue_results.remove_waiting_task_id(task.id); - json data = result.data; + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } - const uint64_t n_prompt_tokens_processed = data.at("n_prompt_tokens_processed"); - const uint64_t t_prompt_processing = data.at("t_prompt_processing"); - - const uint64_t n_tokens_predicted = data.at("n_tokens_predicted"); - const uint64_t t_tokens_generation = data.at("t_tokens_generation"); - - const uint64_t n_decode_total = data.at("n_decode_total"); - const uint64_t n_busy_slots_total = data.at("n_busy_slots_total"); - - const int32_t kv_cache_used_cells = data.at("kv_cache_used_cells"); + // TODO: get rid of this dynamic_cast + auto res_metrics = dynamic_cast(result.get()); + GGML_ASSERT(res_metrics != nullptr); // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names json all_metrics_def = json { {"counter", {{ {"name", "prompt_tokens_total"}, {"help", "Number of prompt tokens processed."}, - {"value", (uint64_t) data.at("n_prompt_tokens_processed_total")} + {"value", (uint64_t) res_metrics->n_prompt_tokens_processed_total} }, { {"name", "prompt_seconds_total"}, {"help", "Prompt process time"}, - {"value", (uint64_t) data.at("t_prompt_processing_total") / 1.e3} + {"value", (uint64_t) res_metrics->t_prompt_processing_total / 1.e3} }, { {"name", "tokens_predicted_total"}, {"help", "Number of generation tokens processed."}, - {"value", (uint64_t) data.at("n_tokens_predicted_total")} + {"value", (uint64_t) res_metrics->n_tokens_predicted_total} }, { {"name", "tokens_predicted_seconds_total"}, {"help", "Predict process time"}, - {"value", (uint64_t) data.at("t_tokens_generation_total") / 1.e3} + {"value", (uint64_t) res_metrics->t_tokens_generation_total / 1.e3} }, { {"name", "n_decode_total"}, {"help", "Total number of llama_decode() calls"}, - {"value", n_decode_total} + {"value", res_metrics->n_decode_total} }, { {"name", "n_busy_slots_per_decode"}, {"help", "Average number of busy slots per llama_decode() call"}, - {"value", (float) n_busy_slots_total / (float) n_decode_total} + {"value", (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total} }}}, {"gauge", {{ {"name", "prompt_tokens_seconds"}, {"help", "Average prompt throughput in tokens/s."}, - {"value", n_prompt_tokens_processed ? 1.e3 / t_prompt_processing * n_prompt_tokens_processed : 0.} + {"value", res_metrics->n_prompt_tokens_processed ? 1.e3 / res_metrics->t_prompt_processing * res_metrics->n_prompt_tokens_processed : 0.} },{ {"name", "predicted_tokens_seconds"}, {"help", "Average generation throughput in tokens/s."}, - {"value", n_tokens_predicted ? 1.e3 / t_tokens_generation * n_tokens_predicted : 0.} + {"value", res_metrics->n_tokens_predicted ? 1.e3 / res_metrics->t_tokens_generation * res_metrics->n_tokens_predicted : 0.} },{ {"name", "kv_cache_usage_ratio"}, {"help", "KV-cache usage. 1 means 100 percent usage."}, - {"value", 1. * kv_cache_used_cells / params.n_ctx} + {"value", 1. * res_metrics->kv_cache_used_cells / params.n_ctx} },{ {"name", "kv_cache_tokens"}, {"help", "KV-cache tokens."}, - {"value", (uint64_t) data.at("kv_cache_tokens_count")} + {"value", (uint64_t) res_metrics->kv_cache_tokens_count} },{ {"name", "requests_processing"}, {"help", "Number of request processing."}, - {"value", (uint64_t) data.at("processing")} + {"value", (uint64_t) res_metrics->n_processing_slots} },{ {"name", "requests_deferred"}, {"help", "Number of request deferred."}, - {"value", (uint64_t) data.at("deferred")} + {"value", (uint64_t) res_metrics->n_tasks_deferred} }}} }; @@ -2741,8 +3282,7 @@ int main(int argc, char ** argv) { } } - const int64_t t_start = data.at("t_start"); - res.set_header("Process-Start-Time-Unix", std::to_string(t_start)); + res.set_header("Process-Start-Time-Unix", std::to_string(res_metrics->t_start)); res.set_content(prometheus.str(), "text/plain; version=0.0.4"); res.status = 200; // HTTP OK @@ -2757,25 +3297,24 @@ int main(int argc, char ** argv) { } std::string filepath = params.slot_save_path + filename; - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_SAVE; - task.data = { - { "id_slot", id_slot }, - { "filename", filename }, - { "filepath", filepath }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_SAVE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; + task.slot_action.filename = filename; + task.slot_action.filepath = filepath; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + res_ok(res, result->to_json()); }; const auto handle_slots_restore = [&ctx_server, &res_error, &res_ok, ¶ms](const httplib::Request & req, httplib::Response & res, int id_slot) { @@ -2787,45 +3326,45 @@ int main(int argc, char ** argv) { } std::string filepath = params.slot_save_path + filename; - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_RESTORE; - task.data = { - { "id_slot", id_slot }, - { "filename", filename }, - { "filepath", filepath }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_RESTORE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; + task.slot_action.filename = filename; + task.slot_action.filepath = filepath; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; const auto handle_slots_erase = [&ctx_server, &res_error, &res_ok](const httplib::Request & /* req */, httplib::Response & res, int id_slot) { - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_ERASE; - task.data = { - { "id_slot", id_slot }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_ERASE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; const auto handle_slots_action = [¶ms, &res_error, &handle_slots_save, &handle_slots_restore, &handle_slots_erase](const httplib::Request & req, httplib::Response & res) { @@ -2858,9 +3397,11 @@ int main(int argc, char ** argv) { }; const auto handle_props = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) { + // this endpoint is publicly available, please only return what is safe to be exposed json data = { { "default_generation_settings", ctx_server.default_generation_settings_for_props }, { "total_slots", ctx_server.params_base.n_parallel }, + { "model_path", ctx_server.params_base.model }, { "chat_template", llama_get_chat_template(ctx_server.model) }, }; @@ -2880,13 +3421,50 @@ int main(int argc, char ** argv) { res_ok(res, {{ "success", true }}); }; - const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) { + // handle completion-like requests (completion, chat, infill) + // we can optionally provide a custom format for partial results and final results + const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok]( + server_task_type type, + json & data, + httplib::Response & res, + bool oaicompat = false, + bool oaicompat_chat = false) { + GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL); + if (ctx_server.params_base.embedding) { res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; } - std::vector tasks = ctx_server.create_tasks_inference(data, inf_type); + auto completion_id = gen_chatcmplid(); + std::vector tasks; + + try { + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, data.at("prompt"), true, true); + tasks.reserve(tokenized_prompts.size()); + for (size_t i = 0; i < tokenized_prompts.size(); i++) { + server_task task = server_task(type); + + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + + task.prompt_tokens = std::move(tokenized_prompts[i]); + task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.ctx, ctx_server.params_base, data); + task.id_selected_slot = json_value(data, "id_slot", -1); + + // OAI-compat + task.params.oaicompat = oaicompat; + task.params.oaicompat_chat = oaicompat_chat; + task.params.oaicompat_cmpl_id = completion_id; + // oaicompat_model is already populated by params_from_json_cmpl + + tasks.push_back(task); + } + } catch (const std::exception & e) { + res_error(res, format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST)); + return; + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); @@ -2894,15 +3472,15 @@ int main(int argc, char ** argv) { const auto task_ids = server_task::get_list_id(tasks); if (!stream) { - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { if (results.size() == 1) { // single result - res_ok(res, results[0].data); + res_ok(res, results[0]->to_json()); } else { // multiple results (multitask) json arr = json::array(); - for (const auto & res : results) { - arr.push_back(res.data); + for (auto & res : results) { + arr.push_back(res->to_json()); } res_ok(res, arr); } @@ -2912,12 +3490,26 @@ int main(int argc, char ** argv) { ctx_server.queue_results.remove_waiting_task_ids(task_ids); } else { - const auto chunked_content_provider = [task_ids, &ctx_server](size_t, httplib::DataSink & sink) { - ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool { - return server_sent_event(sink, "data", result.data); + const auto chunked_content_provider = [task_ids, &ctx_server, oaicompat](size_t, httplib::DataSink & sink) { + ctx_server.receive_cmpl_results_stream(task_ids, [&](server_task_result_ptr & result) -> bool { + json res_json = result->to_json(); + if (res_json.is_array()) { + for (const auto & res : res_json) { + if (!server_sent_event(sink, "data", res)) { + return false; + } + } + return true; + } else { + return server_sent_event(sink, "data", res_json); + } }, [&](const json & error_data) { server_sent_event(sink, "error", error_data); }); + if (oaicompat) { + static const std::string ev_done = "data: [DONE]\n\n"; + sink.write(ev_done.data(), ev_done.size()); + } sink.done(); return false; }; @@ -2932,7 +3524,12 @@ int main(int argc, char ** argv) { const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) { json data = json::parse(req.body); - return handle_completions_generic(SERVER_TASK_INF_TYPE_COMPLETION, data, res); + return handle_completions_generic( + SERVER_TASK_TYPE_COMPLETION, + data, + res, + /* oaicompat */ false, + /* oaicompat_chat */ false); }; const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) { @@ -2955,6 +3552,11 @@ int main(int argc, char ** argv) { json data = json::parse(req.body); // validate input + if (data.contains("prompt") && !data.at("prompt").is_string()) { + // prompt is optional + res_error(res, format_error_response("\"prompt\" must be a string", ERROR_TYPE_INVALID_REQUEST)); + } + if (!data.contains("input_prefix")) { res_error(res, format_error_response("\"input_prefix\" is required", ERROR_TYPE_INVALID_REQUEST)); } @@ -2964,9 +3566,11 @@ int main(int argc, char ** argv) { } if (data.contains("input_extra") && !data.at("input_extra").is_array()) { + // input_extra is optional res_error(res, format_error_response("\"input_extra\" must be an array of {\"filename\": string, \"text\": string}", ERROR_TYPE_INVALID_REQUEST)); return; } + json input_extra = json_value(data, "input_extra", json::array()); for (const auto & chunk : input_extra) { // { "text": string, "filename": string } @@ -2982,67 +3586,40 @@ int main(int argc, char ** argv) { } data["input_extra"] = input_extra; // default to empty array if it's not exist - return handle_completions_generic(SERVER_TASK_INF_TYPE_INFILL, data, res); + std::string prompt = json_value(data, "prompt", std::string()); + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true); + SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); + data["prompt"] = format_infill( + ctx_server.ctx, + data.at("input_prefix"), + data.at("input_suffix"), + data.at("input_extra"), + ctx_server.params_base.n_batch, + ctx_server.params_base.n_predict, + ctx_server.slots[0].n_ctx, // TODO: there should be a better way + ctx_server.params_base.spm_infill, + tokenized_prompts[0] + ); + + return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res); }; - // TODO: maybe merge this function with "handle_completions_generic" - const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) { + const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) { if (ctx_server.params_base.embedding) { res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; } json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template); - - std::vector tasks = ctx_server.create_tasks_inference(data, SERVER_TASK_INF_TYPE_COMPLETION); - ctx_server.queue_results.add_waiting_tasks(tasks); - ctx_server.queue_tasks.post(tasks); - - bool stream = json_value(data, "stream", false); - const auto task_ids = server_task::get_list_id(tasks); - const auto completion_id = gen_chatcmplid(); - - if (!stream) { - ctx_server.receive_cmpl_results(task_ids, [&](const std::vector & results) { - // multitask is never support in chat completion, there is only one result - json result_oai = format_final_response_oaicompat(data, results[0].data, completion_id, /*.streaming =*/ false, verbose); - res_ok(res, result_oai); - }, [&](const json & error_data) { - res_error(res, error_data); - }); - - ctx_server.queue_results.remove_waiting_task_ids(task_ids); - } else { - const auto chunked_content_provider = [task_ids, &ctx_server, completion_id](size_t, httplib::DataSink & sink) { - ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool { - std::vector result_array = format_partial_response_oaicompat(result.data, completion_id); - for (auto & event_data : result_array) { - if (event_data.empty()) { - continue; // skip the stop token - } - if (!server_sent_event(sink, "data", event_data)) { - return false; // connection is closed - } - } - return true; // ok - }, [&](const json & error_data) { - server_sent_event(sink, "error", error_data); - }); - static const std::string ev_done = "data: [DONE]\n\n"; - sink.write(ev_done.data(), ev_done.size()); - sink.done(); - return true; - }; - - auto on_complete = [task_ids, &ctx_server] (bool) { - ctx_server.queue_results.remove_waiting_task_ids(task_ids); - }; - - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } + return handle_completions_generic( + SERVER_TASK_TYPE_COMPLETION, + data, + res, + /* oaicompat */ true, + /* oaicompat_chat */ true); }; - const auto handle_models = [¶ms, &ctx_server](const httplib::Request &, httplib::Response & res) { + const auto handle_models = [¶ms, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) { json models = { {"object", "list"}, {"data", { @@ -3056,7 +3633,7 @@ int main(int argc, char ** argv) { }} }; - res.set_content(models.dump(), MIMETYPE_JSON); + res_ok(res, models); }; const auto handle_tokenize = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) { @@ -3112,37 +3689,63 @@ int main(int argc, char ** argv) { res_ok(res, data); }; - const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) { + const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, bool oaicompat) { const json body = json::parse(req.body); - bool is_openai = false; - // an input prompt can be a string or a list of tokens (integer) + if (oaicompat && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) { + res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST)); + return; + } + + // for the shape of input/content, see tokenize_input_prompts() json prompt; if (body.count("input") != 0) { - is_openai = true; prompt = body.at("input"); - } else if (body.count("content") != 0) { - // with "content", we only support single prompt - prompt = std::vector{body.at("content")}; + } else if (body.contains("content")) { + oaicompat = false; + prompt = body.at("content"); } else { res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST)); return; } + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true); + for (const auto & tokens : tokenized_prompts) { + // this check is necessary for models that do not add BOS token to the input + if (tokens.empty()) { + res_error(res, format_error_response("Input content cannot be empty", ERROR_TYPE_INVALID_REQUEST)); + return; + } + } + // create and queue the task json responses = json::array(); bool error = false; { - std::vector tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_EMBEDDING); + std::vector tasks; + for (size_t i = 0; i < tokenized_prompts.size(); i++) { + server_task task = server_task(SERVER_TASK_TYPE_EMBEDDING); + + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + task.prompt_tokens = std::move(tokenized_prompts[i]); + + // OAI-compat + task.params.oaicompat = oaicompat; + + tasks.push_back(task); + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); // get the result std::unordered_set task_ids = server_task::get_list_id(tasks); - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { - for (const auto & res : results) { - responses.push_back(res.data); + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { + for (auto & res : results) { + GGML_ASSERT(dynamic_cast(res.get()) != nullptr); + responses.push_back(res->to_json()); } }, [&](const json & error_data) { res_error(res, error_data); @@ -3157,12 +3760,18 @@ int main(int argc, char ** argv) { } // write JSON response - json root = is_openai - ? format_embeddings_response_oaicompat(body, responses) - : responses[0]; + json root = oaicompat ? format_embeddings_response_oaicompat(body, responses) : json(responses); res_ok(res, root); }; + const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) { + handle_embeddings_impl(req, res, false); + }; + + const auto handle_embeddings_oai = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) { + handle_embeddings_impl(req, res, true); + }; + const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) { if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) { res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED)); @@ -3198,29 +3807,33 @@ int main(int argc, char ** argv) { return; } - // construct prompt object: array of ["query", "doc0", "doc1", ...] - json prompt; - prompt.push_back(query); - for (const auto & doc : documents) { - prompt.push_back(doc); - } - - LOG_DBG("rerank prompt: %s\n", prompt.dump().c_str()); + llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.ctx, query, /* add_special */ false, true)[0]; // create and queue the task json responses = json::array(); bool error = false; { - std::vector tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_RERANK); + std::vector tasks; + std::vector tokenized_docs = tokenize_input_prompts(ctx_server.ctx, documents, /* add_special */ false, true); + tasks.reserve(tokenized_docs.size()); + for (size_t i = 0; i < tokenized_docs.size(); i++) { + server_task task = server_task(SERVER_TASK_TYPE_RERANK); + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + task.prompt_tokens = format_rerank(ctx_server.model, tokenized_query, tokenized_docs[i]); + tasks.push_back(task); + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); // get the result std::unordered_set task_ids = server_task::get_list_id(tasks); - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { - for (const auto & res : results) { - responses.push_back(res.data); + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { + for (auto & res : results) { + GGML_ASSERT(dynamic_cast(res.get()) != nullptr); + responses.push_back(res->to_json()); } }, [&](const json & error_data) { res_error(res, error_data); @@ -3271,36 +3884,50 @@ int main(int argc, char ** argv) { } } - server_task task; - task.type = SERVER_TASK_TYPE_SET_LORA; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + server_task task(SERVER_TASK_TYPE_SET_LORA); + task.id = ctx_server.queue_tasks.get_new_id(); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - res_ok(res, result.data); - res.status = 200; // HTTP OK + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; // // Router // - // register static assets routes - if (!params.public_path.empty()) { - // Set the base directory for serving static files - bool is_found = svr->set_mount_point("/", params.public_path); - if (!is_found) { - LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str()); - return 1; - } + if (!params.webui) { + LOG_INF("Web UI is disabled\n"); } else { - // using embedded static index.html - svr->Get("/", [](const httplib::Request &, httplib::Response & res) { - res.set_content(reinterpret_cast(index_html), index_html_len, "text/html; charset=utf-8"); - return false; - }); + // register static assets routes + if (!params.public_path.empty()) { + // Set the base directory for serving static files + bool is_found = svr->set_mount_point("/", params.public_path); + if (!is_found) { + LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str()); + return 1; + } + } else { + // using embedded static index.html + svr->Get("/", [](const httplib::Request & req, httplib::Response & res) { + if (req.get_header_value("Accept-Encoding").find("gzip") == std::string::npos) { + res.set_content("Error: gzip is not supported by this browser", "text/plain"); + } else { + res.set_header("Content-Encoding", "gzip"); + res.set_content(reinterpret_cast(index_html_gz), index_html_gz_len, "text/html; charset=utf-8"); + } + return false; + }); + } } // register API routes @@ -3318,7 +3945,7 @@ int main(int argc, char ** argv) { svr->Post("/infill", handle_infill); svr->Post("/embedding", handle_embeddings); // legacy svr->Post("/embeddings", handle_embeddings); - svr->Post("/v1/embeddings", handle_embeddings); + svr->Post("/v1/embeddings", handle_embeddings_oai); svr->Post("/rerank", handle_rerank); svr->Post("/reranking", handle_rerank); svr->Post("/v1/rerank", handle_rerank); diff --git a/examples/server/tests/README.md b/examples/server/tests/README.md index 2930a2e0d..fa3d0a2f5 100644 --- a/examples/server/tests/README.md +++ b/examples/server/tests/README.md @@ -44,4 +44,10 @@ To run with stdout/stderr display in real time (verbose output, but useful for d DEBUG=1 ./tests.sh -s -v -x ``` +Hint: You can compile and run test in single command, useful for local developement: + +```shell +cmake --build build -j --target llama-server && ./examples/server/tests/tests.sh +``` + To see all available arguments, please refer to [pytest documentation](https://docs.pytest.org/en/stable/how-to/usage.html) diff --git a/examples/server/tests/tests.sh b/examples/server/tests/tests.sh index 1e285dcda..1e0777de3 100755 --- a/examples/server/tests/tests.sh +++ b/examples/server/tests/tests.sh @@ -1,5 +1,9 @@ #!/bin/bash +# make sure we are in the right directory +SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd ) +cd $SCRIPT_DIR + set -eu if [ $# -lt 1 ] diff --git a/examples/server/tests/unit/test_basic.py b/examples/server/tests/unit/test_basic.py index d82d54a5a..1485de8ce 100644 --- a/examples/server/tests/unit/test_basic.py +++ b/examples/server/tests/unit/test_basic.py @@ -1,4 +1,5 @@ import pytest +import requests from utils import * server = ServerPreset.tinyllama2() @@ -22,7 +23,12 @@ def test_server_props(): server.start() res = server.make_request("GET", "/props") assert res.status_code == 200 + assert ".gguf" in res.body["model_path"] assert res.body["total_slots"] == server.n_slots + default_val = res.body["default_generation_settings"] + assert server.n_ctx is not None and server.n_slots is not None + assert default_val["n_ctx"] == server.n_ctx / server.n_slots + assert default_val["params"]["seed"] == server.seed def test_server_models(): @@ -33,6 +39,31 @@ def test_server_models(): assert len(res.body["data"]) == 1 assert res.body["data"][0]["id"] == server.model_alias + +def test_server_slots(): + global server + + # without slots endpoint enabled, this should return error + server.server_slots = False + server.start() + res = server.make_request("GET", "/slots") + assert res.status_code == 501 # ERROR_TYPE_NOT_SUPPORTED + assert "error" in res.body + server.stop() + + # with slots endpoint enabled, this should return slots info + server.server_slots = True + server.n_slots = 2 + server.start() + res = server.make_request("GET", "/slots") + assert res.status_code == 200 + assert len(res.body) == server.n_slots + assert server.n_ctx is not None and server.n_slots is not None + assert res.body[0]["n_ctx"] == server.n_ctx / server.n_slots + assert "params" in res.body[0] + assert res.body[0]["params"]["seed"] == server.seed + + def test_load_split_model(): global server server.model_hf_repo = "ggml-org/models" @@ -46,3 +77,20 @@ def test_load_split_model(): }) assert res.status_code == 200 assert match_regex("(little|girl)+", res.body["content"]) + + +def test_no_webui(): + global server + # default: webui enabled + server.start() + url = f"http://{server.server_host}:{server.server_port}" + res = requests.get(url) + assert res.status_code == 200 + assert "" in res.text + server.stop() + + # with --no-webui + server.no_webui = True + server.start() + res = requests.get(url) + assert res.status_code == 404 diff --git a/examples/server/tests/unit/test_chat_completion.py b/examples/server/tests/unit/test_chat_completion.py index 8a439f9ef..6573cc17f 100644 --- a/examples/server/tests/unit/test_chat_completion.py +++ b/examples/server/tests/unit/test_chat_completion.py @@ -12,13 +12,13 @@ def create_server(): @pytest.mark.parametrize( - "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated", + "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason", [ - ("llama-2", "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, False), - ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, False), + (None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"), + ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"), ] ) -def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated): +def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason): global server server.start() res = server.make_request("POST", "/chat/completions", data={ @@ -30,29 +30,28 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte ], }) assert res.status_code == 200 + assert "cmpl" in res.body["id"] # make sure the completion id has the expected format + assert res.body["model"] == model if model is not None else server.model_alias assert res.body["usage"]["prompt_tokens"] == n_prompt assert res.body["usage"]["completion_tokens"] == n_predicted choice = res.body["choices"][0] assert "assistant" == choice["message"]["role"] assert match_regex(re_content, choice["message"]["content"]) - if truncated: - assert choice["finish_reason"] == "length" - else: - assert choice["finish_reason"] == "stop" + assert choice["finish_reason"] == finish_reason @pytest.mark.parametrize( - "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated", + "system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason", [ - ("llama-2", "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, False), - ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, False), + ("Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"), + ("You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"), ] ) -def test_chat_completion_stream(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated): +def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason): global server + server.model_alias = None # try using DEFAULT_OAICOMPAT_MODEL server.start() res = server.make_stream_request("POST", "/chat/completions", data={ - "model": model, "max_tokens": max_tokens, "messages": [ {"role": "system", "content": system_prompt}, @@ -61,18 +60,19 @@ def test_chat_completion_stream(model, system_prompt, user_prompt, max_tokens, r "stream": True, }) content = "" + last_cmpl_id = None for data in res: choice = data["choices"][0] + assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future + if last_cmpl_id is None: + last_cmpl_id = data["id"] + assert last_cmpl_id == data["id"] # make sure the completion id is the same for all events in the stream if choice["finish_reason"] in ["stop", "length"]: assert data["usage"]["prompt_tokens"] == n_prompt assert data["usage"]["completion_tokens"] == n_predicted assert "content" not in choice["delta"] assert match_regex(re_content, content) - # FIXME: not sure why this is incorrect in stream mode - # if truncated: - # assert choice["finish_reason"] == "length" - # else: - # assert choice["finish_reason"] == "stop" + assert choice["finish_reason"] == finish_reason else: assert choice["finish_reason"] is None content += choice["delta"]["content"] @@ -93,7 +93,7 @@ def test_chat_completion_with_openai_library(): temperature=0.8, ) print(res) - assert res.choices[0].finish_reason == "stop" + assert res.choices[0].finish_reason == "length" assert res.choices[0].message.content is not None assert match_regex("(Suddenly)+", res.choices[0].message.content) diff --git a/examples/server/tests/unit/test_completion.py b/examples/server/tests/unit/test_completion.py index 2fa30dd03..36aee57dd 100644 --- a/examples/server/tests/unit/test_completion.py +++ b/examples/server/tests/unit/test_completion.py @@ -10,22 +10,29 @@ def create_server(): global server server = ServerPreset.tinyllama2() -@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [ - ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False), - ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False), +@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated,return_tokens", [ + ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False, False), + ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False, True), ]) -def test_completion(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool): +def test_completion(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool, return_tokens: bool): global server server.start() res = server.make_request("POST", "/completion", data={ "n_predict": n_predict, "prompt": prompt, + "return_tokens": return_tokens, }) assert res.status_code == 200 assert res.body["timings"]["prompt_n"] == n_prompt assert res.body["timings"]["predicted_n"] == n_predicted assert res.body["truncated"] == truncated + assert type(res.body["has_new_line"]) == bool assert match_regex(re_content, res.body["content"]) + if return_tokens: + assert len(res.body["tokens"]) > 0 + assert all(type(tok) == int for tok in res.body["tokens"]) + else: + assert res.body["tokens"] == [] @pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [ @@ -42,15 +49,42 @@ def test_completion_stream(prompt: str, n_predict: int, re_content: str, n_promp }) content = "" for data in res: + assert "stop" in data and type(data["stop"]) == bool if data["stop"]: assert data["timings"]["prompt_n"] == n_prompt assert data["timings"]["predicted_n"] == n_predicted assert data["truncated"] == truncated + assert data["stop_type"] == "limit" + assert type(data["has_new_line"]) == bool + assert "generation_settings" in data + assert server.n_predict is not None + assert data["generation_settings"]["n_predict"] == min(n_predict, server.n_predict) + assert data["generation_settings"]["seed"] == server.seed assert match_regex(re_content, content) else: + assert len(data["tokens"]) > 0 + assert all(type(tok) == int for tok in data["tokens"]) content += data["content"] +def test_completion_stream_vs_non_stream(): + global server + server.start() + res_stream = server.make_stream_request("POST", "/completion", data={ + "n_predict": 8, + "prompt": "I believe the meaning of life is", + "stream": True, + }) + res_non_stream = server.make_request("POST", "/completion", data={ + "n_predict": 8, + "prompt": "I believe the meaning of life is", + }) + content_stream = "" + for data in res_stream: + content_stream += data["content"] + assert content_stream == res_non_stream.body["content"] + + @pytest.mark.parametrize("n_slots", [1, 2]) def test_consistent_result_same_seed(n_slots: int): global server @@ -221,3 +255,24 @@ def test_completion_parallel_slots(n_slots: int, n_requests: int): assert len(res.body["content"]) > 10 # FIXME: the result is not deterministic when using other slot than slot 0 # assert match_regex(re_content, res.body["content"]) + + +def test_n_probs(): + global server + server.start() + res = server.make_request("POST", "/completion", data={ + "prompt": "I believe the meaning of life is", + "n_probs": 10, + "temperature": 0.0, + "n_predict": 5, + }) + assert res.status_code == 200 + assert "completion_probabilities" in res.body + assert len(res.body["completion_probabilities"]) == 5 + for tok in res.body["completion_probabilities"]: + assert "probs" in tok + assert len(tok["probs"]) == 10 + for prob in tok["probs"]: + assert "prob" in prob + assert "tok_str" in prob + assert 0.0 <= prob["prob"] <= 1.0 diff --git a/examples/server/tests/unit/test_embedding.py b/examples/server/tests/unit/test_embedding.py index fc7c20064..e32d74582 100644 --- a/examples/server/tests/unit/test_embedding.py +++ b/examples/server/tests/unit/test_embedding.py @@ -14,8 +14,9 @@ def create_server(): def test_embedding_single(): global server + server.pooling = 'last' server.start() - res = server.make_request("POST", "/embeddings", data={ + res = server.make_request("POST", "/v1/embeddings", data={ "input": "I believe the meaning of life is", }) assert res.status_code == 200 @@ -29,8 +30,9 @@ def test_embedding_single(): def test_embedding_multiple(): global server + server.pooling = 'last' server.start() - res = server.make_request("POST", "/embeddings", data={ + res = server.make_request("POST", "/v1/embeddings", data={ "input": [ "I believe the meaning of life is", "Write a joke about AI from a very long prompt which will not be truncated", @@ -45,10 +47,69 @@ def test_embedding_multiple(): assert len(d['embedding']) > 1 -def test_embedding_openai_library_single(): +@pytest.mark.parametrize( + "input,is_multi_prompt", + [ + # single prompt + ("string", False), + ([12, 34, 56], False), + ([12, 34, "string", 56, 78], False), + # multiple prompts + (["string1", "string2"], True), + (["string1", [12, 34, 56]], True), + ([[12, 34, 56], [12, 34, 56]], True), + ([[12, 34, 56], [12, "string", 34, 56]], True), + ] +) +def test_embedding_mixed_input(input, is_multi_prompt: bool): global server server.start() - client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}") + res = server.make_request("POST", "/v1/embeddings", data={"input": input}) + assert res.status_code == 200 + data = res.body['data'] + if is_multi_prompt: + assert len(data) == len(input) + for d in data: + assert 'embedding' in d + assert len(d['embedding']) > 1 + else: + assert 'embedding' in data[0] + assert len(data[0]['embedding']) > 1 + + +def test_embedding_pooling_none(): + global server + server.pooling = 'none' + server.start() + res = server.make_request("POST", "/embeddings", data={ + "input": "hello hello hello", + }) + assert res.status_code == 200 + assert 'embedding' in res.body[0] + assert len(res.body[0]['embedding']) == 5 # 3 text tokens + 2 special + + # make sure embedding vector is not normalized + for x in res.body[0]['embedding']: + assert abs(sum([x ** 2 for x in x]) - 1) > EPSILON + + +def test_embedding_pooling_none_oai(): + global server + server.pooling = 'none' + server.start() + res = server.make_request("POST", "/v1/embeddings", data={ + "input": "hello hello hello", + }) + + # /v1/embeddings does not support pooling type 'none' + assert res.status_code == 400 + + +def test_embedding_openai_library_single(): + global server + server.pooling = 'last' + server.start() + client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1") res = client.embeddings.create(model="text-embedding-3-small", input="I believe the meaning of life is") assert len(res.data) == 1 assert len(res.data[0].embedding) > 1 @@ -56,8 +117,9 @@ def test_embedding_openai_library_single(): def test_embedding_openai_library_multiple(): global server + server.pooling = 'last' server.start() - client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}") + client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1") res = client.embeddings.create(model="text-embedding-3-small", input=[ "I believe the meaning of life is", "Write a joke about AI from a very long prompt which will not be truncated", @@ -71,8 +133,9 @@ def test_embedding_openai_library_multiple(): def test_embedding_error_prompt_too_long(): global server + server.pooling = 'last' server.start() - res = server.make_request("POST", "/embeddings", data={ + res = server.make_request("POST", "/v1/embeddings", data={ "input": "This is a test " * 512, }) assert res.status_code != 200 @@ -80,8 +143,9 @@ def test_embedding_error_prompt_too_long(): def test_same_prompt_give_same_result(): + server.pooling = 'last' server.start() - res = server.make_request("POST", "/embeddings", data={ + res = server.make_request("POST", "/v1/embeddings", data={ "input": [ "I believe the meaning of life is", "I believe the meaning of life is", @@ -97,3 +161,33 @@ def test_same_prompt_give_same_result(): vi = res.body['data'][i]['embedding'] for x, y in zip(v0, vi): assert abs(x - y) < EPSILON + + +@pytest.mark.parametrize( + "content,n_tokens", + [ + ("I believe the meaning of life is", 9), + ("This is a test", 6), + ] +) +def test_embedding_usage_single(content, n_tokens): + global server + server.start() + res = server.make_request("POST", "/v1/embeddings", data={"input": content}) + assert res.status_code == 200 + assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens'] + assert res.body['usage']['prompt_tokens'] == n_tokens + + +def test_embedding_usage_multiple(): + global server + server.start() + res = server.make_request("POST", "/v1/embeddings", data={ + "input": [ + "I believe the meaning of life is", + "I believe the meaning of life is", + ], + }) + assert res.status_code == 200 + assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens'] + assert res.body['usage']['prompt_tokens'] == 2 * 9 diff --git a/examples/server/tests/unit/test_infill.py b/examples/server/tests/unit/test_infill.py index 6a6d40a1c..ad4b8192a 100644 --- a/examples/server/tests/unit/test_infill.py +++ b/examples/server/tests/unit/test_infill.py @@ -13,28 +13,28 @@ def test_infill_without_input_extra(): global server server.start() res = server.make_request("POST", "/infill", data={ - "prompt": "Complete this", - "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_", + "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " int n_threads = llama_", "input_suffix": "}\n", }) assert res.status_code == 200 - assert match_regex("(One|day|she|saw|big|scary|bird)+", res.body["content"]) + assert match_regex("(Ann|small|shiny)+", res.body["content"]) def test_infill_with_input_extra(): global server server.start() res = server.make_request("POST", "/infill", data={ - "prompt": "Complete this", "input_extra": [{ "filename": "llama.h", "text": "LLAMA_API int32_t llama_n_threads();\n" }], - "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_", + "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " int n_threads = llama_", "input_suffix": "}\n", }) assert res.status_code == 200 - assert match_regex("(cuts|Jimmy|mom|came|into|the|room)+", res.body["content"]) + assert match_regex("(Dad|excited|park)+", res.body["content"]) @pytest.mark.parametrize("input_extra", [ @@ -48,10 +48,30 @@ def test_invalid_input_extra_req(input_extra): global server server.start() res = server.make_request("POST", "/infill", data={ - "prompt": "Complete this", "input_extra": [input_extra], - "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_", + "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " int n_threads = llama_", "input_suffix": "}\n", }) assert res.status_code == 400 assert "error" in res.body + + +@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test") +def test_with_qwen_model(): + global server + server.model_file = None + server.model_hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-IQ3_XXS-GGUF" + server.model_hf_file = "qwen2.5-coder-1.5b-iq3_xxs-imat.gguf" + server.start(timeout_seconds=600) + res = server.make_request("POST", "/infill", data={ + "input_extra": [{ + "filename": "llama.h", + "text": "LLAMA_API int32_t llama_n_threads();\n" + }], + "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " int n_threads = llama_", + "input_suffix": "}\n", + }) + assert res.status_code == 200 + assert res.body["content"] == "n_threads();\n printf(\"Number of threads: %d\\n\", n_threads);\n return 0;\n" diff --git a/examples/server/tests/unit/test_rerank.py b/examples/server/tests/unit/test_rerank.py index 189bc4c96..7203d7943 100644 --- a/examples/server/tests/unit/test_rerank.py +++ b/examples/server/tests/unit/test_rerank.py @@ -53,3 +53,26 @@ def test_invalid_rerank_req(documents): }) assert res.status_code == 400 assert "error" in res.body + + +@pytest.mark.parametrize( + "query,doc1,doc2,n_tokens", + [ + ("Machine learning is", "A machine", "Learning is", 19), + ("Which city?", "Machine learning is ", "Paris, capitale de la", 26), + ] +) +def test_rerank_usage(query, doc1, doc2, n_tokens): + global server + server.start() + + res = server.make_request("POST", "/rerank", data={ + "query": query, + "documents": [ + doc1, + doc2, + ] + }) + assert res.status_code == 200 + assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens'] + assert res.body['usage']['prompt_tokens'] == n_tokens diff --git a/examples/server/tests/unit/test_speculative.py b/examples/server/tests/unit/test_speculative.py index 982d6abb4..3bb5733cb 100644 --- a/examples/server/tests/unit/test_speculative.py +++ b/examples/server/tests/unit/test_speculative.py @@ -82,6 +82,37 @@ def test_different_draft_min_draft_max(): last_content = res.body["content"] +def test_slot_ctx_not_exceeded(): + global server + server.n_ctx = 64 + server.start() + res = server.make_request("POST", "/completion", data={ + "prompt": "Hello " * 56, + "temperature": 0.0, + "top_k": 1, + "speculative.p_min": 0.0, + }) + assert res.status_code == 200 + assert len(res.body["content"]) > 0 + + +def test_with_ctx_shift(): + global server + server.n_ctx = 64 + server.start() + res = server.make_request("POST", "/completion", data={ + "prompt": "Hello " * 56, + "temperature": 0.0, + "top_k": 1, + "n_predict": 64, + "speculative.p_min": 0.0, + }) + assert res.status_code == 200 + assert len(res.body["content"]) > 0 + assert res.body["tokens_predicted"] == 64 + assert res.body["truncated"] == True + + @pytest.mark.parametrize("n_slots,n_requests", [ (1, 2), (2, 2), diff --git a/examples/server/tests/utils.py b/examples/server/tests/utils.py index e17a05ff6..277125e88 100644 --- a/examples/server/tests/utils.py +++ b/examples/server/tests/utils.py @@ -64,6 +64,8 @@ class ServerProcess: server_embeddings: bool | None = False server_reranking: bool | None = False server_metrics: bool | None = False + server_slots: bool | None = False + pooling: str | None = None draft: int | None = None api_key: str | None = None response_format: str | None = None @@ -71,6 +73,7 @@ class ServerProcess: disable_ctx_shift: int | None = False draft_min: int | None = None draft_max: int | None = None + no_webui: bool | None = None # session variables process: subprocess.Popen | None = None @@ -91,7 +94,6 @@ class ServerProcess: else: server_path = "../../../build/bin/llama-server" server_args = [ - "--slots", # requires to get slot status via /slots endpoint "--host", self.server_host, "--port", @@ -129,6 +131,10 @@ class ServerProcess: server_args.append("--reranking") if self.server_metrics: server_args.append("--metrics") + if self.server_slots: + server_args.append("--slots") + if self.pooling: + server_args.extend(["--pooling", self.pooling]) if self.model_alias: server_args.extend(["--alias", self.model_alias]) if self.n_ctx: @@ -156,6 +162,8 @@ class ServerProcess: server_args.extend(["--draft-max", self.draft_max]) if self.draft_min: server_args.extend(["--draft-min", self.draft_min]) + if self.no_webui: + server_args.append("--no-webui") args = [str(arg) for arg in [server_path, *server_args]] print(f"bench: starting server with: {' '.join(args)}") @@ -181,7 +189,7 @@ class ServerProcess: start_time = time.time() while time.time() - start_time < timeout_seconds: try: - response = self.make_request("GET", "/slots", headers={ + response = self.make_request("GET", "/health", headers={ "Authorization": f"Bearer {self.api_key}" if self.api_key else None }) if response.status_code == 200: @@ -224,7 +232,7 @@ class ServerProcess: result.headers = dict(response.headers) result.status_code = response.status_code result.body = response.json() if parse_body else None - print("Response from server", result.body) + print("Response from server", json.dumps(result.body, indent=2)) return result def make_stream_request( @@ -245,7 +253,7 @@ class ServerProcess: break elif line.startswith('data: '): data = json.loads(line[6:]) - print("Partial response from server", data) + print("Partial response from server", json.dumps(data, indent=2)) yield data @@ -369,3 +377,6 @@ def match_regex(regex: str, text: str) -> bool: ).search(text) is not None ) + +def is_slow_test_allowed(): + return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON" diff --git a/examples/server/themes/buttons-top/index.html b/examples/server/themes/buttons-top/index.html index 2797c37c9..3fb88fcc8 100644 --- a/examples/server/themes/buttons-top/index.html +++ b/examples/server/themes/buttons-top/index.html @@ -222,7 +222,6 @@ temperature: 0.7, repeat_last_n: 256, // 0 = disable penalty, -1 = context size repeat_penalty: 1.18, // 1.0 = disabled - penalize_nl: false, top_k: 40, // <= 0 to use vocab size top_p: 0.95, // 1.0 = disabled min_p: 0.05, // 0 = disabled @@ -779,7 +778,6 @@ ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })} ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })} ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })} - ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })} ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })} ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })} ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })} diff --git a/examples/server/themes/wild/index.html b/examples/server/themes/wild/index.html index dbe23c402..73f36d4b2 100644 --- a/examples/server/themes/wild/index.html +++ b/examples/server/themes/wild/index.html @@ -225,7 +225,6 @@ temperature: 0.7, repeat_last_n: 256, // 0 = disable penalty, -1 = context size repeat_penalty: 1.18, // 1.0 = disabled - penalize_nl: false, top_k: 40, // <= 0 to use vocab size top_p: 0.95, // 1.0 = disabled min_p: 0.05, // 0 = disabled @@ -782,7 +781,6 @@ ${FloatField({ label: "Temperature", max: 2.0, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature })} ${FloatField({ label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty })} ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })} - ${BoolField({ label: "Penalize repetition of newlines", name: "penalize_nl", value: params.value.penalize_nl })} ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })} ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })} ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })} diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index e4451532c..ffdffe904 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -20,8 +20,9 @@ #include #include #include +#include -#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" +#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo" using json = nlohmann::ordered_json; @@ -40,17 +41,6 @@ using json = nlohmann::ordered_json; #define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) #define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) -// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11 -enum error_type { - ERROR_TYPE_INVALID_REQUEST, - ERROR_TYPE_AUTHENTICATION, - ERROR_TYPE_SERVER, - ERROR_TYPE_NOT_FOUND, - ERROR_TYPE_PERMISSION, - ERROR_TYPE_UNAVAILABLE, // custom error - ERROR_TYPE_NOT_SUPPORTED, // custom error -}; - template static T json_value(const json & body, const std::string & key, const T & default_value) { // Fallback null to default value @@ -148,6 +138,7 @@ static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_ * and multiple prompts (multi-tasks): * - "prompt": ["string1", "string2"] * - "prompt": ["string1", [12, 34, 56]] + * - "prompt": [[12, 34, 56], [78, 90, 12]] * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]] */ static std::vector tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) { @@ -174,6 +165,9 @@ static std::vector tokenize_input_prompts(llama_context * ctx, con } else { throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts"); } + if (result.empty()) { + throw std::runtime_error("\"prompt\" must not be empty"); + } return result; } @@ -337,12 +331,12 @@ static std::string llama_get_chat_template(const struct llama_model * model) { std::string template_key = "tokenizer.chat_template"; // call with NULL buffer to get the total size of the string int32_t res = llama_model_meta_val_str(model, template_key.c_str(), NULL, 0); - if (res < 0) { + if (res < 2) { return ""; } else { - std::vector model_template(res, 0); + std::vector model_template(res + 1, 0); llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size()); - return std::string(model_template.data(), model_template.size()); + return std::string(model_template.data(), model_template.size() - 1); } } @@ -485,48 +479,11 @@ static std::string tokens_to_output_formatted_string(const llama_context * ctx, return out; } -struct completion_token_output { - llama_token tok; - std::string text_to_send; - - struct token_prob { - llama_token tok; - float prob; - }; - - std::vector probs; -}; - -// convert a vector of completion_token_output to json -static json probs_vector_to_json(const llama_context * ctx, const std::vector & probs) { - json out = json::array(); - - for (const auto & prob : probs) { - json probs_for_token = json::array(); - - for (const auto & p : prob.probs) { - const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok); - probs_for_token.push_back(json { - {"tok_str", tok_str}, - {"prob", p.prob}, - }); - } - - const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok); - out.push_back(json { - {"content", tok_str}, - {"probs", probs_for_token}, - }); - } - - return out; -} - static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) { const std::string str = std::string(event) + ": " + data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; // note: these newlines are important (not sure why though, if you know, add a comment to explain) + "\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row). LOG_DBG("data stream, to_send: %s", str.c_str()); @@ -543,8 +500,6 @@ static json oaicompat_completion_params_parse( const std::string & chat_template) { json llama_params; - llama_params["__oaicompat"] = true; - // Apply chat template to the list of messages llama_params["prompt"] = format_chat(model, chat_template, body.at("messages")); @@ -604,166 +559,9 @@ static json oaicompat_completion_params_parse( return llama_params; } -static json format_final_response_oaicompat(const json & request, const json & result, const std::string & completion_id, bool streaming = false, bool verbose = false) { - bool stopped_word = result.count("stopped_word") != 0; - bool stopped_eos = json_value(result, "stopped_eos", false); - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason = "length"; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - - json choices = - streaming ? json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}) - : json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"message", json{{"content", content}, - {"role", "assistant"}}}}}); - - std::time_t t = std::time(0); - - json res = json { - {"choices", choices}, - {"created", t}, - {"model", - json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, - {"usage", json { - {"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens} - }}, - {"id", completion_id} - }; - - // extra fields for debugging purposes - if (verbose) { - res["__verbose"] = result; - } - - if (result.contains("completion_probabilities")) { - res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); - } - - if (result.contains("timings")) { - res.push_back({"timings", json_value(result, "timings", json::object())}); - } - - return res; -} - -// return value is vector as there is one case where we might need to generate two responses -static std::vector format_partial_response_oaicompat(const json & result, const std::string & completion_id) { - if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { - return std::vector({result}); - } - - bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; - std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - - bool stopped_word = json_value(result, "stopped_word", false); - bool stopped_eos = json_value(result, "stopped_eos", false); - bool stopped_limit = json_value(result, "stopped_limit", false); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - if (stopped_limit) { - finish_reason = "length"; - } - - std::time_t t = std::time(0); - - json choices; - - if (!finish_reason.empty()) { - choices = json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}); - } else { - if (first) { - if (content.empty()) { - choices = json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{{"role", "assistant"}}}}}); - } else { - // We have to send this as two updates to conform to openai behavior - json initial_ret = json{{"choices", json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"role", "assistant"} - }}}})}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - json second_ret = json{ - {"choices", json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"content", content}}} - }})}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({initial_ret, second_ret}); - } - } else { - // Some idiosyncrasy in task processing logic makes several trailing calls - // with empty content, we ignore these at the calee site. - if (content.empty()) { - return std::vector({json::object()}); - } - - choices = json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", - json{ - {"content", content}, - }}, - }}); - } - } - - json ret = json { - {"choices", choices}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"} - }; - - if (result.contains("timings")) { - ret.push_back({"timings", json_value(result, "timings", json::object())}); - } - - if (!finish_reason.empty()) { - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - ret.push_back({"usage", json { - {"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens} - }}); - } - - return std::vector({ret}); -} - static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) { json data = json::array(); + int32_t n_tokens = 0; int i = 0; for (const auto & elem : embeddings) { data.push_back(json{ @@ -771,14 +569,16 @@ static json format_embeddings_response_oaicompat(const json & request, const jso {"index", i++}, {"object", "embedding"} }); + + n_tokens += json_value(elem, "tokens_evaluated", 0); } json res = json { {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, {"object", "list"}, - {"usage", json { // TODO: fill - {"prompt_tokens", 0}, - {"total_tokens", 0} + {"usage", json { + {"prompt_tokens", n_tokens}, + {"total_tokens", n_tokens} }}, {"data", data} }; @@ -788,20 +588,23 @@ static json format_embeddings_response_oaicompat(const json & request, const jso static json format_response_rerank(const json & request, const json & ranks) { json data = json::array(); + int32_t n_tokens = 0; int i = 0; for (const auto & rank : ranks) { data.push_back(json{ {"index", i++}, {"relevance_score", json_value(rank, "score", 0.0)}, }); + + n_tokens += json_value(rank, "tokens_evaluated", 0); } json res = json { {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, {"object", "list"}, - {"usage", json { // TODO: fill - {"prompt_tokens", 0}, - {"total_tokens", 0} + {"usage", json { + {"prompt_tokens", n_tokens}, + {"total_tokens", n_tokens} }}, {"results", data} }; @@ -854,42 +657,17 @@ static json format_detokenized_response(const std::string & content) { }; } -static json format_error_response(const std::string & message, const enum error_type type) { - std::string type_str; - int code = 500; - switch (type) { - case ERROR_TYPE_INVALID_REQUEST: - type_str = "invalid_request_error"; - code = 400; - break; - case ERROR_TYPE_AUTHENTICATION: - type_str = "authentication_error"; - code = 401; - break; - case ERROR_TYPE_NOT_FOUND: - type_str = "not_found_error"; - code = 404; - break; - case ERROR_TYPE_SERVER: - type_str = "server_error"; - code = 500; - break; - case ERROR_TYPE_PERMISSION: - type_str = "permission_error"; - code = 403; - break; - case ERROR_TYPE_NOT_SUPPORTED: - type_str = "not_supported_error"; - code = 501; - break; - case ERROR_TYPE_UNAVAILABLE: - type_str = "unavailable_error"; - code = 503; - break; +static json format_logit_bias(const std::vector & logit_bias) { + json data = json::array(); + for (const auto & lb : logit_bias) { + data.push_back(json{ + {"bias", lb.bias}, + {"token", lb.token}, + }); } - return json { - {"code", code}, - {"message", message}, - {"type", type_str}, - }; + return data; +} + +static std::string safe_json_to_str(json data) { + return data.dump(-1, ' ', false, json::error_handler_t::replace); } diff --git a/examples/server/webui/index.html b/examples/server/webui/index.html index c7e18b45e..dcdd41079 100644 --- a/examples/server/webui/index.html +++ b/examples/server/webui/index.html @@ -15,7 +15,7 @@
-
+

Conversations

@@ -120,51 +120,25 @@ {{ messages.length === 0 ? 'Send a message to start' : '' }}
-
-
- - - - -
-
- - -
- - - - - -
+
-
-
- - -
+
+
@@ -227,6 +201,14 @@
Advanced config
+
+ + +
+
+ + Show tokens per second +
+ + + + +