vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into testing/add-lora-tests-workflow

Browse source
This commit is contained in:
ltoniazzi 2024-12-03 20:00:05 +00:00 • committed by GitHub
commit 8743d1e45f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
593 changed files with 105642 additions and 80444 deletions
Download patch file Download diff file Expand all files Collapse all files

161
.clang-format Normal file
View file

@ -0,0 +1,161 @@
---
Language: Cpp
AlignAfterOpenBracket: Align
AlignArrayOfStructures: Left
AlignConsecutiveAssignments: AcrossComments
AlignConsecutiveBitFields: AcrossComments
AlignConsecutiveDeclarations: AcrossComments
AlignConsecutiveMacros: AcrossComments
# AlignConsecutiveShortCaseStatements: AcrossComments
AlignEscapedNewlines: Left # LeftWithLastLine
AlignOperands: Align
AlignTrailingComments:
Kind: Always
OverEmptyLines: 1
AllowAllArgumentsOnNextLine: true
AllowAllParametersOfDeclarationOnNextLine: false
# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen
AllowShortBlocksOnASingleLine: Never
AllowShortCaseLabelsOnASingleLine: false
AllowShortFunctionsOnASingleLine: Inline
AllowShortIfStatementsOnASingleLine: Never
AllowShortLambdasOnASingleLine: Inline
AllowShortLoopsOnASingleLine: false
AlwaysBreakBeforeMultilineStrings: true
BinPackArguments: true
BinPackParameters: true # OnePerLine
BitFieldColonSpacing: Both
BreakBeforeBraces: Custom # Attach
BraceWrapping:
AfterCaseLabel: true
AfterClass: false
AfterControlStatement: false
AfterEnum: false
AfterFunction: false
AfterNamespace: false
AfterObjCDeclaration: false
AfterStruct: false
AfterUnion: false
AfterExternBlock: false
BeforeCatch: false
BeforeElse: false
BeforeLambdaBody: false
BeforeWhile: false
IndentBraces: false
SplitEmptyFunction: false
SplitEmptyRecord: false
SplitEmptyNamespace: false
# BreakAdjacentStringLiterals: true
BreakAfterAttributes: Never
BreakBeforeBinaryOperators: None
BreakBeforeInlineASMColon: OnlyMultiline
BreakBeforeTernaryOperators: false
# BreakBinaryOperations: Never
BreakConstructorInitializers: AfterColon
# BreakFunctionDefinitionParameters: false
BreakInheritanceList: AfterComma
BreakStringLiterals: true
# BreakTemplateDeclarations: Yes
ColumnLimit: 120
CommentPragmas: '^ IWYU pragma:'
CompactNamespaces: false
ConstructorInitializerIndentWidth: 4
ContinuationIndentWidth: 4
Cpp11BracedListStyle: false
DerivePointerAlignment: false
DisableFormat: false
EmptyLineBeforeAccessModifier: Leave
EmptyLineAfterAccessModifier: Never
ExperimentalAutoDetectBinPacking: false
FixNamespaceComments: true
IncludeBlocks: Regroup
IncludeCategories:
- Regex: '^<.*\.h>'
Priority: 1
SortPriority: 0
- Regex: '^<.*'
Priority: 2
SortPriority: 0
- Regex: '.*'
Priority: 3
SortPriority: 0
IncludeIsMainRegex: '([-_](test|unittest))?$'
IncludeIsMainSourceRegex: ''
IndentAccessModifiers: false
IndentCaseBlocks: true
IndentCaseLabels: true
IndentExternBlock: NoIndent
IndentGotoLabels: false
IndentPPDirectives: AfterHash
IndentWidth: 4
IndentWrappedFunctionNames: false
InsertBraces: true # NOTE: may lead to incorrect formatting
InsertNewlineAtEOF: true
JavaScriptQuotes: Leave
JavaScriptWrapImports: true
KeepEmptyLinesAtTheStartOfBlocks: false
LambdaBodyIndentation: Signature
LineEnding: LF
MacroBlockBegin: ''
MacroBlockEnd: ''
MaxEmptyLinesToKeep: 1
NamespaceIndentation: None
ObjCBinPackProtocolList: Auto
ObjCBlockIndentWidth: 4
ObjCSpaceAfterProperty: true
ObjCSpaceBeforeProtocolList: true
PPIndentWidth: -1
PackConstructorInitializers: CurrentLine
PenaltyBreakAssignment: 2
PenaltyBreakBeforeFirstCallParameter: 1
PenaltyBreakComment: 300
PenaltyBreakFirstLessLess: 120
PenaltyBreakString: 1000
PenaltyBreakTemplateDeclaration: 10
PenaltyExcessCharacter: 1000000
PenaltyReturnTypeOnItsOwnLine: 200
PointerAlignment: Middle
QualifierAlignment: Left
#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict']
RawStringFormats:
- Language: Cpp
Delimiters:
- cc
- CC
- cpp
- Cpp
- CPP
- 'c++'
- 'C++'
CanonicalDelimiter: ''
ReferenceAlignment: Middle
ReflowComments: false # IndentOnly
SeparateDefinitionBlocks: Always
SortIncludes: CaseInsensitive
SortUsingDeclarations: LexicographicNumeric
SpaceAfterCStyleCast: true
SpaceAfterLogicalNot: false
SpaceAfterTemplateKeyword: true
SpaceBeforeAssignmentOperators: true
SpaceBeforeCpp11BracedList: false
SpaceBeforeCtorInitializerColon: true
SpaceBeforeInheritanceColon: true
SpaceBeforeParens: ControlStatements
SpaceBeforeRangeBasedForLoopColon: true
SpaceInEmptyBlock: false
SpaceInEmptyParentheses: false
SpacesBeforeTrailingComments: 2
SpacesInAngles: Never
SpacesInContainerLiterals: true
SpacesInLineCommentPrefix:
Minimum: 1
Maximum: -1
SpacesInParentheses: false
SpacesInSquareBrackets: false
SpaceBeforeSquareBrackets: false
Standard: c++17
TabWidth: 4
UseTab: Never
WhitespaceSensitiveMacros: ['STRINGIZE']
...

2
.clang-tidy
View file

@ -17,8 +17,10 @@ Checks: >
-clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
performance-*,
portability-*,
-portability-simd-intrinsics,
misc-*,
-misc-const-correctness,
-misc-non-private-member-variables-in-classes,
-misc-no-recursion,
-misc-use-anonymous-namespace,
FormatStyle: none

25
.devops/full-cuda.Dockerfile
View file

@ -1,18 +1,16 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG CUDA_VERSION=11.7.1
ARG CUDA_VERSION=12.6.0
# Target the CUDA build image
ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
# CUDA architecture to build for (defaults to all supported archs)
ARG CUDA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
COPY requirements.txt requirements.txt
COPY requirements requirements
@ -24,13 +22,12 @@ WORKDIR /app
COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV GGML_CUDA=1
# Enable cURL
ENV LLAMA_CURL=1
RUN make -j$(nproc)
# Use the default CUDA archs if not specified
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc) && \
cp build/bin/* .
ENTRYPOINT ["/app/.devops/tools.sh"]

33
.devops/full-musa.Dockerfile Normal file
View file

@ -0,0 +1,33 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc3.1.0
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
FROM ${BASE_MUSA_DEV_CONTAINER} AS build
# MUSA architecture to build for (defaults to all supported archs)
ARG MUSA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
COPY requirements.txt requirements.txt
COPY requirements requirements
RUN pip install --upgrade pip setuptools wheel \
&& pip install -r requirements.txt
WORKDIR /app
COPY . .
# Use the default MUSA archs if not specified
RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release -j$(nproc) && \
cp build/bin/* .
ENTRYPOINT ["/app/.devops/tools.sh"]

6
.devops/full-rocm.Dockerfile
View file

@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
# This is mostly tied to rocBLAS supported archs.
ARG ROCM_DOCKER_ARCH=\
ARG ROCM_DOCKER_ARCH="\
gfx803 \
gfx900 \
gfx906 \
@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
gfx1030 \
gfx1100 \
gfx1101 \
gfx1102
gfx1102"
COPY requirements.txt requirements.txt
COPY requirements requirements
@ -34,7 +34,7 @@ WORKDIR /app
COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang

6
.devops/llama-cli-cann.Dockerfile
View file

@ -1,6 +1,6 @@
ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
FROM cosdt/cann:$ASCEND_VERSION AS build
FROM ascendai/cann:$ASCEND_VERSION AS build
WORKDIR /app
@ -22,11 +22,11 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
RUN echo "Building with static libs" && \
source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \
cmake --build build --config Release --target llama-cli
# TODO: use image with NNRT
FROM cosdt/cann:$ASCEND_VERSION AS runtime
FROM ascendai/cann:$ASCEND_VERSION AS runtime
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENV LC_ALL=C.utf8

25
.devops/llama-cli-cuda.Dockerfile
View file

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG CUDA_VERSION=11.7.1
ARG CUDA_VERSION=12.6.0
# Target the CUDA build image
ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
# Target the CUDA runtime image
@ -8,28 +8,31 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
# CUDA architecture to build for (defaults to all supported archs)
ARG CUDA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential git
apt-get install -y build-essential git cmake
WORKDIR /app
COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV GGML_CUDA=1
RUN make -j$(nproc) llama-cli
# Use the default CUDA archs if not specified
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/llama-cli /llama-cli
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /
ENTRYPOINT [ "/llama-cli" ]

4
.devops/llama-cli-intel.Dockerfile
View file

@ -1,4 +1,4 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with static libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
cmake --build build --config Release --target llama-cli

38
.devops/llama-cli-musa.Dockerfile Normal file
View file

@ -0,0 +1,38 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc3.1.0
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
# Target the MUSA runtime image
ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
FROM ${BASE_MUSA_DEV_CONTAINER} AS build
# MUSA architecture to build for (defaults to all supported archs)
ARG MUSA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential git cmake
WORKDIR /app
COPY . .
# Use the default MUSA archs if not specified
RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENTRYPOINT [ "/llama-cli" ]

6
.devops/llama-cli-rocm.Dockerfile
View file

@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
# This is mostly tied to rocBLAS supported archs.
ARG ROCM_DOCKER_ARCH=\
ARG ROCM_DOCKER_ARCH="\
gfx803 \
gfx900 \
gfx906 \
@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
gfx1030 \
gfx1100 \
gfx1101 \
gfx1102
gfx1102"
COPY requirements.txt requirements.txt
COPY requirements requirements
@ -34,7 +34,7 @@ WORKDIR /app
COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang

2
.devops/llama-cli-vulkan.Dockerfile
View file

@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_VULKAN=1 && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 && \
cmake --build build --config Release --target llama-cli
# Clean up

30
.devops/llama-server-cuda.Dockerfile
View file

@ -1,6 +1,6 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG CUDA_VERSION=11.7.1
ARG CUDA_VERSION=12.6.0
# Target the CUDA build image
ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
# Target the CUDA runtime image
@ -8,31 +8,35 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
FROM ${BASE_CUDA_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
ARG CUDA_DOCKER_ARCH=all
# CUDA architecture to build for (defaults to all supported archs)
ARG CUDA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential git libcurl4-openssl-dev
apt-get install -y build-essential git cmake libcurl4-openssl-dev
WORKDIR /app
COPY . .
# Set nvcc architecture
ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
# Enable CUDA
ENV GGML_CUDA=1
# Enable cURL
ENV LLAMA_CURL=1
RUN make -j$(nproc) llama-server
# Use the default CUDA archs if not specified
RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/llama-server /llama-server
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]

6
.devops/llama-server-intel.Dockerfile
View file

@ -1,4 +1,4 @@
ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
fi && \
echo "Building with dynamic libs" && \
cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
cmake --build build --config Release --target llama-server
FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
@ -26,6 +26,8 @@ RUN apt-get update && \
COPY --from=build /app/build/bin/llama-server /llama-server
ENV LC_ALL=C.utf8
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]

43
.devops/llama-server-musa.Dockerfile Normal file
View file

@ -0,0 +1,43 @@
ARG UBUNTU_VERSION=22.04
# This needs to generally match the container host's environment.
ARG MUSA_VERSION=rc3.1.0
# Target the MUSA build image
ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
# Target the MUSA runtime image
ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
FROM ${BASE_MUSA_DEV_CONTAINER} AS build
# MUSA architecture to build for (defaults to all supported archs)
ARG MUSA_DOCKER_ARCH=default
RUN apt-get update && \
apt-get install -y build-essential git cmake libcurl4-openssl-dev
WORKDIR /app
COPY . .
# Use the default MUSA archs if not specified
RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
ENTRYPOINT [ "/llama-server" ]

8
.devops/llama-server-rocm.Dockerfile
View file

@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
# Unless otherwise specified, we make a fat build.
# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
# This is mostly tied to rocBLAS supported archs.
ARG ROCM_DOCKER_ARCH=\
ARG ROCM_DOCKER_ARCH="\
gfx803 \
gfx900 \
gfx906 \
@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
gfx1030 \
gfx1100 \
gfx1101 \
gfx1102
gfx1102"
COPY requirements.txt requirements.txt
COPY requirements requirements
@ -34,11 +34,13 @@ WORKDIR /app
COPY . .
# Set nvcc architecture
ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
# Enable ROCm
ENV GGML_HIPBLAS=1
ENV CC=/opt/rocm/llvm/bin/clang
ENV CXX=/opt/rocm/llvm/bin/clang++
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
# Enable cURL
ENV LLAMA_CURL=1

4
.devops/llama-server-vulkan.Dockerfile
View file

@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
# Build it
WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
cmake --build build --config Release --target llama-server
# Clean up
@ -23,6 +23,8 @@ RUN cp /app/build/bin/llama-server /llama-server && \
rm -rf /app
ENV LC_ALL=C.utf8
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]

22
.devops/llama-server.Dockerfile
View file

@ -3,24 +3,38 @@ ARG UBUNTU_VERSION=22.04
FROM ubuntu:$UBUNTU_VERSION AS build
RUN apt-get update && \
apt-get install -y build-essential git libcurl4-openssl-dev
apt-get install -y build-essential git cmake libcurl4-openssl-dev
WORKDIR /app
COPY . .
ENV LLAMA_CURL=1
RUN make -j$(nproc) llama-server
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 $(nproc) && \
# Copy the built libraries to /app/lib
mkdir -p /app/lib && \
mv libggml-cpu* /app/lib/ && \
find build -name "*.so" -exec cp {} /app/lib/ \;
FROM ubuntu:$UBUNTU_VERSION AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/llama-server /llama-server
COPY --from=build /app/build/bin/llama-server /llama-server
COPY --from=build /app/lib/ /
ENV LC_ALL=C.utf8
# Must be set to 0.0.0.0 so it can listen to requests from host machine
ENV LLAMA_ARG_HOST=0.0.0.0
HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]

53
.devops/nix/devshells.nix
View file

@ -1,13 +1,52 @@
{ inputs, ... }:
{
perSystem =
{ config, lib, ... }:
{
config,
lib,
system,
...
}:
{
devShells =
lib.concatMapAttrs
(name: package: {
${name} = package.passthru.shell;
${name + "-extra"} = package.passthru.shell-extra;
})
config.packages;
let
pkgs = import inputs.nixpkgs { inherit system; };
stdenv = pkgs.stdenv;
scripts = config.packages.python-scripts;
in
lib.pipe (config.packages) [
(lib.concatMapAttrs (
name: package: {
${name} = pkgs.mkShell {
name = "${name}";
inputsFrom = [ package ];
shellHook = ''
echo "Entering ${name} devShell"
'';
};
"${name}-extra" =
if (name == "python-scripts") then
null
else
pkgs.mkShell {
name = "${name}-extra";
inputsFrom = [
package
scripts
];
# Extra packages that *may* be used by some scripts
packages = [
pkgs.python3Packages.tiktoken
];
shellHook = ''
echo "Entering ${name} devShell"
addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib stdenv.cc.cc}/lib"
'';
};
}
))
(lib.filterAttrs (name: value: value != null))
];
};
}

18
.devops/nix/nixpkgs-instances.nix
View file

@ -26,16 +26,14 @@
config.cudaSupport = true;
config.allowUnfreePredicate =
p:
builtins.all
(
license:
license.free
|| builtins.elem license.shortName [
"CUDA EULA"
"cuDNN EULA"
]
)
(p.meta.licenses or [ p.meta.license ]);
builtins.all (
license:
license.free
|| builtins.elem license.shortName [
"CUDA EULA"
"cuDNN EULA"
]
) (p.meta.licenses or [ p.meta.license ]);
};
# Ensure dependencies use ROCm consistently
pkgsRocm = import inputs.nixpkgs {

36
.devops/nix/package-gguf-py.nix Normal file
View file

@ -0,0 +1,36 @@
{
lib,
llamaVersion,
numpy,
tqdm,
sentencepiece,
pyyaml,
poetry-core,
buildPythonPackage,
pytestCheckHook,
}:
buildPythonPackage {
pname = "gguf";
version = llamaVersion;
pyproject = true;
nativeBuildInputs = [ poetry-core ];
propagatedBuildInputs = [
numpy
tqdm
sentencepiece
pyyaml
];
src = lib.cleanSource ../../gguf-py;
pythonImportsCheck = [
"numpy"
"gguf"
];
nativeCheckInputs = [ pytestCheckHook ];
doCheck = true;
meta = with lib; {
description = "Python package for writing binary files in the GGUF format";
license = licenses.mit;
maintainers = [ maintainers.ditsuke ];
};
}

368
.devops/nix/package.nix
View file

@ -3,31 +3,33 @@
glibc,
config,
stdenv,
mkShell,
runCommand,
cmake,
ninja,
pkg-config,
git,
python3,
mpi,
blas,
cudaPackages,
autoAddDriverRunpath,
darwin,
rocmPackages,
vulkan-headers,
vulkan-loader,
curl,
shaderc,
useBlas ? builtins.all (x: !x) [
useCuda
useMetalKit
useRocm
useVulkan
] && blas.meta.available,
useBlas ?
builtins.all (x: !x) [
useCuda
useMetalKit
useRocm
useVulkan
]
&& blas.meta.available,
useCuda ? config.cudaSupport,
useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
useMpi ? false, # Increases the runtime closure size by ~700M
# Increases the runtime closure size by ~700M
useMpi ? false,
useRocm ? config.rocmSupport,
enableCurl ? true,
useVulkan ? false,
@ -37,8 +39,8 @@
# otherwise we get libstdc++ errors downstream.
effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
enableStatic ? effectiveStdenv.hostPlatform.isStatic,
precompileMetalShaders ? false
}@inputs:
precompileMetalShaders ? false,
}:
let
inherit (lib)
@ -46,7 +48,6 @@ let
cmakeFeature
optionals
strings
versionOlder
;
stdenv = throw "Use effectiveStdenv instead";
@ -62,54 +63,11 @@ let
pnameSuffix =
strings.optionalString (suffices != [ ])
"-${strings.concatMapStringsSep "-" strings.toLower suffices}";
descriptionSuffix =
strings.optionalString (suffices != [ ])
", accelerated with ${strings.concatStringsSep ", " suffices}";
descriptionSuffix = strings.optionalString (
suffices != [ ]
) ", accelerated with ${strings.concatStringsSep ", " suffices}";
executableSuffix = effectiveStdenv.hostPlatform.extensions.executable;
# TODO: package the Python in this repository in a Nix-like way.
# It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
# is PEP 517-compatible, and ensure the correct .dist-info is generated.
# https://peps.python.org/pep-0517/
#
# TODO: Package up each Python script or service appropriately, by making
# them into "entrypoints"
llama-python = python3.withPackages (
ps: [
ps.numpy
ps.sentencepiece
]
);
# TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
llama-python-extra = python3.withPackages (
ps: [
ps.numpy
ps.sentencepiece
ps.tiktoken
ps.torchWithoutCuda
ps.transformers
# server bench
ps.matplotlib
# server tests
ps.openai
ps.behave
ps.prometheus-client
# for examples/pydantic-models-to-grammar-examples.py
ps.docstring-parser
ps.pydantic
# for scripts/compare-llama-bench.py
ps.gitpython
ps.tabulate
]
);
xcrunHost = runCommand "xcrunHost" {} ''
xcrunHost = runCommand "xcrunHost" { } ''
mkdir -p $out/bin
ln -s /usr/bin/xcrun $out/bin
'';
@ -144,181 +102,145 @@ let
];
in
effectiveStdenv.mkDerivation (
finalAttrs: {
pname = "llama-cpp${pnameSuffix}";
version = llamaVersion;
effectiveStdenv.mkDerivation (finalAttrs: {
pname = "llama-cpp${pnameSuffix}";
version = llamaVersion;
# Note: none of the files discarded here are visible in the sandbox or
# affect the output hash. This also means they can be modified without
# triggering a rebuild.
src = lib.cleanSourceWith {
filter =
name: type:
let
noneOf = builtins.all (x: !x);
baseName = baseNameOf name;
in
noneOf [
(lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
(lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
(lib.hasPrefix "." baseName) # Skip hidden files and directories
(baseName == "flake.lock")
];
src = lib.cleanSource ../../.;
};
postPatch = ''
substituteInPlace ./ggml/src/ggml-metal.m \
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
substituteInPlace ./ggml/src/ggml-metal.m \
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
# With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
# `default.metallib` may be compiled with Metal compiler from XCode
# and we need to escape sandbox on MacOS to access Metal compiler.
# `xcrun` is used find the path of the Metal compiler, which is varible
# and not on $PATH
# see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
__noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
nativeBuildInputs =
[
cmake
ninja
pkg-config
git
]
++ optionals useCuda [
cudaPackages.cuda_nvcc
# TODO: Replace with autoAddDriverRunpath
# once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
cudaPackages.autoAddOpenGLRunpathHook
]
++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
glibc.static
] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
xcrunHost
# Note: none of the files discarded here are visible in the sandbox or
# affect the output hash. This also means they can be modified without
# triggering a rebuild.
src = lib.cleanSourceWith {
filter =
name: type:
let
noneOf = builtins.all (x: !x);
baseName = baseNameOf name;
in
noneOf [
(lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
(lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
(lib.hasPrefix "." baseName) # Skip hidden files and directories
(baseName == "flake.lock")
];
src = lib.cleanSource ../../.;
};
buildInputs =
optionals effectiveStdenv.isDarwin darwinBuildInputs
++ optionals useCuda cudaBuildInputs
++ optionals useMpi [ mpi ]
++ optionals useRocm rocmBuildInputs
++ optionals useBlas [ blas ]
++ optionals useVulkan vulkanBuildInputs
++ optionals enableCurl [ curl ];
postPatch = ''
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
cmakeFlags =
[
(cmakeBool "LLAMA_BUILD_SERVER" true)
(cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
(cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
(cmakeBool "LLAMA_CURL" enableCurl)
(cmakeBool "GGML_NATIVE" false)
(cmakeBool "GGML_BLAS" useBlas)
(cmakeBool "GGML_CUDA" useCuda)
(cmakeBool "GGML_HIPBLAS" useRocm)
(cmakeBool "GGML_METAL" useMetalKit)
(cmakeBool "GGML_VULKAN" useVulkan)
(cmakeBool "GGML_STATIC" enableStatic)
]
++ optionals useCuda [
(
with cudaPackages.flags;
cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
)
# With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
# `default.metallib` may be compiled with Metal compiler from XCode
# and we need to escape sandbox on MacOS to access Metal compiler.
# `xcrun` is used find the path of the Metal compiler, which is varible
# and not on $PATH
# see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
__noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
nativeBuildInputs =
[
cmake
ninja
pkg-config
git
]
++ optionals useCuda [
cudaPackages.cuda_nvcc
autoAddDriverRunpath
]
++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
buildInputs =
optionals effectiveStdenv.isDarwin darwinBuildInputs
++ optionals useCuda cudaBuildInputs
++ optionals useMpi [ mpi ]
++ optionals useRocm rocmBuildInputs
++ optionals useBlas [ blas ]
++ optionals useVulkan vulkanBuildInputs
++ optionals enableCurl [ curl ];
cmakeFlags =
[
(cmakeBool "LLAMA_BUILD_SERVER" true)
(cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
(cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
(cmakeBool "LLAMA_CURL" enableCurl)
(cmakeBool "GGML_NATIVE" false)
(cmakeBool "GGML_BLAS" useBlas)
(cmakeBool "GGML_CUDA" useCuda)
(cmakeBool "GGML_HIP" useRocm)
(cmakeBool "GGML_METAL" useMetalKit)
(cmakeBool "GGML_VULKAN" useVulkan)
(cmakeBool "GGML_STATIC" enableStatic)
]
++ optionals useCuda [
(
with cudaPackages.flags;
cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
)
]
++ optionals useRocm [
(cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
(cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
]
++ optionals useMetalKit [
(lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
(cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
];
)
]
++ optionals useRocm [
(cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
(cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
]
++ optionals useMetalKit [
(lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
(cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
];
# Environment variables needed for ROCm
env = optionals useRocm {
ROCM_PATH = "${rocmPackages.clr}";
HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
};
# Environment variables needed for ROCm
env = optionals useRocm {
ROCM_PATH = "${rocmPackages.clr}";
HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
};
# TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
# if they haven't been added yet.
postInstall = ''
mkdir -p $out/include
cp $src/include/llama.h $out/include/
'';
# TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
# if they haven't been added yet.
postInstall = ''
mkdir -p $out/include
cp $src/include/llama.h $out/include/
'';
# Define the shells here, but don't add in the inputsFrom to avoid recursion.
passthru = {
inherit
useBlas
useCuda
useMetalKit
useMpi
useRocm
useVulkan
;
meta = {
# Configurations we don't want even the CI to evaluate. Results in the
# "unsupported platform" messages. This is mostly a no-op, because
# cudaPackages would've refused to evaluate anyway.
badPlatforms = optionals useCuda lib.platforms.darwin;
shell = mkShell {
name = "shell-${finalAttrs.finalPackage.name}";
description = "contains numpy and sentencepiece";
buildInputs = [ llama-python ];
inputsFrom = [ finalAttrs.finalPackage ];
shellHook = ''
addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib"
'';
};
# Configurations that are known to result in build failures. Can be
# overridden by importing Nixpkgs with `allowBroken = true`.
broken = (useMetalKit && !effectiveStdenv.isDarwin);
shell-extra = mkShell {
name = "shell-extra-${finalAttrs.finalPackage.name}";
description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
buildInputs = [ llama-python-extra ];
inputsFrom = [ finalAttrs.finalPackage ];
};
};
description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
homepage = "https://github.com/ggerganov/llama.cpp/";
license = lib.licenses.mit;
meta = {
# Configurations we don't want even the CI to evaluate. Results in the
# "unsupported platform" messages. This is mostly a no-op, because
# cudaPackages would've refused to evaluate anyway.
badPlatforms = optionals useCuda lib.platforms.darwin;
# Accommodates `nix run` and `lib.getExe`
mainProgram = "llama-cli";
# Configurations that are known to result in build failures. Can be
# overridden by importing Nixpkgs with `allowBroken = true`.
broken = (useMetalKit && !effectiveStdenv.isDarwin);
# These people might respond, on the best effort basis, if you ping them
# in case of Nix-specific regressions or for reviewing Nix-specific PRs.
# Consider adding yourself to this list if you want to ensure this flake
# stays maintained and you're willing to invest your time. Do not add
# other people without their consent. Consider removing people after
# they've been unreachable for long periods of time.
description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
homepage = "https://github.com/ggerganov/llama.cpp/";
license = lib.licenses.mit;
# Note that lib.maintainers is defined in Nixpkgs, but you may just add
# an attrset following the same format as in
# https://github.com/NixOS/nixpkgs/blob/f36a80e54da29775c78d7eff0e628c2b4e34d1d7/maintainers/maintainer-list.nix
maintainers = with lib.maintainers; [
philiptaron
SomeoneSerge
];
# Accommodates `nix run` and `lib.getExe`
mainProgram = "llama-cli";
# These people might respond, on the best effort basis, if you ping them
# in case of Nix-specific regressions or for reviewing Nix-specific PRs.
# Consider adding yourself to this list if you want to ensure this flake
# stays maintained and you're willing to invest your time. Do not add
# other people without their consent. Consider removing people after
# they've been unreachable for long periods of time.
# Note that lib.maintainers is defined in Nixpkgs, but you may just add
# an attrset following the same format as in
# https://github.com/NixOS/nixpkgs/blob/f36a80e54da29775c78d7eff0e628c2b4e34d1d7/maintainers/maintainer-list.nix
maintainers = with lib.maintainers; [
philiptaron
SomeoneSerge
];
# Extend `badPlatforms` instead
platforms = lib.platforms.all;
};
}
)
# Extend `badPlatforms` instead
platforms = lib.platforms.all;
};
})

66
.devops/nix/python-scripts.nix Normal file
View file

@ -0,0 +1,66 @@
{
lib,
stdenv,
buildPythonPackage,
poetry-core,
mkShell,
python3Packages,
gguf-py,
}@inputs:
let
llama-python-deps = with python3Packages; [
numpy
sentencepiece
transformers
protobuf
torchWithoutCuda
gguf-py
tqdm
# for scripts/compare-llama-bench.py
gitpython
tabulate
# for examples/pydantic-models-to-grammar-examples.py
docstring-parser
pydantic
];
llama-python-test-deps = with python3Packages; [
# Server bench
matplotlib
# server tests
openai
pytest
prometheus-client
];
in
buildPythonPackage ({
pname = "llama-scripts";
version = "0.0.0";
pyproject = true;
# NOTE: The files filtered out here are not visible in the build sandbox, neither
# do they affect the output hash. They can be modified without triggering a rebuild.
src = lib.cleanSourceWith {
filter =
name: type:
let
any = builtins.any (x: x);
baseName = builtins.baseNameOf name;
in
any [
(lib.hasSuffix ".py" name)
(baseName == "README.md")
(baseName == "pyproject.toml")
];
src = lib.cleanSource ../../.;
};
nativeBuildInputs = [ poetry-core ];
nativeCheckInputs = llama-python-test-deps;
dependencies = llama-python-deps;
})

40
.devops/nix/scope.nix
View file

@ -1,19 +1,41 @@
{
lib,
newScope,
python3,
llamaVersion ? "0.0.0",
}:
let
pythonPackages = python3.pkgs;
buildPythonPackage = pythonPackages.buildPythonPackage;
numpy = pythonPackages.numpy;
tqdm = pythonPackages.tqdm;
sentencepiece = pythonPackages.sentencepiece;
pyyaml = pythonPackages.pyyaml;
poetry-core = pythonPackages.poetry-core;
pytestCheckHook = pythonPackages.pytestCheckHook;
in
# We're using `makeScope` instead of just writing out an attrset
# because it allows users to apply overlays later using `overrideScope'`.
# Cf. https://noogle.dev/f/lib/makeScope
lib.makeScope newScope (
self: {
inherit llamaVersion;
llama-cpp = self.callPackage ./package.nix { };
docker = self.callPackage ./docker.nix { };
docker-min = self.callPackage ./docker.nix { interactive = false; };
sif = self.callPackage ./sif.nix { };
}
)
lib.makeScope newScope (self: {
inherit llamaVersion;
gguf-py = self.callPackage ./package-gguf-py.nix {
inherit
buildPythonPackage
numpy
tqdm
sentencepiece
poetry-core
pyyaml
pytestCheckHook
;
};
python-scripts = self.callPackage ./python-scripts.nix { inherit buildPythonPackage poetry-core; };
llama-cpp = self.callPackage ./package.nix { };
docker = self.callPackage ./docker.nix { };
docker-min = self.callPackage ./docker.nix { interactive = false; };
sif = self.callPackage ./sif.nix { };
})

2
.dockerignore
View file

@ -1,7 +1,7 @@
*.o
*.a
.cache/
.git/
# Do not ignore .git directory, otherwise the reported build number will always be 0
.github/
.gitignore
.vs/

2
.ecrc
View file

@ -1,5 +1,5 @@
{
"Exclude": ["^\\.gitmodules$"],
"Exclude": ["^\\.gitmodules$", "stb_image\\.h"],
"Disable": {
"IndentSize": true
}

10
.editorconfig
View file

@ -24,6 +24,16 @@ insert_final_newline = unset
[examples/server/public/*]
indent_size = 2
[examples/server/public/deps_*]
trim_trailing_whitespace = unset
indent_style = unset
indent_size = unset
[examples/server/deps_*]
trim_trailing_whitespace = unset
indent_style = unset
indent_size = unset
[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
indent_style = tab

50
.github/ISSUE_TEMPLATE/01-bug-low.yml vendored
View file

@ -1,50 +0,0 @@
name: Low Severity Bugs
description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
title: "Bug: "
labels: ["bug-unconfirmed", "low severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

77
.github/ISSUE_TEMPLATE/010-bug-compilation.yml vendored Normal file
View file

@ -0,0 +1,77 @@
name: Bug (compilation)
description: Something goes wrong when trying to compile llama.cpp.
title: "Compile bug: "
labels: ["bug-unconfirmed", "compilation"]
body:
- type: markdown
attributes:
value: >
Thanks for taking the time to fill out this bug report!
This issue template is intended for bug reports where the compilation of llama.cpp fails.
Before opening an issue, please confirm that the compilation still fails with `-DGGML_CCACHE=OFF`.
If the compilation succeeds with ccache disabled you should be able to permanently fix the issue
by clearing `~/.cache/ccache` (on Linux).
- type: textarea
id: commit
attributes:
label: Git commit
description: Which commit are you trying to compile?
placeholder: |
$git rev-parse HEAD
84a07a17b1b08cf2b9747c633a2372782848a27f
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: Operating systems
description: Which operating systems do you know to be affected?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: true
- type: dropdown
id: backends
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
multiple: true
validations:
required: true
- type: textarea
id: info
attributes:
label: Problem description & steps to reproduce
description: >
Please give us a summary of the problem and tell us how to reproduce it.
If you can narrow down the bug to specific compile flags, that information would be very much appreciated by us.
placeholder: >
I'm trying to compile llama.cpp with CUDA support on a fresh install of Ubuntu and get error XY.
Here are the exact commands that I used: ...
validations:
required: true
- type: textarea
id: first_bad_commit
attributes:
label: First Bad Commit
description: >
If the bug was not present on an earlier version: when did it start appearing?
If possible, please do a git bisect and identify the exact commit that introduced the bug.
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: >
Please copy and paste any relevant log output, including the command that you entered and any generated text.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:
required: true

101
.github/ISSUE_TEMPLATE/011-bug-results.yml vendored Normal file
View file

@ -0,0 +1,101 @@
name: Bug (model use)
description: Something goes wrong when using a model (in general, not specific to a single llama.cpp module).
title: "Eval bug: "
labels: ["bug-unconfirmed", "model evaluation"]
body:
- type: markdown
attributes:
value: >
Thanks for taking the time to fill out this bug report!
This issue template is intended for bug reports where the model evaluation results
(i.e. the generated text) are incorrect or llama.cpp crashes during model evaluation.
If you encountered the issue while using an external UI (e.g. ollama),
please reproduce your issue using one of the examples/binaries in this repository.
The `llama-cli` binary can be used for simple and reproducible model inference.
- type: textarea
id: version
attributes:
label: Name and Version
description: Which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: Operating systems
description: Which operating systems do you know to be affected?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: true
- type: dropdown
id: backends
attributes:
label: GGML backends
description: Which GGML backends do you know to be affected?
options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
multiple: true
validations:
required: true
- type: textarea
id: hardware
attributes:
label: Hardware
description: Which CPUs/GPUs are you using?
placeholder: >
e.g. Ryzen 5950X + 2x RTX 4090
validations:
required: true
- type: textarea
id: model
attributes:
label: Models
description: >
Which model(s) at which quantization were you using when encountering the bug?
If you downloaded a GGUF file off of Huggingface, please provide a link.
placeholder: >
e.g. Meta LLaMA 3.1 Instruct 8b q4_K_M
validations:
required: false
- type: textarea
id: info
attributes:
label: Problem description & steps to reproduce
description: >
Please give us a summary of the problem and tell us how to reproduce it.
If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
that information would be very much appreciated by us.
placeholder: >
e.g. when I run llama-cli with -ngl 99 I get garbled outputs.
When I use -ngl 0 it works correctly.
Here are the exact commands that I used: ...
validations:
required: true
- type: textarea
id: first_bad_commit
attributes:
label: First Bad Commit
description: >
If the bug was not present on an earlier version: when did it start appearing?
If possible, please do a git bisect and identify the exact commit that introduced the bug.
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: >
Please copy and paste any relevant log output, including the command that you entered and any generated text.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:
required: true

81
.github/ISSUE_TEMPLATE/019-bug-misc.yml vendored Normal file
View file

@ -0,0 +1,81 @@
name: Bug (misc.)
description: Something is not working the way it should (and it's not covered by any of the above cases).
title: "Misc. bug: "
labels: ["bug-unconfirmed"]
body:
- type: markdown
attributes:
value: >
Thanks for taking the time to fill out this bug report!
This issue template is intended for miscellaneous bugs that don't fit into any other category.
If you encountered the issue while using an external UI (e.g. ollama),
please reproduce your issue using one of the examples/binaries in this repository.
- type: textarea
id: version
attributes:
label: Name and Version
description: Which version of our software is affected? (You can use `--version` to get a version string.)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: Operating systems
description: Which operating systems do you know to be affected?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: dropdown
id: module
attributes:
label: Which llama.cpp modules do you know to be affected?
multiple: true
options:
- Documentation/Github
- libllama (core library)
- llama-cli
- llama-server
- llama-bench
- llama-quantize
- Python/Bash scripts
- Test code
- Other (Please specify in the next section)
validations:
required: false
- type: textarea
id: info
attributes:
label: Problem description & steps to reproduce
description: >
Please give us a summary of the problem and tell us how to reproduce it (if applicable).
validations:
required: true
- type: textarea
id: first_bad_commit
attributes:
label: First Bad Commit
description: >
If the bug was not present on an earlier version and it's not trivial to track down: when did it start appearing?
If possible, please do a git bisect and identify the exact commit that introduced the bug.
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: >
If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
This will be automatically formatted into code, so no need for backticks.
render: shell
validations:
required: false

50
.github/ISSUE_TEMPLATE/02-bug-medium.yml vendored
View file

@ -1,50 +0,0 @@
name: Medium Severity Bug
description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
title: "Bug: "
labels: ["bug-unconfirmed", "medium severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

2
.github/ISSUE_TEMPLATE/05-enhancement.yml → .github/ISSUE_TEMPLATE/020-enhancement.yml vendored
View file

@ -1,5 +1,5 @@
name: Enhancement
description: Used to request enhancements for llama.cpp
description: Used to request enhancements for llama.cpp.
title: "Feature Request: "
labels: ["enhancement"]
body:

50
.github/ISSUE_TEMPLATE/03-bug-high.yml vendored
View file

@ -1,50 +0,0 @@
name: High Severity Bug
description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
title: "Bug: "
labels: ["bug-unconfirmed", "high severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

2
.github/ISSUE_TEMPLATE/06-research.yml → .github/ISSUE_TEMPLATE/030-research.yml vendored
View file

@ -1,5 +1,5 @@
name: Research
description: Track new technical research area
description: Track new technical research area.
title: "Research: "
labels: ["research 🔬"]
body:

50
.github/ISSUE_TEMPLATE/04-bug-critical.yml vendored
View file

@ -1,50 +0,0 @@
name: Critical Severity Bug
description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
title: "Bug: "
labels: ["bug-unconfirmed", "critical severity"]
body:
- type: markdown
attributes:
value: |
Thanks for taking the time to fill out this bug report!
Please include information about your system, the steps to reproduce the bug,
and the version of llama.cpp that you are using.
If possible, please provide a minimal code example that reproduces the bug.
- type: textarea
id: what-happened
attributes:
label: What happened?
description: Also tell us, what did you expect to happen?
placeholder: Tell us what you see!
validations:
required: true
- type: textarea
id: version
attributes:
label: Name and Version
description: Which executable and which version of our software are you running? (use `--version` to get a version string)
placeholder: |
$./llama-cli --version
version: 2999 (42b4109e)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
validations:
required: true
- type: dropdown
id: operating-system
attributes:
label: What operating system are you seeing the problem on?
multiple: true
options:
- Linux
- Mac
- Windows
- BSD
- Other? (Please let us know in description)
validations:
required: false
- type: textarea
id: logs
attributes:
label: Relevant log output
description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
render: shell

2
.github/ISSUE_TEMPLATE/07-refactor.yml → .github/ISSUE_TEMPLATE/040-refactor.yml vendored
View file

@ -1,5 +1,5 @@
name: Refactor (Maintainers)
description: Used to track refactoring opportunities
description: Used to track refactoring opportunities.
title: "Refactor: "
labels: ["refactor"]
body:

15
.github/labeler.yml vendored
View file

@ -3,19 +3,18 @@ Kompute:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-kompute.h
- ggml/src/ggml-kompute.cpp
- ggml/src/ggml-kompute/**
- README-kompute.md
Apple Metal:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-metal.h
- ggml/src/ggml-metal.cpp
- ggml/src/ggml-metal/**
- README-metal.md
SYCL:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml-sycl.h
- ggml/src/ggml-sycl.cpp
- ggml/src/ggml-sycl/**
- docs/backend/SYCL.md
- examples/sycl/**
@ -27,8 +26,8 @@ Nvidia GPU:
Vulkan:
- changed-files:
- any-glob-to-any-file:
- ggml/ggml_vk_generate_shaders.py
- ggml/src/ggml-vulkan*
- ggml/include/ggml-vulkan.h
- ggml/src/ggml-vulkan/**
documentation:
- changed-files:
- any-glob-to-any-file:
@ -75,11 +74,7 @@ server:
ggml:
- changed-files:
- any-glob-to-any-file:
- ggml/include/ggml*.h
- ggml/src/ggml*.c
- ggml/src/ggml*.cpp
- ggml/src/ggml*.h
- ggml-cuda/**
- ggml/**
nix:
- changed-files:
- any-glob-to-any-file:

8
.github/pull_request_template.md vendored
View file

@ -1,7 +1 @@
- [x] I have read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md)
- Self-reported review complexity:
- [ ] Low
- [ ] Medium
- [ ] High
*Make sure to read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*

4
.github/workflows/bench.yml.disabled vendored
View file

@ -27,10 +27,10 @@ on:
push:
branches:
- master
paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
pull_request_target:
types: [opened, synchronize, reopened]
paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
schedule:
- cron: '04 2 * * *'

454
.github/workflows/build.yml vendored
View file

@ -19,10 +19,18 @@ concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
# Fine-grant permission
# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
permissions:
contents: write # for creating release
env:
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
macOS-latest-cmake-arm64:
@ -47,7 +55,13 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
cmake .. \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -84,7 +98,7 @@ jobs:
name: llama-bin-macos-arm64.zip
macOS-latest-cmake-x64:
runs-on: macos-12
runs-on: macos-13
steps:
- name: Clone
@ -105,7 +119,12 @@ jobs:
sysctl -a
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -141,66 +160,6 @@ jobs:
path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
name: llama-bin-macos-x64.zip
ubuntu-focal-make:
runs-on: ubuntu-20.04
env:
LLAMA_NODE_AVAILABLE: true
LLAMA_PYTHON_AVAILABLE: true
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential gcc-8
- uses: actions/setup-node@v4
with:
node-version: "20"
- uses: actions/setup-python@v5
with:
python-version: "3.11"
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
run: |
CC=gcc-8 make -j $(nproc)
- name: Test
id: make_test
run: |
CC=gcc-8 make tests -j $(nproc)
make test -j $(nproc)
ubuntu-focal-make-curl:
runs-on: ubuntu-20.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Dependencies
id: depends
run: |
sudo apt-get update
sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
LLAMA_CURL: 1
run: |
CC=gcc-8 make -j $(nproc)
ubuntu-latest-cmake:
runs-on: ubuntu-latest
@ -375,7 +334,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -386,22 +345,43 @@ jobs:
- name: Build with native CMake HIP support
id: cmake_build
run: |
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
cmake --build build --config Release -j $(nproc)
- name: Build with legacy HIP support
id: cmake_build_legacy_hip
run: |
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIP=ON
cmake --build build2 --config Release -j $(nproc)
ubuntu-22-cmake-musa:
runs-on: ubuntu-22.04
container: mthreads/musa:rc3.1.0-devel-ubuntu22.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Dependencies
id: depends
run: |
apt-get update
apt-get install -y build-essential git cmake libcurl4-openssl-dev
- name: Build with native CMake MUSA support
id: cmake_build
run: |
cmake -B build -S . -DGGML_MUSA=ON
cmake --build build --config Release -j $(nproc)
ubuntu-22-cmake-sycl:
runs-on: ubuntu-22.04
continue-on-error: true
steps:
- uses: actions/checkout@v2
- uses: actions/checkout@v4
- name: add oneAPI to apt
shell: bash
@ -442,7 +422,7 @@ jobs:
continue-on-error: true
steps:
- uses: actions/checkout@v2
- uses: actions/checkout@v4
- name: add oneAPI to apt
shell: bash
@ -477,36 +457,6 @@ jobs:
cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON ..
cmake --build . --config Release -j $(nproc)
# TODO: build with GGML_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
macOS-latest-make:
runs-on: macos-latest
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Dependencies
id: depends
continue-on-error: true
run: |
brew update
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
run: |
GGML_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
- name: Test
id: make_test
run: |
GGML_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
GGML_NO_METAL=1 make test -j $(sysctl -n hw.logicalcpu)
# TODO: build with GGML_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
# how to debug it.
# ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
@ -546,7 +496,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -561,6 +511,7 @@ jobs:
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 \
@ -576,7 +527,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -591,6 +542,7 @@ jobs:
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 \
@ -600,33 +552,35 @@ jobs:
-DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
macOS-latest-swift:
runs-on: macos-latest
strategy:
matrix:
destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
- name: Dependencies
id: depends
continue-on-error: true
run: |
brew update
- name: xcodebuild for swift package
id: xcodebuild
run: |
xcodebuild -scheme llama -destination "${{ matrix.destination }}"
- name: Build Swift Example
id: make_build_swift_example
run: |
make swift
# 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 "${{ matrix.destination }}"
#
# - name: Build Swift Example
# id: make_build_swift_example
# run: |
# make swift
windows-msys2:
runs-on: windows-latest
@ -653,21 +607,6 @@ jobs:
mingw-w64-${{matrix.env}}-cmake
mingw-w64-${{matrix.env}}-openblas
- name: Build using make
shell: msys2 {0}
run: |
make -j $(nproc)
- name: Clean after building using make
shell: msys2 {0}
run: |
make clean
- name: Build using make w/ OpenBLAS
shell: msys2 {0}
run: |
make GGML_OPENBLAS=1 -j $(nproc)
- name: Build using CMake
shell: msys2 {0}
run: |
@ -686,7 +625,7 @@ jobs:
cmake --build build --config ${{ matrix.build }} -j $(nproc)
windows-latest-cmake:
runs-on: windows-2019
runs-on: windows-latest
env:
OPENBLAS_VERSION: 0.3.23
@ -726,7 +665,7 @@ jobs:
id: clone_kompute
if: ${{ matrix.build == 'kompute-x64' }}
run: |
git submodule update --init ggml/src/kompute
git submodule update --init ggml/src/ggml-kompute/kompute
- name: Download OpenBLAS
id: get_openblas
@ -829,12 +768,33 @@ jobs:
path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
name: llama-bin-win-${{ matrix.build }}.zip
windows-latest-cmake-cuda:
ubuntu-latest-cmake-cuda:
runs-on: ubuntu-latest
container: nvidia/cuda:12.6.2-devel-ubuntu24.04
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
- name: Install dependencies
env:
DEBIAN_FRONTEND: noninteractive
run: |
apt update
apt install -y cmake build-essential ninja-build libgomp1 git
- name: Build with CMake
run: |
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
cmake --build build
windows-2019-cmake-cuda:
runs-on: windows-2019
strategy:
matrix:
cuda: ['12.2.0', '11.7.1']
cuda: ['12.4', '11.7']
build: ['cuda']
steps:
@ -842,24 +802,83 @@ jobs:
id: checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
fetch-depth: 0
- name: Install CUDA toolkit
id: cuda-toolkit
uses: Jimver/cuda-toolkit@v0.2.15
- name: Install Cuda Toolkit 11.7
if: ${{ matrix.cuda == '11.7' }}
run: |
mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
choco install unzip -y
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
- name: Install Cuda Toolkit 12.4
if: ${{ matrix.cuda == '12.4' }}
run: |
mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
choco install unzip -y
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
- name: Install ccache
uses: hendrikmuhs/ccache-action@v1.2
with:
cuda: ${{ matrix.cuda }}
method: 'network'
sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }}
- name: Install Ninja
id: install_ninja
run: |
choco install ninja
- name: Build
id: cmake_build
shell: cmd
run: |
mkdir build
cd build
cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
cmake --build build --config Release -j %NINJA_JOBS% -t ggml
cmake --build build --config Release
- name: Determine tag name
id: tag
@ -888,10 +907,12 @@ jobs:
name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
- name: Copy and pack Cuda runtime
if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
run: |
echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
echo "Cuda install location: ${{ env.CUDA_PATH }}"
$dst='.\build\bin\cudart\'
robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
- name: Upload Cuda runtime
@ -909,8 +930,8 @@ jobs:
shell: bash
env:
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
steps:
- name: Clone
@ -920,7 +941,8 @@ jobs:
fetch-depth: 0
- name: Install
run: scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
run: |
scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
- name: Build
id: cmake_build
@ -939,24 +961,33 @@ jobs:
echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
fi
- name: Pack artifacts
- name: Build the release package
id: pack_artifacts
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
run: |
echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
echo "cp oneAPI running time dll files to ./build/bin done"
7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
- name: Upload artifacts
- name: Upload the release package
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
uses: actions/upload-artifact@v4
with:
@ -964,19 +995,63 @@ jobs:
name: llama-bin-win-sycl-x64.zip
windows-latest-cmake-hip:
if: ${{ github.event.inputs.create_release != 'true' }}
runs-on: windows-latest
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
uses: actions/checkout@v4
- name: Install
id: depends
run: |
$ErrorActionPreference = "Stop"
write-host "Downloading AMD HIP SDK Installer"
Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
write-host "Installing AMD HIP SDK"
Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
write-host "Completed AMD HIP SDK installation"
- name: Verify ROCm
id: verify
run: |
& 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
- name: Install ccache
uses: hendrikmuhs/ccache-action@v1.2
with:
key: ${{ github.job }}
- name: Build
id: cmake_build
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
windows-latest-cmake-hip-release:
if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
runs-on: windows-latest
strategy:
matrix:
gpu_target: [gfx1100, gfx1101, gfx1030]
steps:
- name: Clone
id: checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Install
id: depends
run: |
$ErrorActionPreference = "Stop"
write-host "Downloading AMD HIP SDK Installer"
Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
write-host "Installing AMD HIP SDK"
Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
write-host "Completed AMD HIP SDK installation"
@ -991,8 +1066,36 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
cmake --build build --config Release
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
md "build\bin\rocblas\library\"
cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
- name: Determine tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
else
SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
fi
- name: Pack artifacts
id: pack_artifacts
run: |
7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
- name: Upload artifacts
uses: actions/upload-artifact@v4
with:
path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
ios-xcode-build:
runs-on: macos-latest
@ -1051,12 +1154,11 @@ jobs:
runs-on: ubuntu-latest
needs:
- ubuntu-focal-make
- ubuntu-latest-cmake
- macOS-latest-make
- macOS-latest-cmake
- windows-latest-cmake
- windows-latest-cmake-cuda
- windows-2019-cmake-cuda
- windows-latest-cmake-hip-release
- macOS-latest-cmake-arm64
- macOS-latest-cmake-x64

5
.github/workflows/close-issue.yml vendored
View file

@ -3,6 +3,11 @@ on:
schedule:
- cron: "42 0 * * *"
# Fine-grant permission
# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
permissions:
issues: write
jobs:
close-issues:
runs-on: ubuntu-latest

91
.github/workflows/docker.yml vendored
View file

@ -10,20 +10,23 @@
name: Publish Docker image
on:
#pull_request:
push:
branches:
- master
paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
workflow_dispatch: # allows manual triggering
schedule:
# Rebuild daily rather than on every push because it is expensive
- cron: '12 4 * * *'
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
# Fine-grant permission
# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
permissions:
packages: write
jobs:
push_to_registry:
name: Push Docker image to Docker Hub
#if: github.event.pull_request.draft == false
runs-on: ubuntu-latest
env:
@ -37,15 +40,20 @@ jobs:
- { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
# Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
- { tag: "light-musa", dockerfile: ".devops/llama-cli-musa.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-musa", dockerfile: ".devops/llama-server-musa.Dockerfile", platforms: "linux/amd64" }
- { tag: "full-musa", dockerfile: ".devops/full-musa.Dockerfile", platforms: "linux/amd64" }
# Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
#- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
#- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
#- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
- { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
- { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
steps:
- name: Check out the repo
uses: actions/checkout@v4
with:
fetch-depth: 0 # preserve git history, so we can determine the build number
- name: Set up QEMU
uses: docker/setup-qemu-action@v2
@ -60,6 +68,34 @@ jobs:
username: ${{ github.repository_owner }}
password: ${{ secrets.GITHUB_TOKEN }}
- name: Determine tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}" # to lower case
REPO_NAME="${{ github.event.repository.name }}"
# determine tag name postfix (build number, commit hash)
if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
TAG_POSTFIX="b${BUILD_NUMBER}"
else
SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
fi
# list all tags possible
TAGS=""
TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
echo "output_tags=$TAGS" # print out for debugging
env:
GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
# https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
- name: Free Disk Space (Ubuntu)
uses: jlumbroso/free-disk-space@main
@ -77,40 +113,13 @@ jobs:
docker-images: true
swap-storage: true
- name: Determine tag name
id: tag
shell: bash
run: |
BUILD_NUMBER="$(git rev-list --count HEAD)"
SHORT_HASH="$(git rev-parse --short=7 HEAD)"
if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
else
SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
fi
- name: Downcase github.repository_owner
run: |
echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
env:
GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
- name: Build and push Docker image (versioned)
if: github.event_name == 'push'
uses: docker/build-push-action@v4
- name: Build and push Docker image (tagged + versioned)
if: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
uses: docker/build-push-action@v6
with:
context: .
push: true
platforms: ${{ matrix.config.platforms }}
tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
file: ${{ matrix.config.dockerfile }}
- name: Build and push Docker image (tagged)
uses: docker/build-push-action@v4
with:
context: .
push: ${{ github.event_name == 'push' }}
platforms: ${{ matrix.config.platforms }}
tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
# tag list is generated from step above
tags: ${{ steps.tag.outputs.output_tags }}
file: ${{ matrix.config.dockerfile }}

65
.github/workflows/nix-ci-aarch64.yml vendored
View file

@ -1,65 +0,0 @@
name: Nix aarch64 builds
on:
workflow_dispatch: # allows manual triggering
schedule:
# Rebuild daily rather than on every push because QEMU is expensive (e.g.
# 1.5h instead of minutes with the cold cache).
#
# randint(0, 59), randint(0, 23)
- cron: '26 12 * * *'
# But also rebuild if we touched any of the Nix expressions:
push:
branches:
- master
paths: ['**/*.nix', 'flake.lock']
pull_request:
types: [opened, synchronize, reopened]
paths: ['**/*.nix', 'flake.lock']
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
nix-build-aarch64:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install QEMU
# Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
run: |
sudo apt-get update
sudo apt-get install -y qemu-user-static qemu-system-aarch64
sudo usermod -a -G kvm $USER
- name: Install Nix
uses: DeterminateSystems/nix-installer-action@v9
with:
github-token: ${{ secrets.GITHUB_TOKEN }}
extra-conf: |
extra-platforms = aarch64-linux
extra-system-features = nixos-test kvm
extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
- uses: DeterminateSystems/magic-nix-cache-action@v2
with:
upstream-cache: https://${{ matrix.cachixName }}.cachix.org
- name: Set-up cachix to push the results to
uses: cachix/cachix-action@v13
with:
authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
name: llama-cpp
- name: Show all output paths
run: >
nix run github:nix-community/nix-eval-jobs
-- --gc-roots-dir gcroot
--flake
".#packages.aarch64-linux"
- name: Build
run: >
nix run github:Mic92/nix-fast-build
-- --skip-cached --no-nom
--systems aarch64-linux
--flake
".#checks.aarch64-linux"

72
.github/workflows/nix-ci.yml vendored
View file

@ -1,72 +0,0 @@
name: Nix CI
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
pull_request:
types: [opened, synchronize, reopened]
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
cancel-in-progress: true
jobs:
nix-eval:
strategy:
fail-fast: false
matrix:
os: [ ubuntu-latest, macos-latest ]
runs-on: ${{ matrix.os }}
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install Nix
uses: DeterminateSystems/nix-installer-action@v9
with:
github-token: ${{ secrets.GITHUB_TOKEN }}
extra-conf: |
extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
- uses: DeterminateSystems/magic-nix-cache-action@v2
with:
upstream-cache: https://${{ matrix.cachixName }}.cachix.org
- name: List all flake outputs
run: nix flake show --all-systems
- name: Show all output paths
run: >
nix run github:nix-community/nix-eval-jobs
-- --gc-roots-dir gcroot
--flake
".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
nix-build:
strategy:
fail-fast: false
matrix:
os: [ ubuntu-latest, macos-latest ]
runs-on: ${{ matrix.os }}
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install Nix
uses: DeterminateSystems/nix-installer-action@v9
with:
github-token: ${{ secrets.GITHUB_TOKEN }}
extra-conf: |
extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
- uses: DeterminateSystems/magic-nix-cache-action@v2
with:
upstream-cache: https://${{ matrix.cachixName }}.cachix.org
- name: Set-up cachix to push the results to
uses: cachix/cachix-action@v13
with:
authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
name: llama-cpp
- name: Build
run: >
nix run github:Mic92/nix-fast-build
-- --skip-cached --no-nom
--flake
".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"

22
.github/workflows/nix-flake-update.yml vendored
View file

@ -1,22 +0,0 @@
name: update-flake-lock
on:
workflow_dispatch:
schedule:
- cron: '0 0 * * 0' # runs weekly on Sunday at 00:00
jobs:
lockfile:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install Nix
uses: DeterminateSystems/nix-installer-action@main
- name: Update flake.lock
uses: DeterminateSystems/update-flake-lock@main
with:
pr-title: "nix: update flake.lock"
pr-labels: |
nix
pr-reviewers: philiptaron,SomeoneSerge
token: ${{ secrets.FLAKE_TOKEN }}

36
.github/workflows/nix-publish-flake.yml vendored
View file

@ -1,36 +0,0 @@
# Make the flake discoverable on https://flakestry.dev and https://flakehub.com/flakes
name: "Publish a flake to flakestry & flakehub"
on:
push:
tags:
- "*"
workflow_dispatch:
inputs:
tag:
description: "The existing tag to publish"
type: "string"
required: true
jobs:
flakestry-publish:
runs-on: ubuntu-latest
permissions:
id-token: "write"
contents: "read"
steps:
- uses: flakestry/flakestry-publish@main
with:
version: "${{ inputs.tag || github.ref_name }}"
flakehub-publish:
runs-on: "ubuntu-latest"
permissions:
id-token: "write"
contents: "read"
steps:
- uses: "actions/checkout@v4"
with:
ref: "${{ (inputs.tag != null) && format('refs/tags/{0}', inputs.tag) || '' }}"
- uses: "DeterminateSystems/nix-installer-action@main"
- uses: "DeterminateSystems/flakehub-push@main"
with:
visibility: "public"
tag: "${{ inputs.tag }}"

9
.github/workflows/python-lint.yml vendored
View file

@ -1,6 +1,13 @@
name: flake8 Lint
on: [push, pull_request]
on:
push:
branches:
- master
paths: ['.github/workflows/python-lint.yml', '**/*.py']
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/python-lint.yml', '**/*.py']
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

4
.github/workflows/python-type-check.yml vendored
View file

@ -4,11 +4,13 @@ on:
push:
paths:
- '.github/workflows/python-type-check.yml'
- 'pyrightconfig.json'
- '**.py'
- '**/requirements*.txt'
pull_request:
paths:
- '.github/workflows/python-type-check.yml'
- 'pyrightconfig.json'
- '**.py'
- '**/requirements*.txt'
@ -33,6 +35,6 @@ jobs:
- name: Type-check with Pyright
uses: jakebailey/pyright-action@v2
with:
version: 1.1.370
version: 1.1.382
level: warning
warnings: true

42
.github/workflows/server.yml vendored
View file

@ -20,6 +20,12 @@ on:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
env:
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
LLAMA_LOG_VERBOSITY: 10
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
cancel-in-progress: true
@ -70,20 +76,26 @@ jobs:
run: |
pip install -r examples/server/tests/requirements.txt
- name: Verify server deps
id: verify_server_deps
# Setup nodejs (to be used for verifying bundled index.html)
- uses: actions/setup-node@v4
with:
node-version: 22
- name: Verify bundled index.html
id: verify_server_index_html
run: |
git config --global --add safe.directory $(realpath .)
cd examples/server
git ls-files --others --modified
cd examples/server/webui
git status
./deps.sh
npm ci
npm run build
git status
not_ignored_files="$(git ls-files --others --modified)"
echo "Modified files: ${not_ignored_files}"
if [ -n "${not_ignored_files}" ]; then
echo "Repository is dirty or server deps are not built as expected"
echo "${not_ignored_files}"
modified_files="$(git status -s)"
echo "Modified files: ${modified_files}"
if [ -n "${modified_files}" ]; then
echo "Repository is dirty or server/webui is not built as expected"
echo "Hint: You may need to follow Web UI build guide in server/README.md"
echo "${modified_files}"
exit 1
fi
@ -116,14 +128,14 @@ jobs:
id: server_integration_tests
run: |
cd examples/server/tests
PORT=8888 ./tests.sh
./tests.sh
- name: Slow tests
id: server_integration_tests_slow
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
run: |
cd examples/server/tests
PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
SLOW_TESTS=1 ./tests.sh
server-windows:
@ -173,11 +185,13 @@ jobs:
if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
run: |
cd examples/server/tests
behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
$env:PYTHONIOENCODING = ":replace"
pytest -v -x
- name: Slow tests
id: server_integration_tests_slow
if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
run: |
cd examples/server/tests
behave.exe --stop --no-skipped --no-capture --tags slow
$env:SLOW_TESTS = "1"
pytest -v -x

10
.gitignore vendored
View file

@ -3,6 +3,7 @@
*.a
*.bat
*.bin
*.d
*.dll
*.dot
*.etag
@ -61,6 +62,7 @@ llama-batched-swift
/rpc-server
out/
tmp/
autogen-*.md
# Deprecated
@ -102,6 +104,10 @@ examples/server/*.mjs.hpp
!examples/sycl/*.bat
!examples/sycl/*.sh
# Server Web UI temporary files
node_modules
examples/server/webui/dist
# Python
/.venv
@ -132,3 +138,7 @@ poetry.toml
# Test models for lora adapters
/lora-tests
# Local scripts
/run-vim.sh
/run-chat.sh

2
.gitmodules vendored
View file

@ -1,3 +1,3 @@
[submodule "kompute"]
path = ggml/src/kompute
path = ggml/src/ggml-kompute/kompute
url = https://github.com/nomic-ai/kompute.git

186
AUTHORS
View file

@ -1,4 +1,4 @@
# date: Wed Jun 26 19:36:34 EEST 2024
# date: Thu Nov 28 20:46:15 EET 2024
# this file is auto-generated by scripts/gen-authors.sh
0cc4m <picard12@live.de>
@ -7,6 +7,7 @@
2f38b454 <dxf@protonmail.com>
3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
44670 <44670@users.noreply.github.com>
65a <10104049+65a@users.noreply.github.com>
AN Long <aisk@users.noreply.github.com>
AT <manyoso@users.noreply.github.com>
Aarni Koskela <akx@iki.fi>
@ -19,20 +20,28 @@ Adithya Balaji <adithya.b94@gmail.com>
AdithyanI <adithyan.i4internet@gmail.com>
Adrian <smith.adriane@gmail.com>
Adrian Hesketh <a-h@users.noreply.github.com>
Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com>
Ahmet Zeer <ahmed.zeer@std.yildiz.edu.tr>
AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
AidanBeltonS <aidan.belton@codeplay.com>
Aisuko <urakiny@gmail.com>
Akarshan Biswas <akarshan.biswas@gmail.com>
Akarshan Biswas <akarshanbiswas@fedoraproject.org>
Al Mochkin <14274697+amochkin@users.noreply.github.com>
Albert Jin <albert.jin@gmail.com>
Alberto <57916483+albbus-stack@users.noreply.github.com>
Alberto Cabrera Pérez <alberto.cabrera@codeplay.com>
Alberto Cabrera Pérez <alberto.cabrera@intel.com>
Alex <awhill19@icloud.com>
Alex Azarov <alex@azarov.by>
Alex Azarov <alexander.azarov@mapbox.com>
Alex Klinkhamer <from.github.com.917@grencez.dev>
Alex Klinkhamer <git@grencez.dev>
Alex Nguyen <tiendung@users.noreply.github.com>
Alex O'Connell <35843486+acon96@users.noreply.github.com>
Alex Petenchea <alex.petenchea@gmail.com>
Alex Renda <alexrenda@users.noreply.github.com>
Alex Tuddenham <61622354+AlexsCode@users.noreply.github.com>
Alex von Gluck IV <kallisti5@unixzen.com>
Alexey Parfenov <zxed@alkatrazstudio.net>
Ali Chraghi <63465728+alichraghi@users.noreply.github.com>
@ -45,18 +54,25 @@ AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
Ananta Bastola <anantarajbastola@gmail.com>
Anas Ahouzi <112881240+aahouzi@users.noreply.github.com>
András Salamon <ott2@users.noreply.github.com>
Andreas (Andi) Kunar <andreask@msn.com>
Andrei <abetlen@gmail.com>
Andrew Canis <andrew.canis@gmail.com>
Andrew Downing <andrew2085@gmail.com>
Andrew Duffy <a10y@users.noreply.github.com>
Andrew Godfrey <AndrewGodfrey@users.noreply.github.com>
Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com>
Andy Salerno <andysalerno@gmail.com>
Andy Tai <andy-tai@users.noreply.github.com>
Anthony Van de Gejuchte <anthonyvdgent@gmail.com>
Antonis Makropoulos <benuix@gmail.com>
Arik Poznanski <arikpoz@users.noreply.github.com>
Armen Kaleshian <kriation@users.noreply.github.com>
Artem <guinmoon@gmail.com>
Artem Zinnatullin <ceo@abstractny.gay>
Artyom Lebedev <vagran.ast@gmail.com>
Asbjørn Olling <asbjornolling@gmail.com>
Ásgeir Bjarni Ingvarsson <asgeir@fundinn.org>
Asghar Ghorbani <a-ghorbani@users.noreply.github.com>
Ashish <1856117+ashishdatta@users.noreply.github.com>
Ashok Gelal <401055+ashokgelal@users.noreply.github.com>
Ashraful Islam <ashraful.meche@gmail.com>
@ -76,12 +92,16 @@ Ben Williams <ben@719ben.com>
Benjamin Findley <39356821+Kartoffelsaft@users.noreply.github.com>
Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com>
Bernat Vadell <hounter.caza@gmail.com>
Bert Wagner <github@bertwagner.com>
Bingan <70050083+binganao@users.noreply.github.com>
Bjarke Viksøe <164612031+bviksoe@users.noreply.github.com>
Bodo Graumann <mail@bodograumann.de>
Bono Lv <lvscar@users.noreply.github.com>
Borislav Stanimirov <b.stanimirov@abv.bg>
Branden Butler <bwtbutler@hotmail.com>
Brandon Squizzato <35474886+bsquizz@users.noreply.github.com>
Brian <mofosyne@gmail.com>
Brian Cunnie <brian.cunnie@gmail.com>
Bruce MacDonald <brucewmacdonald@gmail.com>
Bryan Honof <bryanhonof@gmail.com>
CJ Pais <cj@cjpais.com>
@ -90,32 +110,47 @@ Calvin Laurenson <calvin@laurenson.dev>
Cameron <csteele@steelecameron.com>
Cameron Kaiser <classilla@users.noreply.github.com>
Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
CarryFun <76023481+CarryFun@users.noreply.github.com>
Carsten Kragelund Jørgensen <carsten@kragelund.me>
CarterLi999 <664681047@qq.com>
Casey Primozic <casey@cprimozic.net>
Casey Primozic <me@ameo.link>
CausalLM <148736309+CausalLM@users.noreply.github.com>
Cebtenzzre <cebtenzzre@gmail.com>
Chad Brewbaker <crb002@gmail.com>
Changyeon Kim <cyzero.kim@samsung.com>
Chao Jiang <jc19chaoj@zoho.com>
Charles Xu <63788048+chaxu01@users.noreply.github.com>
Charles Xu <charles.xu@arm.com>
Chen Xi <xi2.chen@intel.com>
Chen Xi <xixichen08@foxmail.com>
Cheng Shao <terrorjack@type.dance>
Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
Chris Elrod <elrodc@gmail.com>
Chris Kuehl <ckuehl@ckuehl.me>
Christian Demsar <christian@github.email.demsar.us>
Christian Demsar <crasm@git.vczf.us>
Christian Falch <875252+chrfalch@users.noreply.github.com>
Christian Kögler <ck3d@gmx.de>
Christian Köhnenkamp <cvk5@me.com>
Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com>
Clark Saben <76020733+csaben@users.noreply.github.com>
Clint Herron <hanclinto@gmail.com>
Conrad Kramer <conrad@conradkramer.com>
CrispStrobe <154636388+CrispStrobe@users.noreply.github.com>
Csaba Kecskemeti <csaba.kecskemeti@gmail.com>
Cuong Trinh Manh <nguoithichkhampha@gmail.com>
DAN™ <dranger003@gmail.com>
Damian Stewart <d@damianstewart.com>
Dan Johansson <164997844+eddnjjn@users.noreply.github.com>
Dan Johansson <dan.johansson@arm.com>
Dane Madsen <dane_madsen@hotmail.com>
DaniAndTheWeb <57776841+DaniAndTheWeb@users.noreply.github.com>
Daniel Bevenius <daniel.bevenius@gmail.com>
Daniel Drake <drake@endlessos.org>
Daniel Hiltgen <dhiltgen@users.noreply.github.com>
Daniel Illescas Romero <illescas.daniel@protonmail.com>
Daniel Kleine <53251018+d-kleine@users.noreply.github.com>
Daniele <57776841+daniandtheweb@users.noreply.github.com>
DannyDaemonic <DannyDaemonic@gmail.com>
Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
@ -129,19 +164,28 @@ David Pflug <david@pflug.email>
David Renshaw <dwrenshaw@gmail.com>
David Sommers <12738+databyte@users.noreply.github.com>
David Yang <davidyang6us@gmail.com>
DavidKorczynski <david@adalogics.com>
Dawid Potocki <github@dawidpotocki.com>
Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
Dean <Dean.Sinaean@gmail.com>
Deins <deinsegle@gmail.com>
Denis Spasyuk <34203011+dspasyuk@users.noreply.github.com>
Derrick T. Woolworth <dwoolworth@gmail.com>
Deven Mistry <31466137+deven367@users.noreply.github.com>
Dibakar Gope <dibakar.gope@arm.com>
Didzis Gosko <didzis@users.noreply.github.com>
Diego Devesa <slarengh@gmail.com>
Diogo Teles Sant'Anna <diogoteles@google.com>
Djip007 <djip.perois@free.fr>
Don Mahurin <dmahurin@users.noreply.github.com>
DooWoong Lee (David) <manics99@naver.com>
Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
Dou Xinpeng <15529241576@163.com>
Dou Xinpeng <81913537+Dou-Git@users.noreply.github.com>
Douglas Hanley <thesecretaryofwar@gmail.com>
Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
Ebey Abraham <ebey97@gmail.com>
Echo Nolan <echo@echonolan.net>
Ed Lee <edilee@mozilla.com>
Ed Lepedus <ed.lepedus@googlemail.com>
Eddie-Wang <wangjinheng1120@163.com>
@ -151,10 +195,13 @@ Elbios <141279586+Elbios@users.noreply.github.com>
Elton Kola <eltonkola@gmail.com>
Engininja2 <139037756+Engininja2@users.noreply.github.com>
Equim <sayaka@ekyu.moe>
Eric Curtin <ecurtin@redhat.com>
Eric Curtin <ericcurtin17@gmail.com>
Eric Sommerlade <es0m@users.noreply.github.com>
Eric Zhang <34133756+EZForever@users.noreply.github.com>
Erik Garrison <erik.garrison@gmail.com>
Erik Scholz <Green-Sky@users.noreply.github.com>
Esko Toivonen <eskot98@gmail.com>
Ettore Di Giacinto <mudler@users.noreply.github.com>
Evan Jones <evan.q.jones@gmail.com>
Evan Miller <emmiller@gmail.com>
@ -166,19 +213,26 @@ FK <sozforex@gmail.com>
Fabian <cmdrf@users.noreply.github.com>
Fabio R. Sluzala <Fabio3rs@users.noreply.github.com>
Faez Shakil <faez.shakil@gmail.com>
Faisal Zaghloul <faisal.zaghloul@gmail.com>
Faisal Zaghloul <quic_fzaghlou@quicinc.com>
Fan Shupei <dymarkfan@outlook.com>
FantasyGmm <16450052+FantasyGmm@users.noreply.github.com>
Farbod Bijary <110523279+farbodbj@users.noreply.github.com>
Fattire <528174+fat-tire@users.noreply.github.com>
Felix <stenbackfelix@gmail.com>
Finn Voorhees <finnvoorhees@gmail.com>
Firat <firatkiral@gmail.com>
FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
Folko-Ven <71110216+Folko-Ven@users.noreply.github.com>
Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com>
Francisco Melo <43780565+francis2tm@users.noreply.github.com>
Frank Mai <thxcode0824@gmail.com>
FrankHB <frankhb1989@gmail.com>
Frankie Robertson <frankier@users.noreply.github.com>
Fred Douglas <43351173+fredlas@users.noreply.github.com>
Frederik Vogel <Schaltfehler@users.noreply.github.com>
Gabe Goodhart <gabe.l.hart@gmail.com>
Gabe Goodhart <ghart@us.ibm.com>
GainLee <perfecter.gen@gmail.com>
Galunid <karolek1231456@gmail.com>
Gary Linscott <glinscott@gmail.com>
@ -187,11 +241,13 @@ Gavin Zhao <gavinzhaojw@protonmail.com>
Genkagaku.GPT <hlhr202@163.com>
Georgi Gerganov <ggerganov@gmail.com>
Gilad S <giladgd@users.noreply.github.com>
Gilad S. <7817232+giladgd@users.noreply.github.com>
Giuseppe Scrivano <giuseppe@scrivano.org>
GiviMAD <GiviMAD@users.noreply.github.com>
Govlzkoy <gotope@users.noreply.github.com>
Guillaume "Vermeille" Sanchez <Guillaume.V.Sanchez@gmail.com>
Guillaume Wenzek <gwenzek@users.noreply.github.com>
Guoliang Hua <32868157+nbcsm@users.noreply.github.com>
Guoteng <32697156+SolenoidWGT@users.noreply.github.com>
Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
Haggai Nuchi <h.nuchi@gmail.com>
@ -213,11 +269,14 @@ Hong Bo PENG <penghb@cn.ibm.com>
Hongyu Ouyang <96765450+casavaca@users.noreply.github.com>
Howard Su <howard0su@gmail.com>
Hua Jiang <allenhjiang@outlook.com>
Huang Qi <huangqi3@xiaomi.com>
Huawei Lin <huaweilin.cs@gmail.com>
Hugo Roussel <hugo.rous@gmail.com>
Huifeng Ou <79071290+ho2103@users.noreply.github.com>
Ian Bull <irbull@eclipsesource.com>
Ian Bull <irbull@gmail.com>
Ian Scrivener <github@zilogy.asia>
Icecream95 <the.real.icecream95@gmail.com>
Ido S <ido.pluto@gmail.com>
IgnacioFDM <ignaciofdm@gmail.com>
Igor Okulist <okigan@gmail.com>
@ -226,11 +285,15 @@ Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Ionoclast Laboratories <brigham@ionoclast.com>
Isaac McFadyen <isaac@imcf.me>
IsaacDynamo <61521674+IsaacDynamo@users.noreply.github.com>
Ivan <nekotekina@gmail.com>
Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
Ivan Komarov <Ivan.Komarov@dfyz.info>
Ivan Stepanov <ivanstepanovftw@gmail.com>
JH23X <165871467+JH23X@users.noreply.github.com>
Jack Mousseau <jack@software.inc>
Jack Mousseau <jmousseau@users.noreply.github.com>
JackJollimore <130917767+JackJollimore@users.noreply.github.com>
Jaeden Amero <jaeden@patater.com>
Jaemin Son <woalsdnd@gmail.com>
Jag Chadha <jagtesh@gmail.com>
Jakub N <jakubniemczyk97@gmail.com>
@ -243,10 +306,14 @@ Jannis Schönleber <joennlae@gmail.com>
Jared Van Bortel <cebtenzzre@gmail.com>
Jared Van Bortel <jared@nomic.ai>
Jason McCartney <jmac@theroot.org>
Jason Stillerman <jason.t.stillerman@gmail.com>
Jean-Christophe Hoelt <hoelt@fovea.cc>
Jean-Michaël Celerier <jeanmichael.celerier+github@gmail.com>
Jed Fox <git@jedfox.com>
Jeff Bolz <jbolz@nvidia.com>
Jeffrey Morgan <jmorganca@gmail.com>
Jeffrey Quesnelle <emozilla@nousresearch.com>
Jeroen Mostert <jeroen.mostert@cm.com>
Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
Jeximo <jeximo@gmail.com>
Jhen-Jie Hong <iainst0409@gmail.com>
@ -258,6 +325,9 @@ Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
Jiří Sejkora <Sejseloid@gmail.com>
Joan Fontanals <jfontanalsmartinez@gmail.com>
Joan Fontanals <joan.fontanals.martinez@jina.ai>
João Dinis Ferreira <hello@joaof.eu>
Joe Eli McIlvain <joe.eli.mac@gmail.com>
Joe Todd <joe.todd@codeplay.com>
Johan <JohanAR@users.noreply.github.com>
Johannes Gäßler <johannesg@5d6.de>
Johannes Rudolph <johannes.rudolph@gmail.com>
@ -274,7 +344,9 @@ Joyce <joycebrum@google.com>
Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
Judd <foldl@users.noreply.github.com>
Julius Arkenberg <arki05@users.noreply.github.com>
Jun Hee Yoo <contact.jhyoo@gmail.com>
Jun Jie <71215065+junnjiee16@users.noreply.github.com>
Junil Kim <logyourself@gmail.com>
Junyang Lin <justinlin930319@hotmail.com>
Juraj Bednar <juraj@bednar.io>
Justin Parker <jparkerweb@gmail.com>
@ -292,12 +364,14 @@ Karthik Sethuraman <k.seth1993@gmail.com>
Kasumi <90275229+kasumi-1@users.noreply.github.com>
Kawrakow <48489457+ikawrakow@users.noreply.github.com>
Keiichi Tabata <keiichi.tabata@outlook.com>
Keke Han <hankeke303@163.com>
Kenvix ⭐ <kenvixzure@live.com>
Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
Kevin Gibbons <bakkot@gmail.com>
Kevin Ji <1146876+kevinji@users.noreply.github.com>
Kevin Kwok <antimatter15@gmail.com>
Kevin Lo <kevlo@kevlo.org>
Kevin Wang <kevmo314@gmail.com>
Kolen Cheung <ickc@users.noreply.github.com>
Konstantin Herud <konstantin.herud@denkbares.com>
Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
@ -315,22 +389,29 @@ LeonEricsson <70749762+LeonEricsson@users.noreply.github.com>
Leonardo Neumann <leonardo@neumann.dev.br>
Li Tan <tanliboy@gmail.com>
Linwei Wang <wanix1988@gmail.com>
Liu Jia <109258120+Septa2112@users.noreply.github.com>
Liu Jia <jia3.liu@intel.com>
LoganDark <github@logandark.mozmail.com>
Loïc Carrère <loic.carrere@gmail.com>
LostRuins <39025047+LostRuins@users.noreply.github.com>
Luciano <lucianostrika44@gmail.com>
Luo Tian <lt@basecity.com>
Lyle Dean <dean@lyle.dev>
M-A <maruel@gmail.com>
M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
Ma Mingfei <mingfei.ma@intel.com>
Maarten ter Huurne <maarten@treewalker.org>
Mack Straight <eiz@users.noreply.github.com>
Maël Kerbiriou <m431.kerbiriou@gmail.com>
MaggotHATE <clay1326@gmail.com>
Mahesh Madhav <67384846+heshpdx@users.noreply.github.com>
Manuel <44313466+makuche@users.noreply.github.com>
Marc Köhlbrugge <subscriptions@marckohlbrugge.com>
Marco Matthies <71844+marcom@users.noreply.github.com>
Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
Marian Cepok <marian.cepok@gmail.com>
Mark Fairbairn <thebaron88@gmail.com>
Mark Zhuang <zhuangqiubin@gmail.com>
Marko Tasic <mtasic85@gmail.com>
Markus Tavenrath <mtavenrath@users.noreply.github.com>
Martin Delille <martin@delille.org>
@ -342,11 +423,15 @@ MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com>
Mateusz Charytoniuk <mateusz.charytoniuk@protonmail.com>
Matheus C. França <matheus-catarino@hotmail.com>
Matheus Gabriel Alves Silva <matheusgasource@gmail.com>
Mathieu Geli <mathieu.geli@gmail.com>
Mathieu Nayrolles <MathieuNls@users.noreply.github.com>
Mathijs Henquet <mathijs.henquet@gmail.com>
Mathijs de Bruin <mathijs@mathijsfietst.nl>
Matt Clayton <156335168+mattjcly@users.noreply.github.com>
Matt Pulver <matt.pulver@heavy.ai>
Matt Stephenson <mstephenson6@users.noreply.github.com>
Matteo Boschini <12133566+mbosc@users.noreply.github.com>
Matteo Mortari <matteo.mortari@gmail.com>
Mattheus Chediak <shammcity00@gmail.com>
Matthew Tejo <matthew.tejo@gmail.com>
Matvey Soloviev <blackhole89@gmail.com>
@ -356,8 +441,10 @@ Maxime <672982+maximegmd@users.noreply.github.com>
Maximilian Winter <maximilian.winter.91@gmail.com>
Meng Zhang <meng@tabbyml.com>
Meng, Hengyu <hengyu.meng@intel.com>
Mengqing Cao <cmq0113@163.com>
Merrick Christensen <merrick.christensen@gmail.com>
Michael Coppola <m18coppola@gmail.com>
Michael Francis <edude03@gmail.com>
Michael Hueschen <m@mhueschen.dev>
Michael Kesper <mkesper@schokokeks.org>
Michael Klimenko <mklimenko29@gmail.com>
@ -365,41 +452,57 @@ Michael Podvitskiy <podvitskiymichael@gmail.com>
Michael Potter <NanoTekGuy@Gmail.com>
Michael de Gans <michael.john.degans@gmail.com>
Michaël de Vries <vriesdemichael@gmail.com>
Michał Tuszyński <srgtuszy@gmail.com>
Mihai <mihai.chirculescu@yahoo.com>
Mike <ytianhui2004@gmail.com>
Mikko Juola <mikjuo@gmail.com>
Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
Minsoo Cheong <icycle0409@snu.ac.kr>
Mirko185 <mirkosig@gmail.com>
Mirror Azure <54669636+MirrorAzure@users.noreply.github.com>
MistApproach <98988043+MistApproach@users.noreply.github.com>
Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
Mohammadreza Hendiani <mohammad.r.hendiani@gmail.com>
Molly Sophia <mollysophia379@gmail.com>
MorganRO8 <47795945+MorganRO8@users.noreply.github.com>
Murilo Santana <mvrilo@gmail.com>
Musab Gultekin <musabgultekin@users.noreply.github.com>
Nam D. Tran <42194884+namtranase@users.noreply.github.com>
Nathan Epstein <nate2@umbc.edu>
Natsu <chino@hotococoa.moe>
NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
Nebula <infinitewormhole@gmail.com>
Neo Zhang <14088817+arthw@users.noreply.github.com>
Neo Zhang <zhang.jianyu@outlook.com>
Neo Zhang Jianyu <jianyu.zhang@intel.com>
Neuman Vong <neuman.vong@gmail.com>
Nexes the Old <124105151+Nexesenex@users.noreply.github.com>
Nexesenex <124105151+Nexesenex@users.noreply.github.com>
Niall Coates <1349685+Niall-@users.noreply.github.com>
Nicholai Tukanov <nicholaitukanov@gmail.com>
Nico Bosshard <nico@bosshome.ch>
Nicolai Weitkemper <kontakt@nicolaiweitkemper.de>
Nicolás Pérez <nicolas_perez@brown.edu>
Nigel Bosch <pnigelb@gmail.com>
Niklas Korz <niklas@niklaskorz.de>
NikolaiLyssogor <59844691+NikolaiLyssogor@users.noreply.github.com>
Nikolas <127742645+nneubacher@users.noreply.github.com>
Nindaleth <Nindaleth@users.noreply.github.com>
OSecret <135510162+OLSecret@users.noreply.github.com>
Oleksandr Nikitin <oleksandr@tvori.info>
Oleksii Maryshchenko <oleksii.maryshchenko@gmail.com>
Olivier Chafik <ochafik@users.noreply.github.com>
Ondřej Čertík <ondrej@certik.us>
Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
PAB <pierreantoine.bannier@gmail.com>
Pablo Duboue <pablo.duboue@gmail.com>
Pascal Patry <ppatry@mtacitlabs.com>
Patrice Ferlet <metal3d@gmail.com>
Paul Tsochantaris <ptsochantaris@icloud.com>
Pavel Zloi <github.com@drteam.rocks>
Pavol Rusnak <pavol@rusnak.io>
Paweł Wodnicki <151604+32bitmicro@users.noreply.github.com>
Pedro Cuenca <pedro@huggingface.co>
Peter Sugihara <peter@campsh.com>
Phil H <5756783+phiharri@users.noreply.github.com>
@ -407,10 +510,15 @@ Philip Taron <philip.taron@gmail.com>
Phillip Kravtsov <phillip@kravtsov.net>
Pierre Alexandre SCHEMBRI <pa.schembri@gmail.com>
Pierrick Hymbert <pierrick.hymbert@gmail.com>
Pieter Ouwerkerk <pieter.ouwerkerk@gmail.com>
Plamen Minev <pacominev@gmail.com>
Prashant Vithule <119530321+Vithulep@users.noreply.github.com>
Przemysław Pawełczyk <przemoc@gmail.com>
Qin Yue Chen <71813199+chenqiny@users.noreply.github.com>
Qingyou Meng <meng.qingyou@gmail.com>
Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
R0CKSTAR <xiaodong.ye@mthreads.com>
R0CKSTAR <yeahdongcn@gmail.com>
RJ Adriaansen <adriaansen@eshcc.eur.nl>
Radoslav Gerganov <rgerganov@gmail.com>
Radosław Gryta <radek.gryta@gmail.com>
@ -419,11 +527,13 @@ Raj Hammeer Singh Hada <hammeerraj@gmail.com>
Ralph Soika <ralph.soika@imixs.com>
Rand Xie <randxiexyy29@gmail.com>
Randall Fitzgerald <randall@dasaku.net>
Random Fly <renfei8@live.cn>
Reinforce-II <fate@eastal.com>
Ren Xuancheng <jklj077@users.noreply.github.com>
Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com>
RhinoDevel <RhinoDevel@users.noreply.github.com>
Riceball LEE <snowyu.lee@gmail.com>
Rich Dougherty <rich@rd.nz>
Richard Kiss <him@richardkiss.com>
Richard Roberson <richardr1126@gmail.com>
Rick G <26732651+TheFlipbook@users.noreply.github.com>
@ -439,21 +549,30 @@ Robey Holderith <robey@flaminglunchbox.net>
Robyn <robyngraf@users.noreply.github.com>
Roger Meier <r.meier@siemens.com>
Roland <14355895+rbur0425@users.noreply.github.com>
Romain Biessy <romain.biessy@codeplay.com>
Romain D <90720+Artefact2@users.noreply.github.com>
Romain Neutron <romain@neutron.io>
Roman Parykin <donderom@gmail.com>
Ron Evans <ron@hybridgroup.com>
Ron Jailall <rojailal@gmail.com>
Roni <sulpher@gmx.net>
Ronny Brendel <ronnybrendel@gmail.com>
Ronsor <ronsor@ronsor.pw>
Rowan Hart <rowanbhart@gmail.com>
Ruchira Hasaranga <ruchira66@gmail.com>
Ruixin Huang <18860020911@163.com>
Rune <43761327+Rune-AI@users.noreply.github.com>
RunningLeon <maningsheng@sensetime.com>
RunningLeon <mnsheng@yeah.net>
Ryan Landay <rlanday@gmail.com>
Ryder Wishart <ryderwishart@gmail.com>
Ryuei <louixs@users.noreply.github.com>
Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
SRHMorris <69468379+SRHMorris@users.noreply.github.com>
SXX <sxx1136965276@gmail.com>
SakuraUmi <yukinon244@gmail.com>
Salvador E. Tropea <stropea@inti.gob.ar>
Salvatore Mesoraca <s.mesoraca16@gmail.com>
Sam Spilsbury <smspillaz@gmail.com>
Sami Farin <3876865+Safari77@users.noreply.github.com>
Samuel Maynard <samwmaynard@gmail.com>
@ -463,23 +582,29 @@ Sebastián A <sebastian.aedo29@gmail.com>
SebastianApel <13675545+SebastianApel@users.noreply.github.com>
Senemu <10880819+Senemu@users.noreply.github.com>
Sergey Alirzaev <zl29ah@gmail.com>
Sergio López <slp@redhat.com>
Sergio López <slp@sinrega.org>
Sertaç Özercan <852750+sozercan@users.noreply.github.com>
SeungWon Jeong <65549245+redlion0929@users.noreply.github.com>
ShadovvBeast <ShadovvBeast@gmail.com>
Shakhar Dasgupta <shakhardasgupta@gmail.com>
Shane A <shanea@allenai.org>
Shangning Xu <32517059+xushangning@users.noreply.github.com>
Shankar <gshankar.87@gmail.com>
Shanshan Shen <467638484@qq.com>
Shijie <821898965@qq.com>
Shintarou Okada <kokuzen@gmail.com>
Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
Shouzheng Liu <lshzh.hi@gmail.com>
Shuichi Tsutsumi <shuichi0526@gmail.com>
Shupei Fan <dymarkfan@outlook.com>
Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Simon Willison <swillison@gmail.com>
Siwen Yu <yusiwen@gmail.com>
Sky Yan <skyan83@gmail.com>
Slaren <2141330+slaren@users.noreply.github.com>
Slava Primenko <primenko.s@gmail.com>
Small Grass Forest <zixuanxcl@gmail.com>
SoftwareRenderer <138734813+SoftwareRenderer@users.noreply.github.com>
Someone <sergei.kozlukov@aalto.fi>
Someone Serge <sergei.kozlukov@aalto.fi>
@ -491,12 +616,15 @@ Stefan Sydow <stefan@sydow.email>
Steffen Röcker <sroecker@gmail.com>
Stephan Walter <stephan@walter.name>
Stephen Nichols <snichols@users.noreply.github.com>
Steve Bonds <sbonds@gmail.com>
Steve Grubb <ausearch.1@gmail.com>
Steven Prichard <spprichard20@gmail.com>
Steven Roussey <sroussey@gmail.com>
Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
StrangeBytesDev <141275258+StrangeBytesDev@users.noreply.github.com>
Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
SuperUserNameMan <yoann@terminajones.com>
Sutou Kouhei <kou@cozmixng.org>
Tai Duc Nguyen <taiducnguyen.drexel@gmail.com>
Taikono-Himazin <kazu@po.harenet.ne.jp>
Tameem <113388789+AhmadTameem@users.noreply.github.com>
@ -507,7 +635,9 @@ Theia Vogel <theia@vgel.me>
Thérence <13496987+Royalphax@users.noreply.github.com>
Thibault Terrasson <thibault.terrasson@gmail.com>
Thomas Klausner <wiz@gatalith.at>
Thorsten Sommer <SommerEngineering@users.noreply.github.com>
Tim Miller <drasticactions@users.noreply.github.com>
Tim Wang <overocean@gmail.com>
Timmy Knight <r2d2fish@gmail.com>
Timothy Cronin <40186632+4imothy@users.noreply.github.com>
Ting Lou <ting.lou@gmail.com>
@ -517,24 +647,31 @@ Tom C <tom.corelis@gmail.com>
Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Tomas <tom.tomas.36478119@gmail.com>
Tomáš Pazdiora <tomas.pazdiora@gmail.com>
Tony Wasserka <4840017+neobrain@users.noreply.github.com>
Tristan Druyen <tristan@vault81.mozmail.com>
Tristan Ross <rosscomputerguy@protonmail.com>
Trivikram Kamat <16024985+trivikr@users.noreply.github.com>
Tungsten842 <886724vf@anonaddy.me>
Tungsten842 <quantmint@protonmail.com>
Tushar <ditsuke@protonmail.com>
UEXTM.com <84163508+uextm@users.noreply.github.com>
Ujjawal Panchal <31011628+Ujjawal-K-Panchal@users.noreply.github.com>
Ulrich Drepper <drepper@gmail.com>
Uzo Nweke <uzoechi@gmail.com>
Vaibhav Srivastav <vaibhavs10@gmail.com>
Val Kharitonov <mail@kharvd.com>
Valentin Konovalov <valle.ketsujin@gmail.com>
Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com>
Vali Malinoiu <0x4139@gmail.com>
Victor Nogueira <felladrin@gmail.com>
Victor Z. Peng <ziliangdotme@gmail.com>
Viet-Anh NGUYEN (Andrew) <vietanh.dev@gmail.com>
Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com>
Vlad <spitfireage@gmail.com>
Vladimir <bogdad@gmail.com>
Vladimir Malyutin <first-leon@yandex.ru>
Vladimir Zorin <vladimir@deviant.guru>
VoidIsVoid <343750470@qq.com>
Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com>
WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com>
Weird Constructor <weirdconstructor@gmail.com>
@ -551,15 +688,22 @@ Xiang (Kevin) Li <kevinli020508@gmail.com>
Xiao-Yong Jin <jinxiaoyong@gmail.com>
XiaotaoChen <chenxiaotao1234@gmail.com>
Xiaoyi Chen <cxychina@gmail.com>
Xie Yanbo <xieyanbo@gmail.com>
Xingchen Song(宋星辰) <xingchensong1996@163.com>
Xinpeng Dou <81913537+Dou-Git@users.noreply.github.com>
Xuan Son Nguyen <thichthat@gmail.com>
Yaiko <elyaiko@hotmail.com>
Yann Follet <131855179+YannFollet@users.noreply.github.com>
Yaroslav <yaroslav.yashin@me.com>
Yazan Agha-Schrader <mountaiin@icloud.com>
Yiming Cui <conandiy@vip.qq.com>
Yishuo Wang <MeouSker77@outlook.com>
Yoshi Suhara <y.suhara@gmail.com>
Yoshi Suhara <ysuhara@nvidia.com>
Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com>
Yui <dev@sleepyyui.com>
Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
Yusuf Kağan Hanoğlu <hanoglu@yahoo.com>
Yuval Peled <31162840+Yuval-Peled@users.noreply.github.com>
ZHAOKAI WANG <sanxianwei@163.com>
@ -568,6 +712,8 @@ Zay <95888118+isaiahbjork@users.noreply.github.com>
Zenix <zenixls2@gmail.com>
Zhang Peiyuan <a1286225768@gmail.com>
Zheng.Deng <32841220+dengzheng-cloud@users.noreply.github.com>
Zhenwei Jin <109658203+kylo5aby@users.noreply.github.com>
Zhiyuan Li <lizhiyuan@uniartisan.com>
ZhouYuChen <zhouyuchen@naver.com>
Ziad Ben Hadj-Alouane <zied.benhadjalouane@gmail.com>
Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com>
@ -581,6 +727,7 @@ alexpinel <93524949+alexpinel@users.noreply.github.com>
alonfaraj <alonfaraj@gmail.com>
alwqx <kenan3015@gmail.com>
amd-lalithnc <lalithnc@amd.com>
amritahs-ibm <amritahs@linux.vnet.ibm.com>
andrijdavid <david@geek.mg>
anon998 <131767832+anon998@users.noreply.github.com>
anzz1 <anzz1@live.com>
@ -588,14 +735,18 @@ apaz <aarpazdera@gmail.com>
apcameron <37645737+apcameron@users.noreply.github.com>
arch-btw <57669023+arch-btw@users.noreply.github.com>
arcrank <arcrank@gmail.com>
ardfork <134447697+ardfork@users.noreply.github.com>
arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
at8u <129688334+at8u@users.noreply.github.com>
automaticcat <daogiatuank54@gmail.com>
awatuna <23447591+awatuna@users.noreply.github.com>
b4b4o <zwbao@foxmail.com>
bandoti <141645996+bandoti@users.noreply.github.com>
beiller <beiller@gmail.com>
bhubbb <79117352+bhubbb@users.noreply.github.com>
bmwl <brian.marshall@tolko.com>
bobqianic <129547291+bobqianic@users.noreply.github.com>
brucepro <git@brucepro.net>
bryanSwk <93190252+bryanSwk@users.noreply.github.com>
bsilvereagle <bsilvereagle@users.noreply.github.com>
bssrdf <merlintiger@hotmail.com>
@ -614,10 +765,14 @@ cpumaxx <163466046+cpumaxx@users.noreply.github.com>
crasm <crasm@git.vczf.net>
crasm <crasm@git.vczf.us>
daboe01 <daboe01@googlemail.com>
daghanerdonmez <44506702+daghanerdonmez@users.noreply.github.com>
daminho <37615795+daminho@users.noreply.github.com>
david raistrick <keen99@users.noreply.github.com>
ddh0 <dylanhalladay02@icloud.com>
ddpasa <112642920+ddpasa@users.noreply.github.com>
deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
devojony <61173062+devojony@users.noreply.github.com>
ditsuke <ditsuke@protonmail.com>
divinity76 <divinity76@gmail.com>
dm4 <sunrisedm4@gmail.com>
dotpy314 <33351922+dotpy314@users.noreply.github.com>
@ -629,14 +784,18 @@ ebraminio <ebraminio@gmail.com>
eiery <19350831+eiery@users.noreply.github.com>
eric8607242 <e0928021388@gmail.com>
fairydreaming <166155368+fairydreaming@users.noreply.github.com>
fengerhu1 <2748250768@qq.com>
fraxy-v <65565042+fraxy-v@users.noreply.github.com>
github-actions[bot] <github-actions[bot]@users.noreply.github.com>
gliptic <gliptic@users.noreply.github.com>
goerch <jhr.walter@t-online.de>
grahameth <96447521+grahameth@users.noreply.github.com>
gtygo <gtydoit@gmail.com>
gwjr <502526+gwjr@users.noreply.github.com>
h-h-h-h <13482553+h-h-h-h@users.noreply.github.com>
hankcs <cnhankmc@gmail.com>
haopeng <657407891@qq.com>
hipudding <huafengchun@gmail.com>
hoangmit <hoangmit@users.noreply.github.com>
hongbo.mo <352280764@qq.com>
hopkins385 <98618192+hopkins385@users.noreply.github.com>
@ -649,12 +808,14 @@ hxer7963 <hxer7963@gmail.com>
hydai <z54981220@gmail.com>
iSma <ismail.senhaji@gmail.com>
iacore <74560659+iacore@users.noreply.github.com>
icppWorld <124377669+icppWorld@users.noreply.github.com>
igarnier <igarnier@protonmail.com>
intelmatt <61025942+intelmatt@users.noreply.github.com>
iohub <rickyang.pro@gmail.com>
jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
jaime-m-p <167997752+jaime-m-p@users.noreply.github.com>
jameswu2014 <545426914@qq.com>
jdomke <28772296+jdomke@users.noreply.github.com>
jiez <373447296@qq.com>
jneem <joeneeman@gmail.com>
joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
@ -677,28 +838,35 @@ klosax <131523366+klosax@users.noreply.github.com>
kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
kunnis <kunnis@users.noreply.github.com>
kuronekosaiko <EvanChanJ@163.com>
kustaaya <58045274+kustaaya@users.noreply.github.com>
kuvaus <22169537+kuvaus@users.noreply.github.com>
kwin1412 <42286931+kwin1412@users.noreply.github.com>
l3utterfly <gc.pthzfoldr@gmail.com>
laik <laik.lj@me.com>
ldwang <ftgreat@163.com>
le.chang <cljs118@126.com>
leejet <leejet714@gmail.com>
leo-pony <nengjunma@outlook.com>
limitedAtonement <limitedAtonement@users.noreply.github.com>
liuwei-git <14815172+liuwei-git@users.noreply.github.com>
lon <114724657+longregen@users.noreply.github.com>
loonerin <132926317+loonerin@users.noreply.github.com>
ltoniazzi <61414566+ltoniazzi@users.noreply.github.com>
luoyu-intel <yu.luo@intel.com>
m3ndax <adrian.goessl@outlook.com>
maddes8cht <55592906+maddes8cht@users.noreply.github.com>
makomk <makosoft@googlemail.com>
manikbhandari <mbbhandarimanik2@gmail.com>
maor-ps <154728172+maor-ps@users.noreply.github.com>
matiaslin <45382001+matiaslin@users.noreply.github.com>
matteo <matteogeniaccio@yahoo.it>
mdrokz <mohammadmunshi@gmail.com>
mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
minarchist <minarchist@users.noreply.github.com>
mj-shifu <77107165+mj-shifu@users.noreply.github.com>
mmyjona <jonathan.gonse@gmail.com>
momonga <115213907+mmnga@users.noreply.github.com>
momonga <146910567+mmngays@users.noreply.github.com>
moritzbrantner <31051084+moritzbrantner@users.noreply.github.com>
mzcu <milos.cubrilo@gmail.com>
nanahi <130121847+na-na-hi@users.noreply.github.com>
@ -716,8 +884,10 @@ omahs <73983677+omahs@users.noreply.github.com>
oobabooga <112222186+oobabooga@users.noreply.github.com>
opparco <parco.opaai@gmail.com>
ostix360 <55257054+ostix360@users.noreply.github.com>
pculliton <phillipculliton@gmail.com>
pengxin99 <pengxin.yuan@intel.com>
perserk <perserk@gmail.com>
piDack <104877312+piDack@users.noreply.github.com>
pmysl <piotr.myslinski@outlook.com>
postmasters <namnguyen@google.com>
pudepiedj <pudepiedj@gmail.com>
@ -733,6 +903,7 @@ runfuture <runfuture@users.noreply.github.com>
sandyiscool <sandyiscool@gmail.com>
sasha0552 <admin@sasha0552.org>
semidark <me@semidark.net>
serhii-nakon <57632032+serhii-nakon@users.noreply.github.com>
sharpHL <132747147+sharpHL@users.noreply.github.com>
shibe2 <shibe@tuta.io>
singularity <12184989+singularity-s0@users.noreply.github.com>
@ -741,42 +912,55 @@ sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com>
slaren <2141330+slaren@users.noreply.github.com>
slaren <slarengh@gmail.com>
snadampal <87143774+snadampal@users.noreply.github.com>
standby24x7 <standby24x7@gmail.com>
staviq <staviq@gmail.com>
stduhpf <stephduh@live.fr>
strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
swittk <switt1995@gmail.com>
takov751 <40316768+takov751@users.noreply.github.com>
tarcey <cey.tarik@gmail.com>
tc-mb <157115220+tc-mb@users.noreply.github.com>
texmex76 <40733439+texmex76@users.noreply.github.com>
thement <40525767+thement@users.noreply.github.com>
thewh1teagle <61390950+thewh1teagle@users.noreply.github.com>
tjohnman <tjohnman@users.noreply.github.com>
toyer <2042519524@qq.com>
tslmy <tslmy@users.noreply.github.com>
ubik2 <ubik2@users.noreply.github.com>
uint256_t <konndennsa@gmail.com>
uint256_t <maekawatoshiki1017@gmail.com>
unbounded <haakon@likedan.net>
uvos <devnull@uvos.xyz>
valiray <133289098+valiray@users.noreply.github.com>
vb <vaibhavs10@gmail.com>
vik <vikhyatk@gmail.com>
viric <viric@viric.name>
vodkaslime <646329483@qq.com>
vvhg1 <94630311+vvhg1@users.noreply.github.com>
vxiiduu <73044267+vxiiduu@users.noreply.github.com>
wangshuai09 <391746016@qq.com>
wbpxre150 <100937007+wbpxre150@users.noreply.github.com>
whoreson <139810751+whoreson@users.noreply.github.com>
woachk <24752637+woachk@users.noreply.github.com>
wonjun Jang <strutive07@gmail.com>
woodx <124784234+woodx9@users.noreply.github.com>
wwoodsTM <104587230+wwoodsTM@users.noreply.github.com>
wzy <32936898+Freed-Wu@users.noreply.github.com>
xaedes <xaedes@gmail.com>
xaedes <xaedes@googlemail.com>
xctan <axunlei@gmail.com>
xloem <0xloem@gmail.com>
yangli2 <yangli2@gmail.com>
yuiseki <yuiseki@gmail.com>
yuri@FreeBSD <yurivict@users.noreply.github.com>
zakkor <edward.partenie@gmail.com>
zhangkaihuo <zhangkaihuo@gmail.com>
zhentaoyu <zhentao.yu@intel.com>
zhouwg <6889919+zhouwg@users.noreply.github.com>
zhouwg <zhouwg2000@gmail.com>
zrm <trustiosity.zrm@gmail.com>
Ștefan-Gabriel Muscalu <legraphista@users.noreply.github.com>
杨朱 · Kiki <baofa.fan@daocloud.io>
源文雨 <41315874+fumiama@users.noreply.github.com>
蕭澧邦 <45505768+shou692199@users.noreply.github.com>
Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>

45
CMakeLists.txt
View file

@ -46,6 +46,13 @@ if (WIN32)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
endif()
if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
add_compile_options("$<$<COMPILE_LANGUAGE:C>:/source-charset:utf-8>")
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/source-charset:utf-8>")
add_compile_options("$<$<COMPILE_LANGUAGE:C>:/execution-charset:utf-8>")
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/execution-charset:utf-8>")
endif()
#
# option list
#
@ -62,6 +69,9 @@ option(LLAMA_SANITIZE_THREAD "llama: enable thread sanitizer" OFF)
option(LLAMA_SANITIZE_ADDRESS "llama: enable address sanitizer" OFF)
option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
# utils
option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
# extra artifacts
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
@ -72,6 +82,7 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
# Required for relocatable CMake package
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
# override ggml options
set(GGML_SANITIZE_THREAD ${LLAMA_SANITIZE_THREAD})
@ -82,11 +93,11 @@ set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
# change the default for these ggml options
if (NOT DEFINED GGML_LLAMAFILE)
set(GGML_LLAMAFILE ON)
set(GGML_LLAMAFILE_DEFAULT ON)
endif()
if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
set(GGML_CUDA_USE_GRAPHS ON)
if (NOT DEFINED GGML_CUDA_GRAPHS)
set(GGML_CUDA_GRAPHS_DEFAULT ON)
endif()
# transition helpers
@ -133,19 +144,27 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")
# At the moment some compile definitions are placed within the ggml/src
# directory but not exported on the `ggml` target. This could be improved by
# determining _precisely_ which defines are necessary for the llama-config
# package.
#
set(GGML_TRANSIENT_DEFINES)
get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
if (GGML_DIR_DEFINES)
list(APPEND GGML_TRANSIENT_DEFINES ${GGML_DIR_DEFINES})
endif()
get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
if (GGML_TARGET_DEFINES)
list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
endif()
get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
# all public headers
set(LLAMA_PUBLIC_HEADERS
${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
install(TARGETS llama LIBRARY PUBLIC_HEADER)
configure_package_config_file(
@ -185,17 +204,19 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
DESTINATION lib/pkgconfig)
#
# programs, examples and tests
# utils, programs, examples and tests
#
add_subdirectory(common)
if (LLAMA_BUILD_COMMON)
add_subdirectory(common)
endif()
if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
include(CTest)
add_subdirectory(tests)
endif ()
endif()
if (LLAMA_BUILD_EXAMPLES)
if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
add_subdirectory(examples)
add_subdirectory(pocs)
endif()

49
CMakePresets.json
View file

@ -24,16 +24,17 @@
"CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
}
},
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } },
{
"name": "arm64-windows-msvc", "hidden": true,
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x86_64", "strategy": "external" },
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x64", "strategy": "external" },
"cacheVariables": {
"CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
}
@ -41,28 +42,44 @@
{
"name": "arm64-windows-llvm", "hidden": true,
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x86_64", "strategy": "external" },
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x64", "strategy": "external" },
"cacheVariables": {
"CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
}
},
{ "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{
"name": "arm64-apple-clang", "hidden": true,
"architecture": { "value": "arm64", "strategy": "external" },
"toolset": { "value": "host=x64", "strategy": "external" },
"cacheVariables": {
"CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
}
},
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug", "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "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-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "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" ] },
{ "name": "x64-windows-sycl-debug" , "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
{ "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
{ "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
{ "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
]
}

3
CODEOWNERS Normal file
View file

@ -0,0 +1,3 @@
# collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
ci/ @ggerganov

18
CONTRIBUTING.md
View file

@ -1,24 +1,25 @@
# Pull requests (for contributors)
- Test your changes:
- Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
- Execute [the full CI locally on your machine](ci/README.md) before publishing
- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
- The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
- Consider allowing write access to your branch for faster review
- Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`)
- If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
- If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
# Pull requests (for collaborators)
- Squash-merge PRs
- Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
- Optionally, pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
- Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
# Coding guidelines
- Avoid adding third-party dependencies, extra files, extra headers, etc.
- Always consider cross-compatibility with other operating systems and architectures
- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
@ -27,3 +28,8 @@
![matmul](media/matmul.png)
# Resources
The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
https://github.com/ggerganov/llama.cpp/projects

714
Makefile
View file

File diff suppressed because it is too large Load diff

29
Package.swift
View file

@ -10,10 +10,16 @@ var sources = [
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.c",
"ggml/src/ggml-quants.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] = []
@ -21,21 +27,26 @@ 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-metal.m")
resources.append(.process("ggml/src/ggml-metal.metal"))
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")
.define("GGML_USE_METAL"),
]
)
#endif
@ -60,13 +71,15 @@ let package = Package(
name: "llama",
path: ".",
exclude: [
"build",
"cmake",
"examples",
"scripts",
"models",
"tests",
"CMakeLists.txt",
"Makefile"
"Makefile",
"ggml/src/ggml-metal-embed.metal"
],
sources: sources,
resources: resources,
@ -75,5 +88,5 @@ let package = Package(
linkerSettings: linkerSettings
)
],
cxxLanguageStandard: .cxx11
cxxLanguageStandard: .cxx17
)

641
README.md
View file

@ -4,62 +4,48 @@
[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
> [!IMPORTANT]
[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
## Recent API changes
- [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006
- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
- [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289)
- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291)
## Hot topics
- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
- Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
- Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017
- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
----
## Description
The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
variety of hardware - locally and in the cloud.
range of hardware - locally and in the cloud.
- Plain C/C++ implementation without any dependencies
- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
- AVX, AVX2 and AVX512 support for x86 architectures
- AVX, AVX2, AVX512 and AMX support for x86 architectures
- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
- Vulkan and SYCL backend support
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
improved significantly thanks to many contributions. It is the main playground for developing new features for the
[ggml](https://github.com/ggerganov/ggml) library.
The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
**Supported models:**
<details>
<summary>Models</summary>
Typically finetunes of the base models below are supported as well.
Instructions for adding support for new models: [HOWTO-add-model.md](docs/development/HOWTO-add-model.md)
#### Text-only
- [X] LLaMA 🦙
- [x] LLaMA 2 🦙🦙
- [x] LLaMA 3 🦙🦙🦙
@ -95,6 +81,8 @@ Typically finetunes of the base models below are supported as well.
- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
- [x] [OLMo](https://allenai.org/olmo)
- [x] [OLMo 2](https://allenai.org/olmo)
- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
@ -107,10 +95,11 @@ Typically finetunes of the base models below are supported as well.
- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
- [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)
(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
**Multimodal models:**
#### Multimodal
- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
- [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
@ -122,12 +111,16 @@ Typically finetunes of the base models below are supported as well.
- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
**Bindings:**
</details>
<details>
<summary>Bindings</summary>
- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
@ -135,347 +128,354 @@ Typically finetunes of the base models below are supported as well.
- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
- C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
- Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
**UI:**
</details>
Unless otherwise noted these projects are open-source with permissive licensing:
- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
- [iohub/collama](https://github.com/iohub/coLLaMA)
- [janhq/jan](https://github.com/janhq/jan) (AGPL)
- [nat/openplayground](https://github.com/nat/openplayground)
- [Faraday](https://faraday.dev/) (proprietary)
- [LMStudio](https://lmstudio.ai/) (proprietary)
- [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
- [ramalama](https://github.com/containers/ramalama) (MIT)
- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
- [ollama/ollama](https://github.com/ollama/ollama)
- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
- [RAGNA Desktop](https://ragna.app/) (proprietary)
- [RecurseChat](https://recurse.chat/) (proprietary)
- [semperai/amica](https://github.com/semperai/amica)
- [withcatai/catai](https://github.com/withcatai/catai)
- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
- [Msty](https://msty.app) (proprietary)
- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file)(Apachev2.0 or later)
- [Dot](https://github.com/alexpinel/Dot) (GPL)
- [MindMac](https://mindmac.app) (proprietary)
- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
- [AIKit](https://github.com/sozercan/aikit) (MIT)
- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
<details>
<summary>UIs</summary>
*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
**Tools:**
- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
- [Dot](https://github.com/alexpinel/Dot) (GPL)
- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
- [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0)
- [janhq/jan](https://github.com/janhq/jan) (AGPL)
- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0)
- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
- [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
- [LARS](https://github.com/abgulati/LARS) (AGPL)
- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
- [LMStudio](https://lmstudio.ai/) (proprietary)
- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
- [MindMac](https://mindmac.app) (proprietary)
- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile) (Apache-2.0)
- [nat/openplayground](https://github.com/nat/openplayground) (MIT)
- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all) (MIT)
- [ollama/ollama](https://github.com/ollama/ollama) (MIT)
- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
- [PocketPal AI](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
- [psugihara/FreeChat](https://github.com/psugihara/FreeChat) (MIT)
- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) (MIT)
- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
- [ramalama](https://github.com/containers/ramalama) (MIT)
- [semperai/amica](https://github.com/semperai/amica) (MIT)
- [withcatai/catai](https://github.com/withcatai/catai) (MIT)
</details>
<details>
<summary>Tools</summary>
- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
**Infrastructure:**
</details>
<details>
<summary>Infrastructure</summary>
- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
</details>
<details>
<summary>Games</summary>
**Games:**
- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
## Demo
<details>
<summary>Typical run using LLaMA v2 13B on M2 Ultra</summary>
```
$ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
I llama.cpp build info:
I UNAME_S: Darwin
I UNAME_P: arm
I UNAME_M: arm64
I CFLAGS: -I. -O3 -std=c11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -pthread -DGGML_USE_K_QUANTS -DGGML_USE_ACCELERATE
I CXXFLAGS: -I. -I./common -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -DGGML_USE_K_QUANTS
I LDFLAGS: -framework Accelerate
I CC: Apple clang version 14.0.3 (clang-1403.0.22.14.1)
I CXX: Apple clang version 14.0.3 (clang-1403.0.22.14.1)
make: Nothing to be done for `default'.
main: build = 1041 (cf658ad)
main: seed = 1692823051
llama_model_loader: loaded meta data with 16 key-value pairs and 363 tensors from models/llama-13b-v2/ggml-model-q4_0.gguf (version GGUF V1 (latest))
llama_model_loader: - type f32: 81 tensors
llama_model_loader: - type q4_0: 281 tensors
llama_model_loader: - type q6_K: 1 tensors
llm_load_print_meta: format = GGUF V1 (latest)
llm_load_print_meta: arch = llama
llm_load_print_meta: vocab type = SPM
llm_load_print_meta: n_vocab = 32000
llm_load_print_meta: n_merges = 0
llm_load_print_meta: n_ctx_train = 4096
llm_load_print_meta: n_ctx = 512
llm_load_print_meta: n_embd = 5120
llm_load_print_meta: n_head = 40
llm_load_print_meta: n_head_kv = 40
llm_load_print_meta: n_layer = 40
llm_load_print_meta: n_rot = 128
llm_load_print_meta: n_gqa = 1
llm_load_print_meta: f_norm_eps = 1.0e-05
llm_load_print_meta: f_norm_rms_eps = 1.0e-05
llm_load_print_meta: n_ff = 13824
llm_load_print_meta: freq_base = 10000.0
llm_load_print_meta: freq_scale = 1
llm_load_print_meta: model type = 13B
llm_load_print_meta: model ftype = mostly Q4_0
llm_load_print_meta: model size = 13.02 B
llm_load_print_meta: general.name = LLaMA v2
llm_load_print_meta: BOS token = 1 '<s>'
llm_load_print_meta: EOS token = 2 '</s>'
llm_load_print_meta: UNK token = 0 '<unk>'
llm_load_print_meta: LF token = 13 '<0x0A>'
llm_load_tensors: ggml ctx size = 0.11 MB
llm_load_tensors: mem required = 7024.01 MB (+ 400.00 MB per state)
...................................................................................................
llama_new_context_with_model: kv self size = 400.00 MB
llama_new_context_with_model: compute buffer total size = 75.41 MB
system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000
generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0
Building a website can be done in 10 simple steps:
Step 1: Find the right website platform.
Step 2: Choose your domain name and hosting plan.
Step 3: Design your website layout.
Step 4: Write your website content and add images.
Step 5: Install security features to protect your site from hackers or spammers
Step 6: Test your website on multiple browsers, mobile devices, operating systems etc…
Step 7: Test it again with people who are not related to you personally – friends or family members will work just fine!
Step 8: Start marketing and promoting the website via social media channels or paid ads
Step 9: Analyze how many visitors have come to your site so far, what type of people visit more often than others (e.g., men vs women) etc…
Step 10: Continue to improve upon all aspects mentioned above by following trends in web design and staying up-to-date on new technologies that can enhance user experience even further!
How does a Website Work?
A website works by having pages, which are made of HTML code. This code tells your computer how to display the content on each page you visit – whether it’s an image or text file (like PDFs). In order for someone else’s browser not only be able but also want those same results when accessing any given URL; some additional steps need taken by way of programming scripts that will add functionality such as making links clickable!
The most common type is called static HTML pages because they remain unchanged over time unless modified manually (either through editing files directly or using an interface such as WordPress). They are usually served up via HTTP protocols – this means anyone can access them without having any special privileges like being part of a group who is allowed into restricted areas online; however, there may still exist some limitations depending upon where one lives geographically speaking.
How to
llama_print_timings: load time = 576.45 ms
llama_print_timings: sample time = 283.10 ms / 400 runs ( 0.71 ms per token, 1412.91 tokens per second)
llama_print_timings: prompt eval time = 599.83 ms / 19 tokens ( 31.57 ms per token, 31.68 tokens per second)
llama_print_timings: eval time = 24513.59 ms / 399 runs ( 61.44 ms per token, 16.28 tokens per second)
llama_print_timings: total time = 25431.49 ms
```
</details>
<details>
<summary>Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook</summary>
And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4
</details>
## Usage
Here are the end-to-end binary build and model conversion steps for most supported models.
### Basic usage
Firstly, you need to get the binary. There are different methods that you can follow:
- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md)
- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md)
- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md)
- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases)
You can run a basic completion using this command:
```bash
llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128
# Output:
# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
```
See [this page](./examples/main/README.md) for a full list of parameters.
### Conversation mode
If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
```bash
llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
# Output:
# > hi, who are you?
# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
#
# > what is 1+1?
# Easy peasy! The answer to 1+1 is... 2!
```
By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
```bash
./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
```
You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
```bash
./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
```
### Web server
[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
Example usage:
```bash
./llama-server -m your_model.gguf --port 8080
# Basic web UI can be accessed via browser: http://localhost:8080
# Chat completion endpoint: http://localhost:8080/v1/chat/completions
```
### Interactive mode
> [!NOTE]
> If you prefer basic usage, please consider using conversation mode instead of interactive mode
In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
Here is an example of a few-shot interaction, invoked with the command
```bash
# default arguments using a 7B model
./examples/chat.sh
# advanced chat with a 13B model
./examples/chat-13B.sh
# custom arguments using a 13B model
./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt
```
Note the use of `--color` to distinguish between user input and generated text. Other parameters are explained in more detail in the [README](examples/main/README.md) for the `llama-cli` example program.
![image](https://user-images.githubusercontent.com/1991296/224575029-2af3c7dc-5a65-4f64-a6bb-517a532aea38.png)
### Persistent Interaction
The prompt, user inputs, and model generations can be saved and resumed across calls to `./llama-cli` by leveraging `--prompt-cache` and `--prompt-cache-all`. The `./examples/chat-persistent.sh` script demonstrates this with support for long-running, resumable chat sessions. To use this example, you must provide a file to cache the initial chat prompt and a directory to save the chat session, and may optionally provide the same variables as `chat-13B.sh`. The same prompt cache can be reused for new chat sessions. Note that both prompt cache and chat directory are tied to the initial prompt (`PROMPT_TEMPLATE`) and the model file.
```bash
# Start a new chat
PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
# Resume that chat
PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/default ./examples/chat-persistent.sh
# Start a different chat with the same prompt/model
PROMPT_CACHE_FILE=chat.prompt.bin CHAT_SAVE_DIR=./chat/another ./examples/chat-persistent.sh
# Different prompt cache for different prompt/model
PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \
CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh
```
### Constrained output with grammars
`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only:
```bash
./llama-cli -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
```
The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md).
For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
## Build
Please refer to [Build llama.cpp locally](./docs/build.md)
## Supported backends
| Backend | Target devices |
| --- | --- |
| [Metal](./docs/build.md#metal-build) | Apple Silicon |
| [BLAS](./docs/build.md#blas-build) | All |
| [BLIS](./docs/backend/BLIS.md) | All |
| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
| [Vulkan](./docs/build.md#vulkan) | GPU |
| [CANN](./docs/build.md#cann) | Ascend NPU |
| [Metal](docs/build.md#metal-build) | Apple Silicon |
| [BLAS](docs/build.md#blas-build) | All |
| [BLIS](docs/backend/BLIS.md) | All |
| [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 |
| [Vulkan](docs/build.md#vulkan) | GPU |
| [CANN](docs/build.md#cann) | Ascend NPU |
## Tools
## Building the project
### Prepare and Quantize
The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
> [!NOTE]
> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
- Clone this repository and build locally, see [how to build](docs/build.md)
- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
- Use a Docker image, see [documentation for Docker](docs/docker.md)
- Download pre-built binaries from [releases](https://github.com/ggerganov/llama.cpp/releases)
To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
## Obtaining and quantizing models
Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
- [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
### Perplexity (measuring model quality)
After downloading a model, use the CLI tools to run it locally - see below.
You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggerganov/llama.cpp/discussions/10123)
- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggerganov/llama.cpp/discussions/9268)
- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggerganov/llama.cpp/discussions/9669)
To learn more about model quantization, [read this documentation](examples/quantize/README.md)
## [`llama-cli`](examples/main)
#### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
- <details open>
<summary>Run simple text completion</summary>
```bash
llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128
# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
```
</details>
- <details>
<summary>Run in conversation mode</summary>
```bash
llama-cli -m model.gguf -p "You are a helpful assistant" -cnv
# > hi, who are you?
# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
#
# > what is 1+1?
# Easy peasy! The answer to 1+1 is... 2!
```
</details>
- <details>
<summary>Run with custom chat template</summary>
```bash
# use the "chatml" template
llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
# use a custom template
llama-cli -m model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
```
[Supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
</details>
- <details>
<summary>Constrain the output with a custom grammar</summary>
```bash
llama-cli -m model.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
# {"appointmentTime": "8pm", "appointmentDetails": "schedule a a call"}
```
The [grammars/](grammars/) folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](grammars/README.md).
For authoring more complex JSON grammars, check out https://grammar.intrinsiclabs.ai/
</details>
## [`llama-server`](examples/server)
#### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.
- <details open>
<summary>Start a local HTTP server with default configuration on port 8080</summary>
```bash
llama-server -m model.gguf --port 8080
# Basic web UI can be accessed via browser: http://localhost:8080
# Chat completion endpoint: http://localhost:8080/v1/chat/completions
```
</details>
- <details>
<summary>Support multiple-users and parallel decoding</summary>
```bash
# up to 4 concurrent requests, each with 4096 max context
llama-server -m model.gguf -c 16384 -np 4
```
</details>
- <details>
<summary>Enable speculative decoding</summary>
```bash
# the draft.gguf model should be a small variant of the target model.gguf
llama-server -m model.gguf -md draft.gguf
```
</details>
- <details>
<summary>Serve an embedding model</summary>
```bash
# use the /embedding endpoint
llama-server -m model.gguf --embedding --pooling cls -ub 8192
```
</details>
- <details>
<summary>Serve a reranking model</summary>
```bash
# use the /reranking endpoint
llama-server -m model.gguf --reranking
```
</details>
- <details>
<summary>Constrain all outputs with a grammar</summary>
```bash
# custom grammar
llama-server -m model.gguf --grammar-file grammar.gbnf
# JSON
llama-server -m model.gguf --grammar-file grammars/json.gbnf
```
</details>
## [`llama-perplexity`](examples/perplexity)
#### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.
- <details open>
<summary>Measure the perplexity over a text file</summary>
```bash
llama-perplexity -m model.gguf -f file.txt
# [1]15.2701,[2]5.4007,[3]5.3073,[4]6.2965,[5]5.8940,[6]5.6096,[7]5.7942,[8]4.9297, ...
# Final estimate: PPL = 5.4007 +/- 0.67339
```
</details>
- <details>
<summary>Measure KL divergence</summary>
```bash
# TODO
```
</details>
[^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)
#### Benchmark the performance of the inference for various parameters.
- <details open>
<summary>Run default benchmark</summary>
```bash
llama-bench -m model.gguf
# Output:
# | model | size | params | backend | threads | test | t/s |
# | ------------------- | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
# | qwen2 1.5B Q4_0 | 885.97 MiB | 1.54 B | Metal,BLAS | 16 | pp512 | 5765.41 ± 20.55 |
# | qwen2 1.5B Q4_0 | 885.97 MiB | 1.54 B | Metal,BLAS | 16 | tg128 | 197.71 ± 0.81 |
#
# build: 3e0ba0e60 (4229)
```
</details>
## [`llama-simple`](examples/simple)
#### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
- <details>
<summary>Basic text completion</summary>
```bash
llama-simple -m model.gguf
# Hello my name is Kaitlyn and I am a 16 year old girl. I am a junior in high school and I am currently taking a class called "The Art of
```
</details>
To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md)
## Contributing
- Contributors can open PRs
- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
- Collaborators will be invited based on contributions
- Any help with managing issues and PRs is very appreciated!
- Any help with managing issues, PRs and projects is very appreciated!
- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
## Other documentations
## Other documentation
- [main (cli)](./examples/main/README.md)
- [server](./examples/server/README.md)
- [jeopardy](./examples/jeopardy/README.md)
- [GBNF grammars](./grammars/README.md)
- [main (cli)](examples/main/README.md)
- [server](examples/server/README.md)
- [GBNF grammars](grammars/README.md)
**Development documentations**
#### Development documentation
- [How to build](./docs/build.md)
- [Running on Docker](./docs/docker.md)
- [Build on Android](./docs/android.md)
- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
- [How to build](docs/build.md)
- [Running on Docker](docs/docker.md)
- [Build on Android](docs/android.md)
- [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
**Seminal papers and background on the models**
#### Seminal papers and background on the models
If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
- LLaMA:
@ -486,3 +486,6 @@ If your issue is with model generation quality, then please at least scan the fo
- GPT-3.5 / InstructGPT / ChatGPT:
- [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
#### References

291
ci/run.sh
View file

@ -1,4 +1,4 @@
#/bin/bash
#!/bin/bash
#
# sample usage:
#
@ -13,6 +13,9 @@
# # with SYCL support
# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
# # with VULKAN support
# GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
#
if [ -z "$2" ]; then
echo "usage: $0 <output-dir> <mnt-dir>"
@ -36,11 +39,11 @@ SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
fi
if [ ! -z ${GG_BUILD_SYCL} ]; then
@ -50,7 +53,11 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
fi
if [ ! -z ${GG_BUILD_VULKAN} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
fi
## helpers
@ -107,7 +114,7 @@ function gg_run_ctest_debug {
gg_check_build_requirements
(time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
(time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
@ -138,7 +145,7 @@ function gg_run_ctest_release {
gg_check_build_requirements
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
if [ -z ${GG_BUILD_LOW_PERF} ]; then
(time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
@ -266,7 +273,6 @@ function gg_sum_ctest_with_model_release {
}
# open_llama_7b_v2
# requires: GG_BUILD_CUDA
function gg_run_open_llama_7b_v2 {
cd ${SRC}
@ -290,8 +296,8 @@ function gg_run_open_llama_7b_v2 {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
@ -320,36 +326,36 @@ function gg_run_open_llama_7b_v2 {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -425,7 +431,7 @@ function gg_run_pythia_1_4b {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
@ -454,34 +460,34 @@ function gg_run_pythia_1_4b {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --model ${model_f16} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-cli --model ${model_f16} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -535,7 +541,6 @@ function gg_sum_pythia_1_4b {
}
# pythia_2_8b
# requires: GG_BUILD_CUDA
function gg_run_pythia_2_8b {
cd ${SRC}
@ -556,8 +561,8 @@ function gg_run_pythia_2_8b {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
@ -586,36 +591,36 @@ function gg_run_pythia_2_8b {
./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
(time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-cli --model ${model_f16} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/llama-save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
(time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
@ -692,7 +697,7 @@ function gg_run_embd_bge_small {
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
@ -701,12 +706,88 @@ function gg_run_embd_bge_small {
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
set +e
}
function gg_sum_embd_bge_small {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'BGE Small (BERT):\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
}
# rerank_tiny
function gg_run_rerank_tiny {
cd ${SRC}
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
path_models="../models-mnt/rerank-tiny"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j$(nproc) ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
model_f16="${path_models}/ggml-model-f16.gguf"
# for this model, the SEP token is "</s>"
(time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
# sample output
# rerank score 0: 0.029
# rerank score 1: 0.029
# rerank score 2: 0.135
# check that the score is in the range [$3, $4]
function check_score {
qnt="$1"
score=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
if [ $(echo "$score < $3" | bc) -eq 1 ] || [ $(echo "$score > $4" | bc) -eq 1 ]; then
printf ' - %s @ %s (FAIL: score not in range [%s, %s])\n' "$qnt" "$score" "$3" "$4"
return 20
fi
printf ' - %s @ %s OK\n' "$qnt" "$score"
return 0
}
check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log
set +e
}
function gg_sum_rerank_tiny {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'Rerank Tiny (Jina):\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-rk-f16.log)"
}
function gg_check_build_requirements {
if ! command -v cmake &> /dev/null; then
gg_printf 'cmake not found, please install'
@ -721,17 +802,11 @@ function gg_check_build_requirements {
fi
}
function gg_sum_embd_bge_small {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'BGE Small (BERT):\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
}
## main
export LLAMA_LOG_PREFIX=1
export LLAMA_LOG_TIMESTAMPS=1
if [ -z ${GG_BUILD_LOW_PERF} ]; then
# Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
rm -rf ${SRC}/models-mnt
@ -740,7 +815,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
ln -sfn ${mnt_models} ${SRC}/models-mnt
# Create a fresh python3 venv and enter it
python3 -m venv "$MNT/venv"
if ! python3 -m venv "$MNT/venv"; then
echo "Error: Failed to create Python virtual environment at $MNT/venv."
exit 1
fi
source "$MNT/venv/bin/activate"
pip install -r ${SRC}/requirements.txt --disable-pip-version-check
@ -754,6 +832,7 @@ test $ret -eq 0 && gg_run ctest_release
if [ -z ${GG_BUILD_LOW_PERF} ]; then
test $ret -eq 0 && gg_run embd_bge_small
test $ret -eq 0 && gg_run rerank_tiny
if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
test $ret -eq 0 && gg_run test_scripts_debug
@ -761,7 +840,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
fi
if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
if [ -z ${GG_BUILD_CUDA} ]; then
if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
test $ret -eq 0 && gg_run pythia_1_4b
else
test $ret -eq 0 && gg_run pythia_2_8b

16
cmake/arm64-apple-clang.cmake Normal file
View file

@ -0,0 +1,16 @@
set( CMAKE_SYSTEM_NAME Darwin )
set( CMAKE_SYSTEM_PROCESSOR arm64 )
set( target arm64-apple-darwin-macho )
set( CMAKE_C_COMPILER clang )
set( CMAKE_CXX_COMPILER clang++ )
set( CMAKE_C_COMPILER_TARGET ${target} )
set( CMAKE_CXX_COMPILER_TARGET ${target} )
set( arch_c_flags "-march=armv8.4-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function" )
set( CMAKE_C_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )
set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

33
cmake/common.cmake Normal file
View file

@ -0,0 +1,33 @@
function(llama_add_compile_flags)
if (LLAMA_FATAL_WARNINGS)
if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
list(APPEND C_FLAGS -Werror)
list(APPEND CXX_FLAGS -Werror)
elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
add_compile_options(/WX)
endif()
endif()
if (LLAMA_ALL_WARNINGS)
if (NOT MSVC)
list(APPEND C_FLAGS -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-Werror=implicit-int -Werror=implicit-function-declaration)
list(APPEND CXX_FLAGS -Wmissing-declarations -Wmissing-noreturn)
list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
list(APPEND C_FLAGS ${WARNING_FLAGS})
list(APPEND CXX_FLAGS ${WARNING_FLAGS})
ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
"$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
else()
# todo : msvc
set(C_FLAGS "" PARENT_SCOPE)
set(CXX_FLAGS "" PARENT_SCOPE)
endif()
endif()
endfunction()

202
cmake/llama-config.cmake.in
View file

@ -3,18 +3,60 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(GGML_BLAS @GGML_BLAS@)
set(GGML_CUDA @GGML_CUDA@)
set(GGML_METAL @GGML_METAL@)
set(GGML_HIPBLAS @GGML_HIPBLAS@)
set(GGML_STATIC @GGML_STATIC@)
set(GGML_NATIVE @GGML_NATIVE@)
set(GGML_LTO @GGML_LTO@)
set(GGML_CCACHE @GGML_CCACHE@)
set(GGML_AVX @GGML_AVX@)
set(GGML_AVX2 @GGML_AVX2@)
set(GGML_AVX512 @GGML_AVX512@)
set(GGML_AVX512_VBMI @GGML_AVX512_VBMI@)
set(GGML_AVX512_VNNI @GGML_AVX512_VNNI@)
set(GGML_AVX512_BF16 @GGML_AVX512_BF16@)
set(GGML_AMX_TILE @GGML_AMX_TILE@)
set(GGML_AMX_INT8 @GGML_AMX_INT8@)
set(GGML_AMX_BF16 @GGML_AMX_BF16@)
set(GGML_FMA @GGML_FMA@)
set(GGML_LASX @GGML_LASX@)
set(GGML_LSX @GGML_LSX@)
set(GGML_RVV @GGML_RVV@)
set(GGML_SVE @GGML_SVE@)
set(GGML_ACCELERATE @GGML_ACCELERATE@)
set(GGML_VULKAN @GGML_VULKAN@)
set(GGML_OPENMP @GGML_OPENMP@)
set(GGML_CPU_HBM @GGML_CPU_HBM@)
set(GGML_BLAS_VENDOR @GGML_BLAS_VENDOR@)
set(GGML_CUDA_FORCE_MMQ @GGML_CUDA_FORCE_MMQ@)
set(GGML_CUDA_FORCE_CUBLAS @GGML_CUDA_FORCE_CUBLAS@)
set(GGML_CUDA_F16 @GGML_CUDA_F16@)
set(GGML_CUDA_PEER_MAX_BATCH_SIZE @GGML_CUDA_PEER_MAX_BATCH_SIZE@)
set(GGML_CUDA_NO_PEER_COPY @GGML_CUDA_NO_PEER_COPY@)
set(GGML_CUDA_NO_VMM @GGML_CUDA_NO_VMM@)
set(GGML_CUDA_FA_ALL_QUANTS @GGML_CUDA_FA_ALL_QUANTS@)
set(GGML_CUDA_GRAPHS @GGML_CUDA_GRAPHS@)
set(GGML_HIP_UMA @GGML_HIP_UMA@)
set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
set(GGML_SYCL @GGML_SYCL@)
set(GGML_OPENMP @GGML_OPENMP@)
set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
set(GGML_VULKAN_SHADER_DEBUG_INFO @GGML_VULKAN_SHADER_DEBUG_INFO@)
set(GGML_VULKAN_PERF @GGML_VULKAN_PERF@)
set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
set(GGML_VULKAN_RUN_TESTS @GGML_VULKAN_RUN_TESTS@)
set(GGML_METAL_USE_BF16 @GGML_METAL_USE_BF16@)
set(GGML_METAL_NDEBUG @GGML_METAL_NDEBUG@)
set(GGML_METAL_SHADER_DEBUG @GGML_METAL_SHADER_DEBUG@)
set(GGML_METAL_EMBED_LIBRARY @GGML_METAL_EMBED_LIBRARY@)
set(GGML_METAL_MACOSX_VERSION_MIN @GGML_METAL_MACOSX_VERSION_MIN@)
set(GGML_METAL_STD @GGML_METAL_STD@)
set(GGML_SYCL_F16 @GGML_SYCL_F16@)
set(GGML_SYCL_TARGET @GGML_SYCL_TARGET@)
set(GGML_SYCL_DEVICE_ARCH @GGML_SYCL_DEVICE_ARCH@)
@PACKAGE_INIT@
@ -22,65 +64,111 @@ set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
set_and_check(LLAMA_LIB_DIR "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
set_and_check(LLAMA_BIN_DIR "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
# Ensure transient dependencies satisfied
find_package(Threads REQUIRED)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
set(_llama_link_deps "")
set(_llama_link_opts "")
foreach(_ggml_lib ggml ggml-base)
string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
find_library(${_ggml_lib_var} ${_ggml_lib}
REQUIRED
HINTS ${LLAMA_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH
)
list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
message(STATUS "Found ${${_ggml_lib_var}}")
endforeach()
foreach(backend amx blas cann cpu cuda hip kompute metal musa rpc sycl vulkan)
string(TOUPPER "GGML_${backend}" backend_id)
set(_ggml_lib "ggml-${backend}")
string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
find_library(${_ggml_lib_var} ${_ggml_lib}
HINTS ${LLAMA_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH
)
if(${_ggml_lib_var})
list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
set(${backend_id} ON)
message(STATUS "Found backend ${${_ggml_lib_var}}")
else()
set(${backend_id} OFF)
endif()
endforeach()
if (NOT LLAMA_SHARED_LIB)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
list(APPEND _llama_link_deps ${ACCELERATE_FRAMEWORK})
endif()
if (GGML_OPENMP)
find_package(OpenMP REQUIRED)
list(APPEND _llama_link_deps OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
list(APPEND _llama_link_deps memkind)
endif()
if (GGML_BLAS)
find_package(BLAS REQUIRED)
list(APPEND _llama_link_deps ${BLAS_LIBRARIES})
list(APPEND _llama_link_opts ${BLAS_LINKER_FLAGS})
endif()
if (GGML_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
list(APPEND _llama_link_deps ${FOUNDATION_LIBRARY}
${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
endif()
if (GGML_VULKAN)
find_package(Vulkan REQUIRED)
list(APPEND _llama_link_deps Vulkan::Vulkan)
endif()
if (GGML_HIP)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
list(APPEND _llama_link_deps hip::host roc::rocblas roc::hipblas)
endif()
if (GGML_SYCL)
find_package(DNNL)
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND _llama_link_deps DNNL::dnnl)
endif()
if (WIN32)
find_package(IntelSYCL REQUIRED)
find_package(MKL REQUIRED)
list(APPEND _llama_link_deps IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
endif()
endif()
endif()
if (GGML_BLAS)
find_package(BLAS REQUIRED)
endif()
if (GGML_CUDA)
find_package(CUDAToolkit REQUIRED)
endif()
if (GGML_METAL)
find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
find_library(METAL_FRAMEWORK Metal REQUIRED)
find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
endif()
if (GGML_VULKAN)
find_package(Vulkan REQUIRED)
endif()
if (GGML_HIPBLAS)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
endif()
if (GGML_SYCL)
find_package(IntelSYCL REQUIRED)
find_package(MKL REQUIRED)
endif()
if (GGML_OPENMP)
find_package(OpenMP REQUIRED)
endif()
find_library(ggml_LIBRARY ggml
REQUIRED
HINTS ${LLAMA_LIB_DIR})
find_library(llama_LIBRARY llama
REQUIRED
HINTS ${LLAMA_LIB_DIR})
set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
HINTS ${LLAMA_LIB_DIR}
NO_CMAKE_FIND_ROOT_PATH
)
add_library(llama UNKNOWN IMPORTED)
set_target_properties(llama
PROPERTIES
INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
INTERFACE_LINK_OPTIONS "${_llama_link_opts}"
INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
IMPORTED_LOCATION "${llama_LIBRARY}"

26
common/CMakeLists.txt
View file

@ -2,6 +2,8 @@
find_package(Threads REQUIRED)
llama_add_compile_flags()
# Build info header
#
@ -51,21 +53,23 @@ endif()
set(TARGET common)
add_library(${TARGET} STATIC
arg.cpp
arg.h
base64.hpp
common.h
common.cpp
sampling.h
sampling.cpp
console.h
common.h
console.cpp
grammar-parser.h
grammar-parser.cpp
json.hpp
console.h
json-schema-to-grammar.cpp
train.h
train.cpp
ngram-cache.h
json.hpp
log.cpp
log.h
ngram-cache.cpp
ngram-cache.h
sampling.cpp
sampling.h
speculative.cpp
speculative.h
)
if (BUILD_SHARED_LIBS)
@ -84,5 +88,5 @@ if (LLAMA_CURL)
endif ()
target_include_directories(${TARGET} PUBLIC .)
target_compile_features (${TARGET} PUBLIC cxx_std_11)
target_compile_features (${TARGET} PUBLIC cxx_std_17)
target_link_libraries (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)

2137
common/arg.cpp Normal file
View file

File diff suppressed because it is too large Load diff

77
common/arg.h Normal file
View file

@ -0,0 +1,77 @@
#pragma once
#include "common.h"
#include <set>
#include <string>
#include <vector>
//
// CLI argument parsing
//
struct common_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
const char * env = nullptr;
std::string help;
bool is_sparam = false; // is current arg a sampling param?
void (*handler_void) (common_params & params) = nullptr;
void (*handler_string) (common_params & params, const std::string &) = nullptr;
void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
void (*handler_int) (common_params & params, int) = nullptr;
common_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(common_params & params, const std::string &)
) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
common_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(common_params & params, int)
) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
common_arg(
const std::initializer_list<const char *> & args,
const std::string & help,
void (*handler)(common_params & params)
) : args(args), help(help), handler_void(handler) {}
// support 2 values for arg
common_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const char * value_hint_2,
const std::string & help,
void (*handler)(common_params & params, const std::string &, const std::string &)
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
common_arg & set_examples(std::initializer_list<enum llama_example> examples);
common_arg & set_env(const char * env);
common_arg & set_sparam();
bool in_example(enum llama_example ex);
bool get_value_from_env(std::string & output);
bool has_value_from_env();
std::string to_string();
};
struct common_params_context {
enum llama_example ex = LLAMA_EXAMPLE_COMMON;
common_params & params;
std::vector<common_arg> options;
void(*print_usage)(int, char **) = nullptr;
common_params_context(common_params & params) : params(params) {}
};
// parse input arguments from CLI
// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
// function to be used by test-arg-parser
common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

2468
common/common.cpp
View file

File diff suppressed because it is too large Load diff

372
common/common.h
View file

@ -4,18 +4,9 @@
#include "llama.h"
#include "sampling.h"
#define LOG_NO_FILE_LINE_FUNCTION
#include "log.h"
#include <cmath>
#include <string>
#include <vector>
#include <random>
#include <thread>
#include <unordered_map>
#include <tuple>
#include <sstream>
#ifdef _WIN32
#define DIRECTORY_SEPARATOR '\\'
@ -33,61 +24,162 @@
#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
struct llama_lora_adapter_info {
struct common_lora_adapter_info {
std::string path;
float scale;
};
struct llama_lora_adapter_container : llama_lora_adapter_info {
struct common_lora_adapter_container : common_lora_adapter_info {
struct llama_lora_adapter * adapter;
};
using llama_tokens = std::vector<llama_token>;
// build info
extern int LLAMA_BUILD_NUMBER;
extern char const * LLAMA_COMMIT;
extern char const * LLAMA_COMPILER;
extern char const * LLAMA_BUILD_TARGET;
struct llama_control_vector_load_info;
struct common_control_vector_load_info;
//
// CPU utils
//
struct cpu_params {
int n_threads = -1;
bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
bool mask_valid = false; // Default: any CPU
enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
bool strict_cpu = false; // Use strict CPU placement
uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling)
};
int32_t cpu_get_num_physical_cores();
int32_t cpu_get_num_math();
//
// CLI argument parsing
// Common params
//
enum llama_example {
LLAMA_EXAMPLE_COMMON,
LLAMA_EXAMPLE_SPECULATIVE,
LLAMA_EXAMPLE_MAIN,
LLAMA_EXAMPLE_INFILL,
LLAMA_EXAMPLE_EMBEDDING,
LLAMA_EXAMPLE_PERPLEXITY,
LLAMA_EXAMPLE_RETRIEVAL,
LLAMA_EXAMPLE_PASSKEY,
LLAMA_EXAMPLE_IMATRIX,
LLAMA_EXAMPLE_BENCH,
LLAMA_EXAMPLE_SERVER,
LLAMA_EXAMPLE_CVECTOR_GENERATOR,
LLAMA_EXAMPLE_EXPORT_LORA,
LLAMA_EXAMPLE_LLAVA,
LLAMA_EXAMPLE_LOOKUP,
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_COUNT,
};
enum common_sampler_type {
COMMON_SAMPLER_TYPE_NONE = 0,
COMMON_SAMPLER_TYPE_DRY = 1,
COMMON_SAMPLER_TYPE_TOP_K = 2,
COMMON_SAMPLER_TYPE_TOP_P = 3,
COMMON_SAMPLER_TYPE_MIN_P = 4,
//COMMON_SAMPLER_TYPE_TFS_Z = 5,
COMMON_SAMPLER_TYPE_TYPICAL_P = 6,
COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
COMMON_SAMPLER_TYPE_XTC = 8,
COMMON_SAMPLER_TYPE_INFILL = 9,
};
// dimensionality reduction methods, used by cvector-generator
enum dimre_method {
DIMRE_METHOD_PCA,
DIMRE_METHOD_MEAN,
};
struct gpt_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed
// sampling parameters
struct common_params_sampling {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_threads = cpu_get_num_math();
int32_t n_threads_draft = -1;
int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads)
int32_t n_threads_batch_draft = -1;
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float xtc_probability = 0.00f; // 0.0 = disabled
float xtc_threshold = 0.10f; // > 0.5 disables XTC
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
float dry_multiplier = 0.0f; // 0.0 = disabled; DRY repetition penalty for tokens extending repetition:
float dry_base = 1.75f; // 0.0 = disabled; multiplier * base ^ (length of sequence before token - allowed length)
int32_t dry_allowed_length = 2; // tokens extending repetitions beyond this receive penalty
int32_t dry_penalty_last_n = -1; // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
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;
std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"}; // default sequence breakers for DRY
std::vector<enum common_sampler_type> samplers = {
COMMON_SAMPLER_TYPE_DRY,
COMMON_SAMPLER_TYPE_TOP_K,
COMMON_SAMPLER_TYPE_TYPICAL_P,
COMMON_SAMPLER_TYPE_TOP_P,
COMMON_SAMPLER_TYPE_MIN_P,
COMMON_SAMPLER_TYPE_XTC,
COMMON_SAMPLER_TYPE_TEMPERATURE,
};
std::string grammar; // optional BNF-like grammar to constrain sampling
std::vector<llama_logit_bias> logit_bias; // logit biases to apply
// print the parameters into a string
std::string print() const;
};
struct common_params_speculative {
std::vector<ggml_backend_dev_t> 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
int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
float p_split = 0.1f; // speculative decoding split probability
float p_min = 0.9f; // minimum speculative decoding probability (greedy)
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
std::string model = ""; // draft model for speculative decoding // NOLINT
};
struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_ctx = 4096; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_ubatch = 512; // physical batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 5; // number of tokens to draft during speculative decoding
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
float p_split = 0.1f; // speculative decoding split probability
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-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
int32_t grp_attn_n = 1; // group-attention factor
int32_t grp_attn_w = 512; // group-attention width
int32_t n_print = -1; // print token count every n tokens (-1 = disabled)
@ -98,47 +190,54 @@ struct gpt_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
float defrag_thold = -1.0f; // KV cache defragmentation threshold
float defrag_thold = 0.1f; // KV cache defragmentation threshold
// offload params
std::vector<ggml_backend_dev_t> 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;
ggml_backend_sched_eval_callback cb_eval = nullptr;
void * cb_eval_user_data = nullptr;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
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
// // sampling parameters
struct llama_sampling_params sparams;
struct common_params_sampling sampling;
struct common_params_speculative speculative;
std::string model = ""; // model path
std::string model_draft = ""; // draft model for speculative decoding
std::string model_alias = "unknown"; // model alias
std::string model_url = ""; // model url to download
std::string hf_token = ""; // HF token
std::string hf_repo = ""; // HF repo
std::string hf_file = ""; // HF file
std::string prompt = "";
std::string prompt_file = ""; // store the external prompt file name
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state
std::string input_prefix = ""; // string to prefix user inputs with
std::string input_suffix = ""; // string to suffix user inputs with
std::string logdir = ""; // directory in which to save YAML log files
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
std::string logits_file = ""; // file for saving *all* logits
std::string rpc_servers = ""; // comma separated list of RPC servers
std::string model = ""; // model path // NOLINT
std::string model_alias = "unknown"; // 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
std::string hf_file = ""; // HF file // NOLINT
std::string prompt = ""; // NOLINT
std::string prompt_file = ""; // store the external prompt file name // NOLINT
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT
std::string input_prefix = ""; // string to prefix user inputs with // NOLINT
std::string input_suffix = ""; // string to suffix user inputs with // NOLINT
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT
std::string logits_file = ""; // file for saving *all* logits // NOLINT
std::string rpc_servers = ""; // comma separated list of RPC servers // NOLINT
std::vector<std::string> in_files; // all input files
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
int32_t verbosity = 0;
int32_t control_vector_layer_start = -1; // layer range for control vector
@ -173,15 +272,15 @@ struct gpt_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = false; // flash attention
bool no_perf = false; // disable performance metrics
bool ctx_shift = true; // context shift on inifinite text generation
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool ignore_eos = false; // ignore generated EOS tokens
bool logits_all = false; // return logits for all tokens in the batch
bool use_mmap = true; // use mmap for faster loads
bool use_mlock = false; // use mlock to keep model in memory
bool verbose_prompt = false; // print prompt tokens before generation
bool display_prompt = true; // print prompt before generation
bool infill = false; // use infill mode
bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes
bool no_kv_offload = false; // disable KV offloading
bool warmup = true; // warmup run
@ -191,33 +290,37 @@ struct gpt_params {
std::string cache_type_v = "f16"; // KV cache data type for the V
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector
std::string mmproj = ""; // path to multimodal projector // NOLINT
std::vector<std::string> image; // path to image file(s)
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
std::string embd_sep = "\n"; // separator of embendings
std::string embd_sep = "\n"; // separator of embeddings
bool reranking = false; // enable reranking support on server
// server params
int32_t port = 8080; // server listens on this network port
int32_t timeout_read = 600; // http read timeout in seconds
int32_t timeout_write = timeout_read; // http write timeout in seconds
int32_t n_threads_http = -1; // number of threads to process HTTP requests
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
std::string hostname = "127.0.0.1";
std::string public_path = "";
std::string chat_template = "";
std::string system_prompt = "";
std::string public_path = ""; // NOLINT
std::string chat_template = ""; // NOLINT
bool enable_chat_template = true;
std::vector<std::string> api_keys;
std::string ssl_file_key = "";
std::string ssl_file_cert = "";
std::string ssl_file_key = ""; // NOLINT
std::string ssl_file_cert = ""; // NOLINT
bool endpoint_slots = true;
// "advanced" endpoints are disabled by default for better security
bool webui = true;
bool endpoint_slots = false;
bool endpoint_props = false; // only control POST requests, not GET
bool endpoint_metrics = false;
bool log_json = false;
@ -265,23 +368,38 @@ struct gpt_params {
bool spm_infill = false; // suffix/prefix/middle pattern for infill
std::string lora_outfile = "ggml-lora-merged-f16.gguf";
// batched-bench params
bool batched_bench_output_jsonl = false;
};
void gpt_params_parse_from_env(gpt_params & params);
void gpt_params_handle_model_default(gpt_params & params);
// call once at the start of a program if it uses libcommon
// initializes the logging system and prints info about the build
void common_init();
bool gpt_params_parse_ex (int argc, char ** argv, gpt_params & params);
bool gpt_params_parse (int argc, char ** argv, gpt_params & params);
bool gpt_params_find_arg (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
std::string common_params_get_system_info(const common_params & params);
std::string gpt_params_get_system_info(const gpt_params & params);
bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
bool set_process_priority(enum ggml_sched_priority prio);
//
// String utils
//
std::vector<std::string> string_split(std::string input, char separator);
#ifdef __GNUC__
#ifdef __MINGW32__
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
#endif
LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
std::string string_format(const char * fmt, ...);
std::string string_strip(const std::string & str);
std::string string_get_sortable_timestamp();
@ -290,6 +408,7 @@ void string_replace_all(std::string & s, const std::string & search, const std::
template<class T>
static std::vector<T> string_split(const std::string & str, char delim) {
static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
std::vector<T> values;
std::istringstream str_stream(str);
std::string token;
@ -302,9 +421,30 @@ static std::vector<T> string_split(const std::string & str, char delim) {
return values;
}
template<>
std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
{
std::vector<std::string> parts;
size_t begin_pos = 0;
size_t separator_pos = input.find(separator);
while (separator_pos != std::string::npos) {
std::string part = input.substr(begin_pos, separator_pos - begin_pos);
parts.emplace_back(part);
begin_pos = separator_pos + 1;
separator_pos = input.find(separator, begin_pos);
}
parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
return parts;
}
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
std::string string_from(bool value);
std::string string_from(const std::vector<int> & values);
std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
//
// Filesystem utils
//
@ -319,47 +459,69 @@ std::string fs_get_cache_file(const std::string & filename);
// Model utils
//
struct llama_init_result {
struct common_init_result {
struct llama_model * model = nullptr;
struct llama_context * context = nullptr;
std::vector<llama_lora_adapter_container> lora_adapters;
std::vector<common_lora_adapter_container> lora_adapters;
};
struct llama_init_result llama_init_from_gpt_params(gpt_params & params);
struct common_init_result common_init_from_params(common_params & params);
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params);
struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
struct llama_model_params common_model_params_to_llama ( common_params & params);
struct llama_context_params common_context_params_to_llama(const common_params & params);
struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
struct llama_model * common_load_model_from_url(
const std::string & model_url,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
struct llama_model * common_load_model_from_hf(
const std::string & repo,
const std::string & remote_path,
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
// clear LoRA adapters from context, then apply new list of adapters
void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
//
// Batch utils
//
void llama_batch_clear(struct llama_batch & batch);
void common_batch_clear(struct llama_batch & batch);
void llama_batch_add(
void common_batch_add(
struct llama_batch & batch,
llama_token id,
llama_pos pos,
const std::vector<llama_seq_id> & seq_ids,
bool logits);
//
// Token utils
//
// longest common prefix
size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
// longet common subsequence
size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
//
// Vocab utils
//
// tokenizes a string into a vector of tokens
// should work similar to Python's `tokenizer.encode`
std::vector<llama_token> llama_tokenize(
std::vector<llama_token> common_tokenize(
const struct llama_context * ctx,
const std::string & text,
bool add_special,
bool parse_special = false);
std::vector<llama_token> llama_tokenize(
std::vector<llama_token> common_tokenize(
const struct llama_model * model,
const std::string & text,
bool add_special,
@ -367,7 +529,7 @@ std::vector<llama_token> llama_tokenize(
// tokenizes a token into a piece, optionally renders special/control tokens
// should work similar to Python's `tokenizer.id_to_piece`
std::string llama_token_to_piece(
std::string common_token_to_piece(
const struct llama_context * ctx,
llama_token token,
bool special = true);
@ -375,7 +537,7 @@ std::string llama_token_to_piece(
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
// optionally renders special/control tokens
std::string llama_detokenize(
std::string common_detokenize(
llama_context * ctx,
const std::vector<llama_token> & tokens,
bool special = true);
@ -385,31 +547,31 @@ std::string llama_detokenize(
//
// same with llama_chat_message, but uses std::string
struct llama_chat_msg {
struct common_chat_msg {
std::string role;
std::string content;
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool llama_chat_verify_template(const std::string & tmpl);
bool common_chat_verify_template(const std::string & tmpl);
// CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml
// If the custom "tmpl" is not supported, we throw an error
std::string llama_chat_apply_template(const struct llama_model * model,
std::string common_chat_apply_template(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & chat,
const std::vector<common_chat_msg> & chat,
bool add_ass);
// Format single message, while taking into account the position of that message in chat history
std::string llama_chat_format_single(const struct llama_model * model,
std::string common_chat_format_single(const struct llama_model * model,
const std::string & tmpl,
const std::vector<llama_chat_msg> & past_msg,
const llama_chat_msg & new_msg,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass);
// Returns an example of formatted chat
std::string llama_chat_format_example(const struct llama_model * model,
std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl);
//
@ -417,31 +579,31 @@ std::string llama_chat_format_example(const struct llama_model * model,
//
// Dump the KV cache view with the number of sequences per cell.
void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
// Dump the KV cache view showing individual sequences in each cell (long output).
void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
//
// Embedding utils
//
void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
//
// Control vector utils
//
struct llama_control_vector_data {
struct common_control_vector_data {
int n_embd;
// stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
std::vector<float> data;
};
struct llama_control_vector_load_info {
struct common_control_vector_load_info {
float strength;
std::string fname;
@ -449,7 +611,7 @@ struct llama_control_vector_load_info {
// Load control vectors, scale each by strength, and add them together.
// On error, returns {-1, empty}
llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos);
//
// Split utils
@ -458,15 +620,3 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
//
// YAML utils
//
void yaml_dump_vector_float (FILE * stream, const char * prop_name, const std::vector<float> & data);
void yaml_dump_vector_int (FILE * stream, const char * prop_name, const std::vector<int> & data);
void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
void yaml_dump_non_result_info(
FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

3
common/console.cpp
View file

@ -94,6 +94,9 @@ namespace console {
simple_io = true;
}
}
if (simple_io) {
_setmode(_fileno(stdin), _O_U8TEXT);
}
#else
// POSIX-specific console initialization
if (!simple_io) {

539
common/grammar-parser.cpp
View file

@ -1,539 +0,0 @@
#include "grammar-parser.h"
#include <cstdint>
#include <cwchar>
#include <string>
#include <utility>
#include <stdexcept>
#include <exception>
namespace grammar_parser {
// NOTE: assumes valid utf8 (but checks for overrun)
// copied from llama.cpp
static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
uint8_t first_byte = static_cast<uint8_t>(*src);
uint8_t highbits = first_byte >> 4;
int len = lookup[highbits];
uint8_t mask = (1 << (8 - len)) - 1;
uint32_t value = first_byte & mask;
const char * end = src + len; // may overrun!
const char * pos = src + 1;
for ( ; pos < end && *pos; pos++) {
value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
}
return std::make_pair(value, pos);
}
static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
return result.first->second;
}
static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
return next_id;
}
static void add_rule(
parse_state & state,
uint32_t rule_id,
const std::vector<llama_grammar_element> & rule) {
if (state.rules.size() <= rule_id) {
state.rules.resize(rule_id + 1);
}
state.rules[rule_id] = rule;
}
static bool is_digit_char(char c) {
return '0' <= c && c <= '9';
}
static bool is_word_char(char c) {
return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
}
static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
const char * pos = src;
const char * end = src + size;
uint32_t value = 0;
for ( ; pos < end && *pos; pos++) {
value <<= 4;
char c = *pos;
if ('a' <= c && c <= 'f') {
value += c - 'a' + 10;
} else if ('A' <= c && c <= 'F') {
value += c - 'A' + 10;
} else if ('0' <= c && c <= '9') {
value += c - '0';
} else {
break;
}
}
if (pos != end) {
throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
}
return std::make_pair(value, pos);
}
static const char * parse_space(const char * src, bool newline_ok) {
const char * pos = src;
while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
(newline_ok && (*pos == '\r' || *pos == '\n'))) {
if (*pos == '#') {
while (*pos && *pos != '\r' && *pos != '\n') {
pos++;
}
} else {
pos++;
}
}
return pos;
}
static const char * parse_name(const char * src) {
const char * pos = src;
while (is_word_char(*pos)) {
pos++;
}
if (pos == src) {
throw std::runtime_error(std::string("expecting name at ") + src);
}
return pos;
}
static const char * parse_int(const char * src) {
const char * pos = src;
while (is_digit_char(*pos)) {
pos++;
}
if (pos == src) {
throw std::runtime_error(std::string("expecting integer at ") + src);
}
return pos;
}
static std::pair<uint32_t, const char *> parse_char(const char * src) {
if (*src == '\\') {
switch (src[1]) {
case 'x': return parse_hex(src + 2, 2);
case 'u': return parse_hex(src + 2, 4);
case 'U': return parse_hex(src + 2, 8);
case 't': return std::make_pair('\t', src + 2);
case 'r': return std::make_pair('\r', src + 2);
case 'n': return std::make_pair('\n', src + 2);
case '\\':
case '"':
case '[':
case ']':
return std::make_pair(src[1], src + 2);
default:
throw std::runtime_error(std::string("unknown escape at ") + src);
}
} else if (*src) {
return decode_utf8(src);
}
throw std::runtime_error("unexpected end of input");
}
const char * parse_alternates(
parse_state & state,
const char * src,
const std::string & rule_name,
uint32_t rule_id,
bool is_nested);
static const char * parse_sequence(
parse_state & state,
const char * src,
const std::string & rule_name,
std::vector<llama_grammar_element> & out_elements,
bool is_nested) {
size_t last_sym_start = out_elements.size();
const char * pos = src;
auto handle_repetitions = [&](int min_times, int max_times) {
if (last_sym_start == out_elements.size()) {
throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
}
// apply transformation to previous symbol (last_sym_start to end) according to
// the following rewrite rules:
// S{m,n} --> S S S (m times) S'(n-m)
// S'(x) ::= S S'(x-1) |
// (... n-m definitions of these S' rules ...)
// S'(1) ::= S |
// S{m,} --> S S S (m times) S'
// S' ::= S S' |
// S* --> S{0,}
// --> S' ::= S S' |
// S+ --> S{1,}
// --> S S'
// S' ::= S S' |
// S? --> S{0,1}
// --> S'
// S' ::= S |
std::vector<llama_grammar_element> previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
if (min_times == 0) {
out_elements.resize(last_sym_start);
} else {
// Repeat the previous elements (min_times - 1) times
for (int i = 1; i < min_times; i++) {
out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
}
}
uint32_t last_rec_rule_id = 0;
auto n_opt = max_times < 0 ? 1 : max_times - min_times;
std::vector<llama_grammar_element> rec_rule(previous_elements);
for (int i = 0; i < n_opt; i++) {
rec_rule.resize(previous_elements.size());
uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
if (i > 0 || max_times < 0) {
rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
}
rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
rec_rule.push_back({LLAMA_GRETYPE_END, 0});
add_rule(state, rec_rule_id, rec_rule);
last_rec_rule_id = rec_rule_id;
}
if (n_opt > 0) {
out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
}
};
while (*pos) {
if (*pos == '"') { // literal string
pos++;
last_sym_start = out_elements.size();
while (*pos != '"') {
if (!*pos) {
throw std::runtime_error("unexpected end of input");
}
auto char_pair = parse_char(pos);
pos = char_pair.second;
out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
}
pos = parse_space(pos + 1, is_nested);
} else if (*pos == '[') { // char range(s)
pos++;
enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
if (*pos == '^') {
pos++;
start_type = LLAMA_GRETYPE_CHAR_NOT;
}
last_sym_start = out_elements.size();
while (*pos != ']') {
if (!*pos) {
throw std::runtime_error("unexpected end of input");
}
auto char_pair = parse_char(pos);
pos = char_pair.second;
enum llama_gretype type = last_sym_start < out_elements.size()
? LLAMA_GRETYPE_CHAR_ALT
: start_type;
out_elements.push_back({type, char_pair.first});
if (pos[0] == '-' && pos[1] != ']') {
if (!pos[1]) {
throw std::runtime_error("unexpected end of input");
}
auto endchar_pair = parse_char(pos + 1);
pos = endchar_pair.second;
out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
}
}
pos = parse_space(pos + 1, is_nested);
} else if (is_word_char(*pos)) { // rule reference
const char * name_end = parse_name(pos);
uint32_t ref_rule_id = get_symbol_id(state, pos, name_end - pos);
pos = parse_space(name_end, is_nested);
last_sym_start = out_elements.size();
out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
} else if (*pos == '(') { // grouping
// parse nested alternates into synthesized rule
pos = parse_space(pos + 1, true);
uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
last_sym_start = out_elements.size();
// output reference to synthesized rule
out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
if (*pos != ')') {
throw std::runtime_error(std::string("expecting ')' at ") + pos);
}
pos = parse_space(pos + 1, is_nested);
} else if (*pos == '.') { // any char
last_sym_start = out_elements.size();
out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
pos = parse_space(pos + 1, is_nested);
} else if (*pos == '*') {
pos = parse_space(pos + 1, is_nested);
handle_repetitions(0, -1);
} else if (*pos == '+') {
pos = parse_space(pos + 1, is_nested);
handle_repetitions(1, -1);
} else if (*pos == '?') {
pos = parse_space(pos + 1, is_nested);
handle_repetitions(0, 1);
} else if (*pos == '{') {
pos = parse_space(pos + 1, is_nested);
if (!is_digit_char(*pos)) {
throw std::runtime_error(std::string("expecting an int at ") + pos);
}
const char * int_end = parse_int(pos);
int min_times = std::stoul(std::string(pos, int_end - pos));
pos = parse_space(int_end, is_nested);
int max_times = -1;
if (*pos == '}') {
max_times = min_times;
pos = parse_space(pos + 1, is_nested);
} else if (*pos == ',') {
pos = parse_space(pos + 1, is_nested);
if (is_digit_char(*pos)) {
const char * int_end = parse_int(pos);
max_times = std::stoul(std::string(pos, int_end - pos));
pos = parse_space(int_end, is_nested);
}
if (*pos != '}') {
throw std::runtime_error(std::string("expecting '}' at ") + pos);
}
pos = parse_space(pos + 1, is_nested);
} else {
throw std::runtime_error(std::string("expecting ',' at ") + pos);
}
handle_repetitions(min_times, max_times);
} else {
break;
}
}
return pos;
}
const char * parse_alternates(
parse_state & state,
const char * src,
const std::string & rule_name,
uint32_t rule_id,
bool is_nested) {
std::vector<llama_grammar_element> rule;
const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
while (*pos == '|') {
rule.push_back({LLAMA_GRETYPE_ALT, 0});
pos = parse_space(pos + 1, true);
pos = parse_sequence(state, pos, rule_name, rule, is_nested);
}
rule.push_back({LLAMA_GRETYPE_END, 0});
add_rule(state, rule_id, rule);
return pos;
}
static const char * parse_rule(parse_state & state, const char * src) {
const char * name_end = parse_name(src);
const char * pos = parse_space(name_end, false);
size_t name_len = name_end - src;
uint32_t rule_id = get_symbol_id(state, src, name_len);
const std::string name(src, name_len);
if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
throw std::runtime_error(std::string("expecting ::= at ") + pos);
}
pos = parse_space(pos + 3, true);
pos = parse_alternates(state, pos, name, rule_id, false);
if (*pos == '\r') {
pos += pos[1] == '\n' ? 2 : 1;
} else if (*pos == '\n') {
pos++;
} else if (*pos) {
throw std::runtime_error(std::string("expecting newline or end at ") + pos);
}
return parse_space(pos, true);
}
parse_state parse(const char * src) {
try {
parse_state state;
const char * pos = parse_space(src, true);
while (*pos) {
pos = parse_rule(state, pos);
}
// Validate the state to ensure that all rules are defined
for (const auto & rule : state.rules) {
if (rule.empty()) {
throw std::runtime_error("Undefined rule");
}
for (const auto & elem : rule) {
if (elem.type == LLAMA_GRETYPE_RULE_REF) {
// Ensure that the rule at that location exists
if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
// Get the name of the rule that is missing
for (const auto & kv : state.symbol_ids) {
if (kv.second == elem.value) {
throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
}
}
}
}
}
}
return state;
} catch (const std::exception & err) {
fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
return parse_state();
}
}
static void print_grammar_char(FILE * file, uint32_t c) {
if (0x20 <= c && c <= 0x7f) {
fprintf(file, "%c", static_cast<char>(c));
} else {
// cop out of encoding UTF-8
fprintf(file, "<U+%04X>", c);
}
}
static bool is_char_element(llama_grammar_element elem) {
switch (elem.type) {
case LLAMA_GRETYPE_CHAR: return true;
case LLAMA_GRETYPE_CHAR_NOT: return true;
case LLAMA_GRETYPE_CHAR_ALT: return true;
case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
case LLAMA_GRETYPE_CHAR_ANY: return true;
default: return false;
}
}
static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
for (auto elem : rule) {
switch (elem.type) {
case LLAMA_GRETYPE_END: fprintf(file, "END"); break;
case LLAMA_GRETYPE_ALT: fprintf(file, "ALT"); break;
case LLAMA_GRETYPE_RULE_REF: fprintf(file, "RULE_REF"); break;
case LLAMA_GRETYPE_CHAR: fprintf(file, "CHAR"); break;
case LLAMA_GRETYPE_CHAR_NOT: fprintf(file, "CHAR_NOT"); break;
case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
case LLAMA_GRETYPE_CHAR_ALT: fprintf(file, "CHAR_ALT"); break;
case LLAMA_GRETYPE_CHAR_ANY: fprintf(file, "CHAR_ANY"); break;
}
switch (elem.type) {
case LLAMA_GRETYPE_END:
case LLAMA_GRETYPE_ALT:
case LLAMA_GRETYPE_RULE_REF:
fprintf(file, "(%u) ", elem.value);
break;
case LLAMA_GRETYPE_CHAR:
case LLAMA_GRETYPE_CHAR_NOT:
case LLAMA_GRETYPE_CHAR_RNG_UPPER:
case LLAMA_GRETYPE_CHAR_ALT:
case LLAMA_GRETYPE_CHAR_ANY:
fprintf(file, "(\"");
print_grammar_char(file, elem.value);
fprintf(file, "\") ");
break;
}
}
fprintf(file, "\n");
}
static void print_rule(
FILE * file,
uint32_t rule_id,
const std::vector<llama_grammar_element> & rule,
const std::map<uint32_t, std::string> & symbol_id_names) {
if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
throw std::runtime_error(
"malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
}
fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
llama_grammar_element elem = rule[i];
switch (elem.type) {
case LLAMA_GRETYPE_END:
throw std::runtime_error(
"unexpected end of rule: " + std::to_string(rule_id) + "," +
std::to_string(i));
case LLAMA_GRETYPE_ALT:
fprintf(file, "| ");
break;
case LLAMA_GRETYPE_RULE_REF:
fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
break;
case LLAMA_GRETYPE_CHAR:
fprintf(file, "[");
print_grammar_char(file, elem.value);
break;
case LLAMA_GRETYPE_CHAR_NOT:
fprintf(file, "[^");
print_grammar_char(file, elem.value);
break;
case LLAMA_GRETYPE_CHAR_RNG_UPPER:
if (i == 0 || !is_char_element(rule[i - 1])) {
throw std::runtime_error(
"LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
std::to_string(rule_id) + "," + std::to_string(i));
}
fprintf(file, "-");
print_grammar_char(file, elem.value);
break;
case LLAMA_GRETYPE_CHAR_ALT:
if (i == 0 || !is_char_element(rule[i - 1])) {
throw std::runtime_error(
"LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
std::to_string(rule_id) + "," + std::to_string(i));
}
print_grammar_char(file, elem.value);
break;
case LLAMA_GRETYPE_CHAR_ANY:
fprintf(file, ".");
break;
}
if (is_char_element(elem)) {
switch (rule[i + 1].type) {
case LLAMA_GRETYPE_CHAR_ALT:
case LLAMA_GRETYPE_CHAR_RNG_UPPER:
case LLAMA_GRETYPE_CHAR_ANY:
break;
default:
fprintf(file, "] ");
}
}
}
fprintf(file, "\n");
}
void print_grammar(FILE * file, const parse_state & state) {
try {
std::map<uint32_t, std::string> symbol_id_names;
for (const auto & kv : state.symbol_ids) {
symbol_id_names[kv.second] = kv.first;
}
for (size_t i = 0, end = state.rules.size(); i < end; i++) {
// fprintf(file, "%zu: ", i);
// print_rule_binary(file, state.rules[i]);
print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
// fprintf(file, "\n");
}
} catch (const std::exception & err) {
fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
}
}
std::vector<const llama_grammar_element *> parse_state::c_rules() {
std::vector<const llama_grammar_element *> ret;
ret.reserve(rules.size());
for (const auto & rule : rules) {
ret.push_back(rule.data());
}
return ret;
}
}

29
common/grammar-parser.h
View file

@ -1,29 +0,0 @@
// Implements a parser for an extended Backus-Naur form (BNF), producing the
// binary context-free grammar format specified by llama.h. Supports character
// ranges, grouping, and repetition operators. As an example, a grammar for
// arithmetic might look like:
//
// root ::= expr
// expr ::= term ([-+*/] term)*
// term ::= num | "(" space expr ")" space
// num ::= [0-9]+ space
// space ::= [ \t\n]*
#pragma once
#include "llama.h"
#include <vector>
#include <map>
#include <cstdint>
#include <string>
namespace grammar_parser {
struct parse_state {
std::map<std::string, uint32_t> symbol_ids;
std::vector<std::vector<llama_grammar_element>> rules;
std::vector<const llama_grammar_element *> c_rules();
};
parse_state parse(const char * src);
void print_grammar(FILE * file, const parse_state & state);
}

2
common/json-schema-to-grammar.cpp
View file

@ -611,7 +611,7 @@ private:
}
return join_seq();
};
return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space");
}
/*

401
common/log.cpp Normal file
View file

@ -0,0 +1,401 @@
#include "log.h"
#include <condition_variable>
#include <cstdarg>
#include <cstdio>
#include <mutex>
#include <sstream>
#include <thread>
#include <vector>
int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
void common_log_set_verbosity_thold(int verbosity) {
common_log_verbosity_thold = verbosity;
}
#define LOG_COL_DEFAULT "\033[0m"
#define LOG_COL_BOLD "\033[1m"
#define LOG_COL_RED "\033[31m"
#define LOG_COL_GREEN "\033[32m"
#define LOG_COL_YELLOW "\033[33m"
#define LOG_COL_BLUE "\033[34m"
#define LOG_COL_MAGENTA "\033[35m"
#define LOG_COL_CYAN "\033[36m"
#define LOG_COL_WHITE "\033[37m"
static int64_t t_us() {
return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
}
// colors
enum common_log_col : int {
COMMON_LOG_COL_DEFAULT = 0,
COMMON_LOG_COL_BOLD,
COMMON_LOG_COL_RED,
COMMON_LOG_COL_GREEN,
COMMON_LOG_COL_YELLOW,
COMMON_LOG_COL_BLUE,
COMMON_LOG_COL_MAGENTA,
COMMON_LOG_COL_CYAN,
COMMON_LOG_COL_WHITE,
};
// disable colors by default
static std::vector<const char *> g_col = {
"",
"",
"",
"",
"",
"",
"",
"",
"",
};
struct common_log_entry {
enum ggml_log_level level;
bool prefix;
int64_t timestamp;
std::vector<char> msg;
// signals the worker thread to stop
bool is_end;
void print(FILE * file = nullptr) const {
FILE * fcur = file;
if (!fcur) {
// stderr displays DBG messages only when their verbosity level is not higher than the threshold
// these messages will still be logged to a file
if (level == GGML_LOG_LEVEL_DEBUG && common_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
return;
}
fcur = stdout;
if (level != GGML_LOG_LEVEL_NONE) {
fcur = stderr;
}
}
if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
if (timestamp) {
// [M.s.ms.us]
fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
g_col[COMMON_LOG_COL_BLUE],
(int) (timestamp / 1000000 / 60),
(int) (timestamp / 1000000 % 60),
(int) (timestamp / 1000 % 1000),
(int) (timestamp % 1000),
g_col[COMMON_LOG_COL_DEFAULT]);
}
switch (level) {
case GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[COMMON_LOG_COL_GREEN], g_col[COMMON_LOG_COL_DEFAULT]); break;
case GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[COMMON_LOG_COL_MAGENTA], "" ); break;
case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[COMMON_LOG_COL_RED], "" ); break;
case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[COMMON_LOG_COL_YELLOW], "" ); break;
default:
break;
}
}
fprintf(fcur, "%s", msg.data());
if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
fprintf(fcur, "%s", g_col[COMMON_LOG_COL_DEFAULT]);
}
fflush(fcur);
}
};
struct common_log {
// default capacity - will be expanded if needed
common_log() : common_log(256) {}
common_log(size_t capacity) {
file = nullptr;
prefix = false;
timestamps = false;
running = false;
t_start = t_us();
// initial message size - will be expanded if longer messages arrive
entries.resize(capacity);
for (auto & entry : entries) {
entry.msg.resize(256);
}
head = 0;
tail = 0;
resume();
}
~common_log() {
pause();
if (file) {
fclose(file);
}
}
private:
std::mutex mtx;
std::thread thrd;
std::condition_variable cv;
FILE * file;
bool prefix;
bool timestamps;
bool running;
int64_t t_start;
// ring buffer of entries
std::vector<common_log_entry> entries;
size_t head;
size_t tail;
// worker thread copies into this
common_log_entry cur;
public:
void add(enum ggml_log_level level, const char * fmt, va_list args) {
std::lock_guard<std::mutex> lock(mtx);
if (!running) {
// discard messages while the worker thread is paused
return;
}
auto & entry = entries[tail];
{
// cannot use args twice, so make a copy in case we need to expand the buffer
va_list args_copy;
va_copy(args_copy, args);
#if 1
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args_copy);
}
#else
// hack for bolding arguments
std::stringstream ss;
for (int i = 0; fmt[i] != 0; i++) {
if (fmt[i] == '%') {
ss << LOG_COL_BOLD;
while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
ss << LOG_COL_DEFAULT;
if (fmt[i] == 0) break;
}
ss << fmt[i];
}
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
}
#endif
}
entry.level = level;
entry.prefix = prefix;
entry.timestamp = 0;
if (timestamps) {
entry.timestamp = t_us() - t_start;
}
entry.is_end = false;
tail = (tail + 1) % entries.size();
if (tail == head) {
// expand the buffer
std::vector<common_log_entry> new_entries(2*entries.size());
size_t new_tail = 0;
do {
new_entries[new_tail] = std::move(entries[head]);
head = (head + 1) % entries.size();
new_tail = (new_tail + 1);
} while (head != tail);
head = 0;
tail = new_tail;
for (size_t i = tail; i < new_entries.size(); i++) {
new_entries[i].msg.resize(256);
}
entries = std::move(new_entries);
}
cv.notify_one();
}
void resume() {
std::lock_guard<std::mutex> lock(mtx);
if (running) {
return;
}
running = true;
thrd = std::thread([this]() {
while (true) {
{
std::unique_lock<std::mutex> lock(mtx);
cv.wait(lock, [this]() { return head != tail; });
cur = entries[head];
head = (head + 1) % entries.size();
}
if (cur.is_end) {
break;
}
cur.print(); // stdout and stderr
if (file) {
cur.print(file);
}
}
});
}
void pause() {
{
std::lock_guard<std::mutex> lock(mtx);
if (!running) {
return;
}
running = false;
// push an entry to signal the worker thread to stop
{
auto & entry = entries[tail];
entry.is_end = true;
tail = (tail + 1) % entries.size();
}
cv.notify_one();
}
thrd.join();
}
void set_file(const char * path) {
pause();
if (file) {
fclose(file);
}
if (path) {
file = fopen(path, "w");
} else {
file = nullptr;
}
resume();
}
void set_colors(bool colors) {
pause();
if (colors) {
g_col[COMMON_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
g_col[COMMON_LOG_COL_BOLD] = LOG_COL_BOLD;
g_col[COMMON_LOG_COL_RED] = LOG_COL_RED;
g_col[COMMON_LOG_COL_GREEN] = LOG_COL_GREEN;
g_col[COMMON_LOG_COL_YELLOW] = LOG_COL_YELLOW;
g_col[COMMON_LOG_COL_BLUE] = LOG_COL_BLUE;
g_col[COMMON_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
g_col[COMMON_LOG_COL_CYAN] = LOG_COL_CYAN;
g_col[COMMON_LOG_COL_WHITE] = LOG_COL_WHITE;
} else {
for (size_t i = 0; i < g_col.size(); i++) {
g_col[i] = "";
}
}
resume();
}
void set_prefix(bool prefix) {
std::lock_guard<std::mutex> lock(mtx);
this->prefix = prefix;
}
void set_timestamps(bool timestamps) {
std::lock_guard<std::mutex> lock(mtx);
this->timestamps = timestamps;
}
};
//
// public API
//
struct common_log * common_log_init() {
return new common_log;
}
struct common_log * common_log_main() {
static struct common_log log;
return &log;
}
void common_log_pause(struct common_log * log) {
log->pause();
}
void common_log_resume(struct common_log * log) {
log->resume();
}
void common_log_free(struct common_log * log) {
delete log;
}
void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...) {
va_list args;
va_start(args, fmt);
log->add(level, fmt, args);
va_end(args);
}
void common_log_set_file(struct common_log * log, const char * file) {
log->set_file(file);
}
void common_log_set_colors(struct common_log * log, bool colors) {
log->set_colors(colors);
}
void common_log_set_prefix(struct common_log * log, bool prefix) {
log->set_prefix(prefix);
}
void common_log_set_timestamps(struct common_log * log, bool timestamps) {
log->set_timestamps(timestamps);
}

788
common/log.h
View file

@ -1,724 +1,92 @@
#pragma once
#include <chrono>
#include <cstring>
#include <sstream>
#include <iostream>
#include <thread>
#include <vector>
#include <algorithm>
#include <cinttypes>
#include "ggml.h" // for ggml_log_level
// --------------------------------
//
// Basic usage:
//
// --------
//
// The LOG() and LOG_TEE() macros are ready to go by default
// they do not require any initialization.
//
// LOGLN() and LOG_TEELN() are variants which automatically
// include \n character at the end of the log string.
//
// LOG() behaves exactly like printf, by default writing to a logfile.
// LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
//
// Default logfile is named
// "llama.<threadID>.log"
// Default LOG_TEE() secondary output target is
// stderr
//
// Logs can be dynamically disabled or enabled using functions:
// log_disable()
// and
// log_enable()
//
// A log target can be changed with:
// log_set_target( string )
// creating and opening, or re-opening a file by string filename
// or
// log_set_target( FILE* )
// allowing to point at stderr, stdout, or any valid FILE* file handler.
//
// --------
//
// End of Basic usage.
//
// --------------------------------
// Specifies a log target.
// default uses log_handler() with "llama.log" log file
// this can be changed, by defining LOG_TARGET
// like so:
//
// #define LOG_TARGET (a valid FILE*)
// #include "log.h"
//
// or it can be simply redirected to stdout or stderr
// like so:
//
// #define LOG_TARGET stderr
// #include "log.h"
//
// The log target can also be redirected to a different function
// like so:
//
// #define LOG_TARGET log_handler_different()
// #include "log.h"
//
// FILE* log_handler_different()
// {
// return stderr;
// }
//
// or:
//
// #define LOG_TARGET log_handler_another_one("somelog.log")
// #include "log.h"
//
// FILE* log_handler_another_one(char*filename)
// {
// static FILE* logfile = nullptr;
// (...)
// if( !logfile )
// {
// fopen(...)
// }
// (...)
// return logfile
// }
//
#ifndef LOG_TARGET
#define LOG_TARGET log_handler()
#endif
#ifndef LOG_TEE_TARGET
#define LOG_TEE_TARGET stderr
#endif
// Utility for synchronizing log configuration state
// since std::optional was introduced only in c++17
enum LogTriState
{
LogTriStateSame,
LogTriStateFalse,
LogTriStateTrue
};
// Utility to obtain "pid" like unique process id and use it when creating log files.
inline std::string log_get_pid()
{
static std::string pid;
if (pid.empty())
{
// std::this_thread::get_id() is the most portable way of obtaining a "process id"
// it's not the same as "pid" but is unique enough to solve multiple instances
// trying to write to the same log.
std::stringstream ss;
ss << std::this_thread::get_id();
pid = ss.str();
}
return pid;
}
// Utility function for generating log file names with unique id based on thread id.
// invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
// where the number is a runtime id of the current thread.
#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
// INTERNAL, DO NOT USE
inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
{
static bool _multilog = false;
if (multilog != LogTriStateSame)
{
_multilog = multilog == LogTriStateTrue;
}
std::stringstream buf;
buf << log_file_basename;
if (_multilog)
{
buf << ".";
buf << log_get_pid();
}
buf << ".";
buf << log_file_extension;
return buf.str();
}
#ifndef LOG_DEFAULT_FILE_NAME
#define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
#endif
// Utility for turning #define values into string literals
// so we can have a define for stderr and
// we can print "stderr" instead of literal stderr, etc.
#define LOG_STRINGIZE1(s) #s
#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
// Allows disabling timestamps.
// in order to disable, define LOG_NO_TIMESTAMPS
// like so:
//
// #define LOG_NO_TIMESTAMPS
// #include "log.h"
//
#ifndef LOG_NO_TIMESTAMPS
#ifndef _MSC_VER
#define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#else
#define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#endif
#ifndef __GNUC__
# define LOG_ATTRIBUTE_FORMAT(...)
#elif defined(__MINGW32__)
# define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LOG_TIMESTAMP_FMT "%s"
#define LOG_TIMESTAMP_VAL ,""
# define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#ifdef LOG_TEE_TIMESTAMPS
#ifndef _MSC_VER
#define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#else
#define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#endif
#else
#define LOG_TEE_TIMESTAMP_FMT "%s"
#define LOG_TEE_TIMESTAMP_VAL ,""
#endif
#define LOG_DEFAULT_DEBUG 1
#define LOG_DEFAULT_LLAMA 0
// Allows disabling file/line/function prefix
// in order to disable, define LOG_NO_FILE_LINE_FUNCTION
// like so:
// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
// set via common_log_set_verbosity()
extern int common_log_verbosity_thold;
void common_log_set_verbosity_thold(int verbosity); // not thread-safe
// the common_log uses an internal worker thread to print/write log messages
// when the worker thread is paused, incoming log messages are discarded
struct common_log;
struct common_log * common_log_init();
struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
void common_log_pause (struct common_log * log); // pause the worker thread, not thread-safe
void common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
void common_log_free (struct common_log * log);
LOG_ATTRIBUTE_FORMAT(3, 4)
void common_log_add(struct common_log * log, enum ggml_log_level level, const char * fmt, ...);
// defaults: file = NULL, colors = false, prefix = false, timestamps = false
//
// #define LOG_NO_FILE_LINE_FUNCTION
// #include "log.h"
// regular log output:
//
#ifndef LOG_NO_FILE_LINE_FUNCTION
#ifndef _MSC_VER
#define LOG_FLF_FMT "[%24s:%5d][%24s] "
#define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
#else
#define LOG_FLF_FMT "[%24s:%5ld][%24s] "
#define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
#endif
#else
#define LOG_FLF_FMT "%s"
#define LOG_FLF_VAL ,""
#endif
#ifdef LOG_TEE_FILE_LINE_FUNCTION
#ifndef _MSC_VER
#define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
#define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
#else
#define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
#define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
#endif
#else
#define LOG_TEE_FLF_FMT "%s"
#define LOG_TEE_FLF_VAL ,""
#endif
// INTERNAL, DO NOT USE
// USE LOG() INSTEAD
// ggml_backend_metal_log_allocated_size: allocated buffer, size = 6695.84 MiB, ( 6695.91 / 21845.34)
// llm_load_tensors: ggml ctx size = 0.27 MiB
// llm_load_tensors: offloading 32 repeating layers to GPU
// llm_load_tensors: offloading non-repeating layers to GPU
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TARGET); \
} \
// with prefix = true, timestamps = true, the log output will look like this:
//
// 0.00.035.060 D ggml_backend_metal_log_allocated_size: allocated buffer, size = 6695.84 MiB, ( 6695.91 / 21845.34)
// 0.00.035.064 I llm_load_tensors: ggml ctx size = 0.27 MiB
// 0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
// 0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
//
// I - info (stdout, V = 0)
// W - warning (stderr, V = 0)
// E - error (stderr, V = 0)
// D - debug (stderr, V = LOG_DEFAULT_DEBUG)
//
void common_log_set_file (struct common_log * log, const char * file); // not thread-safe
void common_log_set_colors (struct common_log * log, bool colors); // not thread-safe
void common_log_set_prefix (struct common_log * log, bool prefix); // whether to output prefix to each log
void common_log_set_timestamps(struct common_log * log, bool timestamps); // whether to output timestamps in the prefix
// helper macros for logging
// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
//
// for example:
//
// LOG_DBG("this is a debug message: %d\n", expensive_function());
//
// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > common_log_verbosity_thold
//
#define LOG_TMPL(level, verbosity, ...) \
do { \
if ((verbosity) <= common_log_verbosity_thold) { \
common_log_add(common_log_main(), (level), __VA_ARGS__); \
} \
} while (0)
#else
#define LOG_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TARGET); \
} \
} while (0)
#endif
// INTERNAL, DO NOT USE
// USE LOG_TEE() INSTEAD
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_TEE_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TARGET); \
} \
if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \
{ \
fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TEE_TARGET); \
} \
} while (0)
#else
#define LOG_TEE_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TARGET); \
} \
if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \
{ \
fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TEE_TARGET); \
} \
} while (0)
#endif
#define LOG(...) LOG_TMPL(GGML_LOG_LEVEL_NONE, 0, __VA_ARGS__)
#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
// The '\0' as a last argument, is a trick to bypass the silly
// "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
// so we can have a single macro which can be called just like printf.
#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, 0, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, 0, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT, 0, __VA_ARGS__)
// Main LOG macro.
// behaves like printf, and supports arguments the exact same way.
//
#if !defined(_MSC_VER) || defined(__clang__)
#define LOG(...) LOG_IMPL(__VA_ARGS__, "")
#else
#define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
#endif
// Main TEE macro.
// does the same as LOG
// and
// simultaneously writes stderr.
//
// Secondary target can be changed just like LOG_TARGET
// by defining LOG_TEE_TARGET
//
#if !defined(_MSC_VER) || defined(__clang__)
#define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
#else
#define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
#endif
// LOG macro variants with auto endline.
#if !defined(_MSC_VER) || defined(__clang__)
#define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
#define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
#else
#define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
#define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
#endif
// INTERNAL, DO NOT USE
inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
{
static bool _initialized = false;
static bool _append = false;
static bool _disabled = filename.empty() && target == nullptr;
static std::string log_current_filename{filename};
static FILE *log_current_target{target};
static FILE *logfile = nullptr;
if (change)
{
if (append != LogTriStateSame)
{
_append = append == LogTriStateTrue;
return logfile;
}
if (disable == LogTriStateTrue)
{
// Disable primary target
_disabled = true;
}
// If previously disabled, only enable, and keep previous target
else if (disable == LogTriStateFalse)
{
_disabled = false;
}
// Otherwise, process the arguments
else if (log_current_filename != filename || log_current_target != target)
{
_initialized = false;
}
}
if (_disabled)
{
// Log is disabled
return nullptr;
}
if (_initialized)
{
// with fallback in case something went wrong
return logfile ? logfile : stderr;
}
// do the (re)initialization
if (target != nullptr)
{
if (logfile != nullptr && logfile != stdout && logfile != stderr)
{
fclose(logfile);
}
log_current_filename = LOG_DEFAULT_FILE_NAME;
log_current_target = target;
logfile = target;
}
else
{
if (log_current_filename != filename)
{
if (logfile != nullptr && logfile != stdout && logfile != stderr)
{
fclose(logfile);
}
}
logfile = fopen(filename.c_str(), _append ? "a" : "w");
}
if (!logfile)
{
// Verify whether the file was opened, otherwise fallback to stderr
logfile = stderr;
fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
fflush(stderr);
// At this point we let the init flag be to true below, and let the target fallback to stderr
// otherwise we would repeatedly fopen() which was already unsuccessful
}
_initialized = true;
return logfile ? logfile : stderr;
}
// INTERNAL, DO NOT USE
inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
{
return log_handler1_impl(change, append, disable, filename, target);
}
// Disables logs entirely at runtime.
// Makes LOG() and LOG_TEE() produce no output,
// until enabled back.
#define log_disable() log_disable_impl()
// INTERNAL, DO NOT USE
inline FILE *log_disable_impl()
{
return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
}
// Enables logs at runtime.
#define log_enable() log_enable_impl()
// INTERNAL, DO NOT USE
inline FILE *log_enable_impl()
{
return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
}
// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
#define log_set_target(target) log_set_target_impl(target)
// INTERNAL, DO NOT USE
inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
// INTERNAL, DO NOT USE
inline FILE *log_handler() { return log_handler1_impl(); }
// Enable or disable creating separate log files for each run.
// can ONLY be invoked BEFORE first log use.
#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
// Enable or disable append mode for log file.
// can ONLY be invoked BEFORE first log use.
#define log_append(enable) log_append_impl(enable)
// INTERNAL, DO NOT USE
inline FILE *log_append_impl(bool enable)
{
return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
}
inline void log_test()
{
log_disable();
LOG("01 Hello World to nobody, because logs are disabled!\n");
log_enable();
LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
log_set_target(stderr);
LOG("04 Hello World to stderr!\n");
LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
log_set_target(LOG_DEFAULT_FILE_NAME);
LOG("06 Hello World to default log file!\n");
log_set_target(stdout);
LOG("07 Hello World to stdout!\n");
log_set_target(LOG_DEFAULT_FILE_NAME);
LOG("08 Hello World to default log file again!\n");
log_disable();
LOG("09 Hello World _1_ into the void!\n");
log_enable();
LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
log_disable();
log_set_target("llama.anotherlog.log");
LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
log_enable();
LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
log_set_target("llama.yetanotherlog.log");
LOG("13 Hello World this time in yet new file?\n");
log_set_target(log_filename_generator("llama_autonamed", "log"));
LOG("14 Hello World in log with generated filename!\n");
#ifdef _MSC_VER
LOG_TEE("15 Hello msvc TEE without arguments\n");
LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
LOG_TEELN("17 Hello msvc TEELN without arguments\n");
LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
LOG("19 Hello msvc LOG without arguments\n");
LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
LOGLN("21 Hello msvc LOGLN without arguments\n");
LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
#endif
}
inline bool log_param_single_parse(const std::string & param)
{
if ( param == "--log-test")
{
log_test();
return true;
}
if ( param == "--log-disable")
{
log_disable();
return true;
}
if ( param == "--log-enable")
{
log_enable();
return true;
}
if (param == "--log-new")
{
log_multilog(true);
return true;
}
if (param == "--log-append")
{
log_append(true);
return true;
}
return false;
}
inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
{
if ( param == "--log-file")
{
if (!check_but_dont_parse)
{
log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
}
return true;
}
return false;
}
inline void log_print_usage()
{
printf("log options:\n");
/* format
printf(" -h, --help show this help message and exit\n");*/
/* spacing
printf("__-param----------------Description\n");*/
printf(" --log-test Run simple logging test\n");
printf(" --log-disable Disable trace logs\n");
printf(" --log-enable Enable trace logs\n");
printf(" --log-file Specify a log filename (without extension)\n");
printf(" --log-new Create a separate new log file on start. "
"Each log file will have unique name: \"<name>.<ID>.log\"\n");
printf(" --log-append Don't truncate the old log file.\n");
printf("\n");
}
#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
// INTERNAL, DO NOT USE
inline void log_dump_cmdline_impl(int argc, char **argv)
{
std::stringstream buf;
for (int i = 0; i < argc; ++i)
{
if (std::string(argv[i]).find(' ') != std::string::npos)
{
buf << " \"" << argv[i] <<"\"";
}
else
{
buf << " " << argv[i];
}
}
LOGLN("Cmd:%s", buf.str().c_str());
}
#define log_tostr(var) log_var_to_string_impl(var).c_str()
inline std::string log_var_to_string_impl(bool var)
{
return var ? "true" : "false";
}
inline std::string log_var_to_string_impl(std::string var)
{
return var;
}
inline std::string log_var_to_string_impl(const std::vector<int> & var)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (auto e : var)
{
if (first)
{
first = false;
}
else
{
buf << ", ";
}
buf << std::to_string(e);
}
buf << " ]";
return buf.str();
}
template <typename C, typename T>
inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (const auto & token : tokens)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, token);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "'" << detokenized << "'"
<< ":" << std::to_string(token);
}
buf << " ]";
return buf.str();
}
template <typename C, typename B>
inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (int i = 0; i < batch.n_tokens; ++i)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "\n" << std::to_string(i)
<< ":token '" << detokenized << "'"
<< ":pos " << std::to_string(batch.pos[i])
<< ":n_seq_id " << std::to_string(batch.n_seq_id[i])
<< ":seq_id " << std::to_string(batch.seq_id[i][0])
<< ":logits " << std::to_string(batch.logits[i]);
}
buf << " ]";
return buf.str();
}
#ifdef LOG_DISABLE_LOGS
#undef LOG
#define LOG(...) // dummy stub
#undef LOGLN
#define LOGLN(...) // dummy stub
#undef LOG_TEE
#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
#undef LOG_TEELN
#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
#undef LOG_DISABLE
#define LOG_DISABLE() // dummy stub
#undef LOG_ENABLE
#define LOG_ENABLE() // dummy stub
#undef LOG_ENABLE
#define LOG_ENABLE() // dummy stub
#undef LOG_SET_TARGET
#define LOG_SET_TARGET(...) // dummy stub
#undef LOG_DUMP_CMDLINE
#define LOG_DUMP_CMDLINE(...) // dummy stub
#endif // LOG_DISABLE_LOGS
#define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO, verbosity, __VA_ARGS__)
#define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN, verbosity, __VA_ARGS__)
#define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
#define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
#define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT, verbosity, __VA_ARGS__)

75
common/ngram-cache.cpp
View file

@ -2,10 +2,13 @@
#include "common.h"
#include "log.h"
#include <cinttypes>
#include <cstdint>
#include <cstdio>
#include <fstream>
#include <thread>
void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
std::vector<llama_token> & inp, int nnew, bool print_progress) {
const int64_t t_start_ms = ggml_time_ms();
const int64_t inp_size = inp.size();
@ -17,16 +20,16 @@ void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, in
const int64_t i_start = std::max(inp_size - nnew, ngram_size);
for (int64_t i = i_start; i < inp_size; ++i) {
const int64_t ngram_start = i - ngram_size;
llama_ngram ngram(&inp[ngram_start], ngram_size);
common_ngram ngram(&inp[ngram_start], ngram_size);
const llama_token token = inp[i];
llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
common_ngram_cache::iterator part_it = ngram_cache.find(ngram);
if (part_it == ngram_cache.end()) {
llama_ngram_cache_part part;
common_ngram_cache_part part;
part.emplace(token, 1);
ngram_cache.emplace(ngram, part);
} else {
llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
common_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
if (token_count_it == part_it->second.end()) {
part_it->second.emplace(token, 1);
} else {
@ -59,12 +62,12 @@ constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4, 3, 2, 2};
constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
// Helper function that tries to draft a token from only the static ngram cache:
static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
if (part_static_it == nc_static.end()) {
return -1;
}
const llama_ngram_cache_part part_static = part_static_it->second;
const common_ngram_cache_part part_static = part_static_it->second;
int max_count_static = 0;
int sum_count_static = 0;
@ -92,19 +95,19 @@ static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ng
// Try to draft a token from primary cache (context/dynamic), validate with static cache:
static llama_token try_draft(
llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
const int * min_sample_size, const int * min_percent) {
llama_token drafted_token = -1;
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
const llama_ngram ngram_primary = ngrams_primary[i];
const common_ngram ngram_primary = ngrams_primary[i];
llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
if (part_primary_it == nc_primary.end()) {
continue;
}
const llama_ngram_cache_part part_primary = part_primary_it->second;
const common_ngram_cache_part part_primary = part_primary_it->second;
int max_count_primary = 0;
int max_count_static = 0;
@ -114,7 +117,7 @@ static llama_token try_draft(
for (std::pair<llama_token, int> token_count_primary : part_primary) {
const llama_token token = token_count_primary.first;
llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
common_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
const int32_t count_primary = token_count_primary.second;
const int32_t count_static = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
@ -139,9 +142,9 @@ static llama_token try_draft(
return drafted_token;
}
void llama_ngram_cache_draft(
void common_ngram_cache_draft(
std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static
) {
GGML_ASSERT(draft.size() == 1);
const int inp_size = inp.size();
@ -154,21 +157,21 @@ void llama_ngram_cache_draft(
llama_token drafted_token = -1;
const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
llama_ngram ngram_static;
common_ngram ngram_static;
for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
}
llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
llama_ngram_cache_part part_static;
common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
common_ngram_cache_part part_static;
if (part_static_it != nc_static.end()) {
part_static = part_static_it->second;
}
// cd = context + dynamic
std::vector<llama_ngram> ngrams_cd;
std::vector<common_ngram> ngrams_cd;
for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
llama_ngram ngram_cd;
common_ngram ngram_cd;
for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
}
@ -193,16 +196,16 @@ void llama_ngram_cache_draft(
}
}
void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
std::ofstream file_out(filename, std::ios::binary);
for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
const llama_ngram ngram = item.first;
llama_ngram_cache_part token_counts = item.second;
for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
const common_ngram ngram = item.first;
common_ngram_cache_part token_counts = item.second;
GGML_ASSERT(!token_counts.empty());
const int32_t ntokens = token_counts.size();
GGML_ASSERT(ntokens > 0);
file_out.write(reinterpret_cast<const char *>(&ngram), sizeof(llama_ngram));
file_out.write(reinterpret_cast<const char *>(&ngram), sizeof(common_ngram));
file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
for (std::pair<llama_token, int32_t> item2 : token_counts) {
const llama_token token = item2.first;
@ -216,14 +219,14 @@ void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filen
}
llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
common_ngram_cache common_ngram_cache_load(std::string & filename) {
std::ifstream hashmap_file(filename, std::ios::binary);
if (!hashmap_file) {
throw std::ifstream::failure("Unable to open file " + filename);
}
llama_ngram_cache ngram_cache;
common_ngram_cache ngram_cache;
llama_ngram ngram;
common_ngram ngram;
int32_t ntokens;
llama_token token;
int32_t count;
@ -232,11 +235,11 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
char * ntokensc = reinterpret_cast<char*>(&ntokens);
char * tokenc = reinterpret_cast<char*>(&token);
char * countc = reinterpret_cast<char*>(&count);
while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
while(hashmap_file.read(ngramc, sizeof(common_ngram))) {
GGML_ASSERT(!hashmap_file.eof());
GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
GGML_ASSERT(ntokens > 0);
llama_ngram_cache_part token_counts;
common_ngram_cache_part token_counts;
for (int i = 0; i < ntokens; ++i) {
GGML_ASSERT(!hashmap_file.eof());
@ -254,12 +257,12 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
return ngram_cache;
}
void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
const llama_ngram ngram = ngram_part.first;
llama_ngram_cache_part part = ngram_part.second;
void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add) {
for (std::pair<common_ngram, common_ngram_cache_part> ngram_part : ngram_cache_add) {
const common_ngram ngram = ngram_part.first;
common_ngram_cache_part part = ngram_part.second;
llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
common_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
if (part_merged_it == ngram_cache_target.end()) {
ngram_cache_target.emplace(ngram, part);
continue;
@ -270,7 +273,7 @@ void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram
const int32_t count = token_count.second;
GGML_ASSERT(count > 0);
llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
common_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
if (token_count_merged_it == part_merged_it->second.end()) {
part_merged_it->second.emplace(token, count);
continue;

38
common/ngram-cache.h
View file

@ -12,22 +12,22 @@
// Data structures to map n-grams to empirical token probabilities:
struct llama_ngram {
struct common_ngram {
llama_token tokens[LLAMA_NGRAM_MAX];
llama_ngram() {
common_ngram() {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = -1;
}
}
llama_ngram(const llama_token * input, const int ngram_size) {
common_ngram(const llama_token * input, const int ngram_size) {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = i < ngram_size ? input[i] : -1;
}
}
bool operator==(const llama_ngram & other) const {
bool operator==(const common_ngram & other) const {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
if (tokens[i] != other.tokens[i]) {
return false;
@ -37,28 +37,28 @@ struct llama_ngram {
}
};
struct llama_token_hash_function {
struct common_token_hash_function {
size_t operator()(const llama_token token) const {
// see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
return token * 11400714819323198485llu;
}
};
struct llama_ngram_hash_function {
size_t operator()(const llama_ngram & ngram) const {
size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
struct common_ngram_hash_function {
size_t operator()(const common_ngram & ngram) const {
size_t hash = common_token_hash_function{}(ngram.tokens[0]);
for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
hash ^= llama_token_hash_function{}(ngram.tokens[i]);
hash ^= common_token_hash_function{}(ngram.tokens[i]);
}
return hash;
}
};
// token -> number of times token has been seen
typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
typedef std::unordered_map<llama_token, int32_t> common_ngram_cache_part;
// n-gram -> empirical distribution of following tokens
typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
typedef std::unordered_map<common_ngram, common_ngram_cache_part, common_ngram_hash_function> common_ngram_cache;
// Update an ngram cache with tokens.
@ -70,8 +70,8 @@ typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash
//
// In order to get correct results inp_data can ONLY BE APPENDED TO.
// Changes in the middle need a complete rebuild.
void llama_ngram_cache_update(
llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
void common_ngram_cache_update(
common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
// Try to draft tokens from ngram caches.
// inp: the tokens generated so far.
@ -81,21 +81,21 @@ void llama_ngram_cache_update(
// nc_context: ngram cache based on current context.
// nc_dynamic: ngram cache based on previous user generations.
// nc_static: ngram cache generated from a large text corpus, used for validation.
void llama_ngram_cache_draft(
void common_ngram_cache_draft(
std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static);
// Save an ngram cache to a file.
// ngram_cache: the ngram cache to save.
// filename: the path under which to save the ngram cache.
void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
// Load an ngram cache saved with llama_ngram_cache_save.
// Load an ngram cache saved with common_ngram_cache_save.
// filename: the path from which to load the ngram cache.
// returns: an ngram cache containing the information saved to filename.
llama_ngram_cache llama_ngram_cache_load(std::string & filename);
common_ngram_cache common_ngram_cache_load(std::string & filename);
// Merge two ngram caches.
// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
// ngram_cache_add: the ngram cache to add to ngram_cache_target.
void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);
void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add);

871
common/sampling.cpp
View file

@ -1,460 +1,505 @@
#define LLAMA_API_INTERNAL
#include "sampling.h"
#include <random>
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
struct llama_sampling_context * result = new llama_sampling_context();
#include "common.h"
result->params = params;
result->grammar = nullptr;
#include <cmath>
#include <unordered_map>
// if there is a grammar, parse it
if (!params.grammar.empty()) {
result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// the ring buffer works similarly to std::deque, but with a fixed capacity
// TODO: deduplicate with llama-impl.h
template<typename T>
struct ring_buffer {
ring_buffer(size_t cap) : capacity(cap), data(cap) {}
// will be empty (default) if there are parse errors
if (result->parsed_grammar.rules.empty()) {
fprintf(stderr, "%s: failed to parse grammar\n", __func__);
delete result;
return nullptr;
T & front() {
if (sz == 0) {
throw std::runtime_error("ring buffer is empty");
}
return data[first];
}
const T & front() const {
if (sz == 0) {
throw std::runtime_error("ring buffer is empty");
}
return data[first];
}
T & back() {
if (sz == 0) {
throw std::runtime_error("ring buffer is empty");
}
return data[pos];
}
const T & back() const {
if (sz == 0) {
throw std::runtime_error("ring buffer is empty");
}
return data[pos];
}
void push_back(const T & value) {
if (sz == capacity) {
// advance the start when buffer is full
first = (first + 1) % capacity;
} else {
sz++;
}
data[pos] = value;
pos = (pos + 1) % capacity;
}
T pop_front() {
if (sz == 0) {
throw std::runtime_error("ring buffer is empty");
}
T value = data[first];
first = (first + 1) % capacity;
sz--;
return value;
}
const T & rat(size_t i) const {
if (i >= sz) {
throw std::runtime_error("ring buffer: index out of bounds");
}
return data[(first + sz - i - 1) % capacity];
}
std::vector<T> to_vector() const {
std::vector<T> result;
result.reserve(sz);
for (size_t i = 0; i < sz; i++) {
result.push_back(data[(first + i) % capacity]);
}
return result;
}
void clear() {
// here only reset the status of the buffer
sz = 0;
first = 0;
pos = 0;
}
bool empty() const {
return sz == 0;
}
size_t size() const {
return sz;
}
size_t capacity = 0;
size_t sz = 0;
size_t first = 0;
size_t pos = 0;
std::vector<T> data;
};
struct common_sampler {
common_params_sampling params;
struct llama_sampler * grmr;
struct llama_sampler * chain;
ring_buffer<llama_token> prev;
std::vector<llama_token_data> cur;
llama_token_data_array cur_p;
void set_logits(struct llama_context * ctx, int idx) {
const auto * logits = llama_get_logits_ith(ctx, idx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
// Ensure that there is a "root" node.
if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
delete result;
return nullptr;
}
std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
struct llama_grammar * grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
}
result->grammar = grammar;
cur_p = { cur.data(), cur.size(), -1, false };
}
};
result->prev.resize(params.n_prev);
result->n_valid = 0;
llama_sampling_set_rng_seed(result, params.seed);
return result;
}
void llama_sampling_free(struct llama_sampling_context * ctx) {
if (ctx->grammar != NULL) {
llama_grammar_free(ctx->grammar);
}
delete ctx;
}
void llama_sampling_reset(llama_sampling_context * ctx) {
if (ctx->grammar != NULL) {
llama_grammar_free(ctx->grammar);
ctx->grammar = NULL;
}
if (!ctx->parsed_grammar.rules.empty()) {
std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
struct llama_grammar * grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
}
ctx->grammar = grammar;
}
std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
ctx->cur.clear();
ctx->n_valid = 0;
}
void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
if (seed == LLAMA_DEFAULT_SEED) {
seed = std::random_device{}();
}
ctx->rng.seed(seed);
}
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
if (dst->grammar) {
llama_grammar_free(dst->grammar);
dst->grammar = nullptr;
}
if (src->grammar) {
dst->grammar = llama_grammar_copy(src->grammar);
}
dst->prev = src->prev;
}
llama_token llama_sampling_last(llama_sampling_context * ctx) {
return ctx->prev.back();
}
std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
const int size = ctx_sampling->prev.size();
n = std::min(n, size);
std::string result;
for (int i = size - n; i < size; i++) {
result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
}
return result;
}
std::string llama_sampling_print(const llama_sampling_params & params) {
std::string common_params_sampling::print() const {
char result[1024];
snprintf(result, sizeof(result),
"\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
"\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
"\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
"\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
params.mirostat, params.mirostat_eta, params.mirostat_tau);
penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
mirostat, mirostat_eta, mirostat_tau);
return std::string(result);
}
std::string llama_sampling_order_print(const llama_sampling_params & params) {
std::string result = "CFG -> Penalties ";
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = params.no_perf;
auto * result = new common_sampler {
/* .params = */ params,
/* .grmr = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
/* .chain = */ llama_sampler_chain_init(lparams),
/* .prev = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
/* .cur = */ {},
/* .cur_p = */ {},
};
llama_sampler_chain_add(result->chain,
llama_sampler_init_logit_bias(
llama_n_vocab(model),
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 (auto sampler_type : params.samplers_sequence) {
const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
if (!sampler_type_name.empty()) {
result += "-> " + sampler_type_name + " ";
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
{
std::vector<const char*> c_breakers;
c_breakers.reserve(params.dry_sequence_breakers.size());
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;
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break;
case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_XTC:
llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
} else if (params.mirostat == 2) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
} else {
result += "-> mirostat ";
GGML_ASSERT(false && "unknown mirostat version");
}
return result;
}
std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
switch (sampler_type) {
case llama_sampler_type::TOP_K: return "top_k";
case llama_sampler_type::TFS_Z: return "tfs_z";
case llama_sampler_type::TYPICAL_P: return "typical_p";
case llama_sampler_type::TOP_P: return "top_p";
case llama_sampler_type::MIN_P: return "min_p";
case llama_sampler_type::TEMPERATURE: return "temperature";
void common_sampler_free(struct common_sampler * gsmpl) {
if (gsmpl) {
llama_sampler_free(gsmpl->grmr);
llama_sampler_free(gsmpl->chain);
delete gsmpl;
}
}
void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
if (accept_grammar) {
llama_sampler_accept(gsmpl->grmr, token);
}
llama_sampler_accept(gsmpl->chain, token);
gsmpl->prev.push_back(token);
}
void common_sampler_reset(struct common_sampler * gsmpl) {
llama_sampler_reset(gsmpl->grmr);
llama_sampler_reset(gsmpl->chain);
}
struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
return new common_sampler {
/* .params = */ gsmpl->params,
/* .grmr = */ llama_sampler_clone(gsmpl->grmr),
/* .chain = */ llama_sampler_clone(gsmpl->chain),
/* .prev = */ gsmpl->prev,
/* .cur = */ gsmpl->cur,
/* .cur_p = */ gsmpl->cur_p,
};
}
void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
// TODO: measure grammar performance
if (gsmpl) {
llama_perf_sampler_print(gsmpl->chain);
}
if (ctx) {
llama_perf_context_print(ctx);
}
}
llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
gsmpl->set_logits(ctx, idx);
auto & grmr = gsmpl->grmr;
auto & chain = gsmpl->chain;
auto & cur_p = gsmpl->cur_p; // initialized by set_logits
if (grammar_first) {
llama_sampler_apply(grmr, &cur_p);
}
llama_sampler_apply(chain, &cur_p);
GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
const llama_token id = cur_p.data[cur_p.selected].id;
if (grammar_first) {
return id;
}
// check if it the sampled token fits the grammar
{
llama_token_data single_token_data = { id, 1.0f, 0.0f };
llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
llama_sampler_apply(grmr, &single_token_data_array);
const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
if (is_valid) {
return id;
}
}
// resampling:
// if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
gsmpl->set_logits(ctx, idx);
llama_sampler_apply(grmr, &cur_p);
llama_sampler_apply(chain, &cur_p);
GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
return cur_p.data[cur_p.selected].id;
}
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
std::vector<llama_token> result;
result.reserve(idxs.size());
size_t i = 0;
for (; i < draft.size(); i++) {
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
common_sampler_accept(gsmpl, id, true);
result.push_back(id);
if (draft[i] != id) {
break;
}
}
if (i == draft.size()) {
const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
common_sampler_accept(gsmpl, id, true);
result.push_back(id);
}
return result;
}
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
std::vector<int> idxs(draft.size() + 1);
for (size_t i = 0; i < idxs.size(); ++i) {
idxs[i] = i;
}
return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
}
uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
return llama_sampler_get_seed(gsmpl->chain);
}
// helpers
llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
return &gsmpl->cur_p;
}
llama_token common_sampler_last(const struct common_sampler * gsmpl) {
return gsmpl->prev.rat(0);
}
std::string common_sampler_print(const struct common_sampler * gsmpl) {
std::string result = "logits ";
for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
result += std::string("-> ") + llama_sampler_name(smpl) + " ";
}
return result;
}
std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_main, int n) {
n = std::min(n, (int) gsmpl->prev.size());
if (n <= 0) {
return "";
}
std::string result;
result.reserve(8*n); // 8 is the average length of a token [citation needed], TODO: compute this from the vocab
for (int i = n - 1; i >= 0; i--) {
const llama_token id = gsmpl->prev.rat(i);
GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
result += common_token_to_piece(ctx_main, id);
}
return result;
}
char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY: return 'd';
case COMMON_SAMPLER_TYPE_TOP_K: return 'k';
case COMMON_SAMPLER_TYPE_TYPICAL_P: return 'y';
case COMMON_SAMPLER_TYPE_TOP_P: return 'p';
case COMMON_SAMPLER_TYPE_MIN_P: return 'm';
case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
case COMMON_SAMPLER_TYPE_XTC: return 'x';
case COMMON_SAMPLER_TYPE_INFILL: return 'i';
default : return '?';
}
}
std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY: return "dry";
case COMMON_SAMPLER_TYPE_TOP_K: return "top_k";
case COMMON_SAMPLER_TYPE_TYPICAL_P: return "typ_p";
case COMMON_SAMPLER_TYPE_TOP_P: return "top_p";
case COMMON_SAMPLER_TYPE_MIN_P: return "min_p";
case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
case COMMON_SAMPLER_TYPE_XTC: return "xtc";
case COMMON_SAMPLER_TYPE_INFILL: return "infill";
default : return "";
}
}
std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
{"top_k", llama_sampler_type::TOP_K},
{"top_p", llama_sampler_type::TOP_P},
{"typical_p", llama_sampler_type::TYPICAL_P},
{"min_p", llama_sampler_type::MIN_P},
{"tfs_z", llama_sampler_type::TFS_Z},
{"temperature", llama_sampler_type::TEMPERATURE}
std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
{ "dry", COMMON_SAMPLER_TYPE_DRY },
{ "top_k", COMMON_SAMPLER_TYPE_TOP_K },
{ "top_p", COMMON_SAMPLER_TYPE_TOP_P },
{ "typ_p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min_p", COMMON_SAMPLER_TYPE_MIN_P },
{ "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
{ "xtc", COMMON_SAMPLER_TYPE_XTC },
{ "infill", COMMON_SAMPLER_TYPE_INFILL },
};
// since samplers names are written multiple ways
// make it ready for both system names and input names
std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
{"top-k", llama_sampler_type::TOP_K},
{"top-p", llama_sampler_type::TOP_P},
{"nucleus", llama_sampler_type::TOP_P},
{"typical-p", llama_sampler_type::TYPICAL_P},
{"typical", llama_sampler_type::TYPICAL_P},
{"min-p", llama_sampler_type::MIN_P},
{"tfs-z", llama_sampler_type::TFS_Z},
{"tfs", llama_sampler_type::TFS_Z},
{"temp", llama_sampler_type::TEMPERATURE}
std::unordered_map<std::string, common_sampler_type> sampler_alt_name_map {
{ "top-k", COMMON_SAMPLER_TYPE_TOP_K },
{ "top-p", COMMON_SAMPLER_TYPE_TOP_P },
{ "nucleus", COMMON_SAMPLER_TYPE_TOP_P },
{ "typical-p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typical", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typ-p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "typ", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min-p", COMMON_SAMPLER_TYPE_MIN_P },
{ "temp", COMMON_SAMPLER_TYPE_TEMPERATURE },
};
std::vector<llama_sampler_type> sampler_types;
sampler_types.reserve(names.size());
for (const auto & name : names)
{
auto sampler_item = sampler_canonical_name_map.find(name);
if (sampler_item != sampler_canonical_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
else
{
if (allow_alt_names)
{
sampler_item = sampler_alt_name_map.find(name);
if (sampler_item != sampler_alt_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
}
}
}
return sampler_types;
}
std::vector<common_sampler_type> samplers;
samplers.reserve(names.size());
std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
std::unordered_map<char, llama_sampler_type> sampler_name_map {
{'k', llama_sampler_type::TOP_K},
{'p', llama_sampler_type::TOP_P},
{'y', llama_sampler_type::TYPICAL_P},
{'m', llama_sampler_type::MIN_P},
{'f', llama_sampler_type::TFS_Z},
{'t', llama_sampler_type::TEMPERATURE}
};
std::vector<llama_sampler_type> sampler_types;
sampler_types.reserve(names_string.size());
for (const auto & c : names_string) {
const auto sampler_item = sampler_name_map.find(c);
if (sampler_item != sampler_name_map.end()) {
sampler_types.push_back(sampler_item->second);
}
}
return sampler_types;
}
// no reasons to expose this function in header
static void sampler_queue(
struct llama_context * ctx_main,
const llama_sampling_params & params,
llama_token_data_array & cur_p,
size_t min_keep) {
const float temp = params.temp;
const float dynatemp_range = params.dynatemp_range;
const float dynatemp_exponent = params.dynatemp_exponent;
const int32_t top_k = params.top_k;
const float top_p = params.top_p;
const float min_p = params.min_p;
const float tfs_z = params.tfs_z;
const float typical_p = params.typical_p;
const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
for (auto sampler_type : samplers_sequence) {
switch (sampler_type) {
case llama_sampler_type::TOP_K : llama_sample_top_k (ctx_main, &cur_p, top_k, min_keep); break;
case llama_sampler_type::TFS_Z : llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep); break;
case llama_sampler_type::TYPICAL_P: llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); break;
case llama_sampler_type::TOP_P : llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); break;
case llama_sampler_type::MIN_P : llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); break;
case llama_sampler_type::TEMPERATURE:
if (dynatemp_range > 0) {
float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
} else {
llama_sample_temp(ctx_main, &cur_p, temp);
}
break;
default : break;
}
}
}
static llama_token llama_sampling_sample_impl(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx,
bool is_resampling) {
const llama_sampling_params & params = ctx_sampling->params;
const float temp = params.temp;
const int mirostat = params.mirostat;
const float mirostat_tau = params.mirostat_tau;
const float mirostat_eta = params.mirostat_eta;
std::vector<float> original_logits;
auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
if (ctx_sampling->grammar != NULL && !is_resampling) {
GGML_ASSERT(!original_logits.empty());
}
llama_token id = 0;
if (temp < 0.0) {
// greedy sampling, with probs
llama_sample_softmax(ctx_main, &cur_p);
id = cur_p.data[0].id;
} else if (temp == 0.0) {
// greedy sampling, no probs
id = llama_sample_token_greedy(ctx_main, &cur_p);
} else {
if (mirostat == 1) {
const int mirostat_m = 100;
llama_sample_temp(ctx_main, &cur_p, temp);
id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
} else if (mirostat == 2) {
llama_sample_temp(ctx_main, &cur_p, temp);
id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
for (const auto & name : names) {
auto sampler = sampler_canonical_name_map.find(name);
if (sampler != sampler_canonical_name_map.end()) {
samplers.push_back(sampler->second);
} else {
// temperature sampling
size_t min_keep = std::max(1, params.min_keep);
sampler_queue(ctx_main, params, cur_p, min_keep);
id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
//{
// const int n_top = 10;
// LOG("top %d candidates:\n", n_top);
// for (int i = 0; i < n_top; i++) {
// const llama_token id = cur_p.data[i].id;
// (void)id; // To avoid a warning that id is unused when logging is disabled.
// LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
// }
//}
//LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
}
}
if (ctx_sampling->grammar != NULL && !is_resampling) {
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
// Create an array with a single token data element for the sampled id
llama_token_data single_token_data = {id, logits[id], 0.0f};
llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
// Apply grammar constraints to the single token
llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
// Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
// If the token is not valid according to the grammar, perform resampling
if (!is_valid) {
LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
// Restore logits from the copy
std::copy(original_logits.begin(), original_logits.end(), logits);
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
}
}
ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;
return id;
}
static llama_token_data_array llama_sampling_prepare_impl(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx,
bool apply_grammar,
std::vector<float> * original_logits) {
const llama_sampling_params & params = ctx_sampling->params;
const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
const float penalty_repeat = params.penalty_repeat;
const float penalty_freq = params.penalty_freq;
const float penalty_present = params.penalty_present;
const bool penalize_nl = params.penalize_nl;
auto & prev = ctx_sampling->prev;
auto & cur = ctx_sampling->cur;
// Get a pointer to the logits
float * logits = llama_get_logits_ith(ctx_main, idx);
if (ctx_sampling->grammar != NULL && !apply_grammar) {
GGML_ASSERT(original_logits != NULL);
// Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
*original_logits = {logits, logits + n_vocab};
}
// apply params.logit_bias map
for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
logits[it->first] += it->second;
}
if (ctx_cfg) {
float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
}
cur.resize(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
}
llama_token_data_array cur_p = { cur.data(), cur.size(), false };
// apply penalties
const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
if (penalty_tokens_used_size) {
const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
llama_sample_repetition_penalties(ctx_main, &cur_p,
penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
if (!penalize_nl) {
for (size_t idx = 0; idx < cur_p.size; idx++) {
if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
cur_p.data[idx].logit = nl_logit;
break;
if (allow_alt_names) {
sampler = sampler_alt_name_map.find(name);
if (sampler != sampler_alt_name_map.end()) {
samplers.push_back(sampler->second);
}
}
}
}
// apply grammar checks before sampling logic
if (apply_grammar && ctx_sampling->grammar != NULL) {
llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
return samplers;
}
std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
std::unordered_map<char, common_sampler_type> sampler_name_map = {
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY), COMMON_SAMPLER_TYPE_DRY },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K), COMMON_SAMPLER_TYPE_TOP_K },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P), COMMON_SAMPLER_TYPE_TYPICAL_P },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P), COMMON_SAMPLER_TYPE_TOP_P },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P), COMMON_SAMPLER_TYPE_MIN_P },
{ 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 },
};
std::vector<common_sampler_type> samplers;
samplers.reserve(chars.size());
for (const auto & c : chars) {
const auto sampler = sampler_name_map.find(c);
if (sampler != sampler_name_map.end()) {
samplers.push_back(sampler->second);
}
}
return cur_p;
}
llama_token llama_sampling_sample(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx) {
// Call the implementation function with is_resampling set to false by default
return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
}
llama_token_data_array llama_sampling_prepare(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx,
bool apply_grammar,
std::vector<float> * original_logits) {
return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
}
void llama_sampling_accept(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
llama_token id,
bool apply_grammar) {
ctx_sampling->prev.erase(ctx_sampling->prev.begin());
ctx_sampling->prev.push_back(id);
if (ctx_sampling->grammar != NULL && apply_grammar) {
llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id);
}
return samplers;
}

238
common/sampling.h
View file

@ -2,159 +2,103 @@
#include "llama.h"
#include "grammar-parser.h"
#include <random>
#include <string>
#include <unordered_map>
#include <vector>
// sampler types
enum class llama_sampler_type : char {
TOP_K = 'k',
TOP_P = 'p',
MIN_P = 'm',
TFS_Z = 'f',
TYPICAL_P = 'y',
TEMPERATURE = 't'
};
// sampling parameters
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
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
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
std::vector<llama_sampler_type> samplers_sequence = {
llama_sampler_type::TOP_K,
llama_sampler_type::TFS_Z,
llama_sampler_type::TYPICAL_P,
llama_sampler_type::TOP_P,
llama_sampler_type::MIN_P,
llama_sampler_type::TEMPERATURE
};
std::string grammar; // optional BNF-like grammar to constrain sampling
// Classifier-Free Guidance
// https://arxiv.org/abs/2306.17806
std::string cfg_negative_prompt; // string to help guidance
float cfg_scale = 1.f; // how strong is guidance
std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
std::vector<llama_token> penalty_prompt_tokens;
bool use_penalty_prompt_tokens = false;
} llama_sampling_params;
// general sampler context
// TODO: move to llama.h
struct llama_sampling_context {
// parameters that will be used for sampling
llama_sampling_params params;
// mirostat sampler state
float mirostat_mu;
llama_grammar * grammar;
// internal
grammar_parser::parse_state parsed_grammar;
// TODO: replace with ring-buffer
std::vector<llama_token> prev;
std::vector<llama_token_data> cur;
size_t n_valid; // Number of correct top tokens with correct probabilities.
std::mt19937 rng;
};
#include "common.h"
// Create a new sampling context instance.
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
#include <string>
#include <vector>
void llama_sampling_free(struct llama_sampling_context * ctx);
// Reset the sampler context
// - clear prev tokens
// - reset grammar
void llama_sampling_reset(llama_sampling_context * ctx);
// Set the sampler seed
void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
// Copy the sampler context
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
// Get the last sampled token
llama_token llama_sampling_last(llama_sampling_context * ctx);
// Get a string representation of the last sampled tokens
std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
// Print sampling parameters into a string
std::string llama_sampling_print(const llama_sampling_params & params);
// Print sampling order into a string
std::string llama_sampling_order_print(const llama_sampling_params & params);
std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
// this is a common sampling function used across the examples for convenience
// it can serve as a starting point for implementing your own sampling function
// Note: When using multiple sequences, it is the caller's responsibility to call
// llama_sampling_reset when a sequence ends
// common_sampler extends llama_sampler with additional functionality:
//
// required:
// - ctx_main: context to use for sampling
// - ctx_sampling: sampling-specific context
// - grammar support
// - custom sampler logic based on the parameters
// - history of the last accepted tokens
// - performance metrics
//
// optional:
// - ctx_cfg: context to use for classifier-free guidance
// - idx: sample from llama_get_logits_ith(ctx, idx)
// This goal is to have a common implementation of the sampling logic shared across the examples.
// For example, depending on the temperature, the sampling chain can be very simple (greedy) or more
// complex (top-k, top-p, etc).
//
// returns:
// - token: sampled token
// - candidates: vector of candidate tokens
// Another example is related to the grammar. In general, the grammar constraints applied on the full
// vocabulary can be very taxing. To improve performance, the grammar can be applied only to the sampled
// token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the
// grammar constraints are applied to the full vocabulary and the token is resampled.
//
// The common_sampler also maintains a container with the last accepted tokens. In the future, this can
// be moved into the core llama library.
//
// For convenience, the common_sampler also maintains a container with the current candidate tokens.
// This can be used to access the probabilities of the rest of the non-sampled tokens.
//
// TODO: measure grammar performance
//
llama_token llama_sampling_sample(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
int idx = -1);
// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
llama_token_data_array llama_sampling_prepare(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
int idx = 0,
bool apply_grammar = true,
std::vector<float> * original_logits = nullptr);
struct common_sampler;
void llama_sampling_accept(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
llama_token id,
bool apply_grammar);
// llama_sampler API overloads
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
void common_sampler_free(struct common_sampler * gsmpl);
// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar);
void common_sampler_reset (struct common_sampler * gsmpl);
struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
// arguments can be nullptr to skip printing
void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
// extended sampling implementation:
//
// - set logits
// - apply the configured sampler chain
// - check if the token fits the grammar (if any)
// - if not: resample by first applying the grammar constraints and then sampling again (slower path)
//
// if grammar_first is true, the grammar is applied before the samplers (slower)
// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
//
llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
// generalized version of common_sampler_sample
//
// will cross-reference the sampled tokens with a batch of draft tokens and accept those that match
// if the sampler disagrees at some point, we stop and return the accepted tokens up to now
//
// common_sampler_sample_n(gsmpl, ctx, { idx }, {});
//
// is equivalent to
//
// common_sampler_sample(gsmpl, ctx, idx);
// common_sampler_accept(gsmpl, token, true);
//
// requires: idxs.size() == draft.size() + 1
//
// returns at least 1 token, up to idxs.size()
//
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
// assume idxs == [ 0, 1, 2, ..., draft.size() ]
std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
// helpers
// access the internal list of current candidate tokens
llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
// get the last accepted token
llama_token common_sampler_last(const struct common_sampler * gsmpl);
// print the sampler chain into a string
std::string common_sampler_print(const struct common_sampler * gsmpl);
// get a string representation of the last accepted tokens
std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx, int n);
char common_sampler_type_to_chr(enum common_sampler_type cnstr);
std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);

270
common/speculative.cpp Normal file
View file

@ -0,0 +1,270 @@
#include "speculative.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include <cstring>
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
struct common_speculative {
struct llama_context * ctx;
struct common_sampler * smpl;
llama_batch batch;
llama_tokens prompt;
};
struct common_speculative * common_speculative_init(
struct llama_context * ctx_dft) {
auto * result = new common_speculative {
/* .ctx = */ ctx_dft,
/* .smpl = */ nullptr,
/* .batch = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
/* .prompt = */ {},
};
// TODO: optimize or pass from outside?
#if 0
{
common_params_sampling params;
params.no_perf = false;
params.top_k = 40;
params.top_p = 0.9;
params.samplers = {
COMMON_SAMPLER_TYPE_TOP_K,
COMMON_SAMPLER_TYPE_TOP_P,
COMMON_SAMPLER_TYPE_INFILL,
};
result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
}
#else
{
common_params_sampling params;
params.no_perf = false;
params.top_k = 10;
params.samplers = {
COMMON_SAMPLER_TYPE_TOP_K,
};
result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
}
#endif
return result;
}
void common_speculative_free(struct common_speculative * spec) {
common_sampler_free(spec->smpl);
llama_batch_free(spec->batch);
delete spec;
}
bool common_speculative_are_compatible(
const struct llama_context * ctx_tgt,
const struct llama_context * ctx_dft) {
const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
const struct llama_model * model_dft = llama_get_model(ctx_dft);
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
"vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
return false;
}
if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
return false;
}
{
const int n_vocab_tgt = llama_n_vocab(model_tgt);
const int n_vocab_dft = llama_n_vocab(model_dft);
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
__func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return false;
}
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but "
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
common_token_to_piece(ctx_tgt, i).c_str(),
common_token_to_piece(ctx_dft, i).c_str());
return false;
}
}
}
return true;
}
llama_tokens common_speculative_gen_draft(
struct common_speculative * spec,
struct common_speculative_params params,
const llama_tokens & prompt_tgt,
llama_token id_last) {
auto & batch = spec->batch;
auto & ctx = spec->ctx;
auto & smpl = spec->smpl;
auto & prompt = spec->prompt;
int reuse_i = 0;
int reuse_n = 0;
const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
// reuse as much as possible from the old draft context
// ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
for (int i = 0; i < (int) prompt.size(); ++i) {
int cur = 0;
while (i_start + cur < (int) prompt_tgt.size() &&
i + cur < (int) prompt.size() &&
prompt_tgt[i_start + cur] == prompt[i + cur]) {
cur++;
}
if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) {
reuse_i = i;
reuse_n = cur;
}
}
LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
llama_tokens result;
result.reserve(params.n_draft);
if (reuse_n == 0) {
llama_kv_cache_clear(ctx);
prompt.clear();
} else {
// this happens when a previous draft has been discarded (for example, due to being too small), but the
// target model agreed with it. in this case, we simply pass back the previous results to save compute
if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
result.push_back(prompt[i]);
if (params.n_draft <= (int) result.size()) {
break;
}
}
return result;
}
if (reuse_i > 0) {
llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
}
if (reuse_n < (int) prompt.size()) {
llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
prompt.erase(prompt.begin() + reuse_n, prompt.end());
}
}
// prepare a batch to evaluate any new tokens in the prompt
common_batch_clear(batch);
for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
//LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
prompt.push_back(prompt_tgt[i]);
}
// we should rarely end-up here during normal decoding
if (batch.n_tokens > 0) {
//LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
llama_decode(ctx, batch);
}
const llama_pos n_past = prompt.size();
LOG_DBG("%s: n_past = %d\n", __func__, n_past);
common_batch_clear(batch);
common_batch_add (batch, id_last, n_past, { 0 }, true);
prompt.push_back(id_last);
//LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
llama_decode(ctx, batch);
common_sampler_reset(smpl);
// sample n_draft tokens from the draft model
for (int i = 0; i < params.n_draft; ++i) {
common_batch_clear(batch);
common_sampler_sample(smpl, ctx, 0, true);
const auto * cur_p = common_sampler_get_candidates(smpl);
for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
}
// add drafted token for each sequence
const llama_token id = cur_p->data[0].id;
// only collect very high-confidence draft tokens
if (cur_p->data[0].p < params.p_min) {
break;
}
common_sampler_accept(smpl, id, true);
result.push_back(id);
if (params.n_draft <= (int) result.size()) {
break;
}
common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
// evaluate the drafted tokens on the draft model
llama_decode(ctx, batch);
prompt.push_back(id);
}
return result;
}

28
common/speculative.h Normal file
View file

@ -0,0 +1,28 @@
#pragma once
#include "llama.h"
#include "common.h"
struct common_speculative;
struct common_speculative_params {
int n_draft = 16; // max drafted tokens
int n_reuse = 256;
float p_min = 0.9f; // min probabiliy required to accept a token in the draft
};
struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);
void common_speculative_free(struct common_speculative * spec);
bool common_speculative_are_compatible(
const struct llama_context * ctx_tgt,
const struct llama_context * ctx_dft);
// sample up to n_draft tokens and add them to the batch using the draft model
llama_tokens common_speculative_gen_draft(
struct common_speculative * spec,
struct common_speculative_params params,
const llama_tokens & prompt,
llama_token id_last);

11662
common/stb_image.h
View file

File diff suppressed because it is too large Load diff

1513
common/train.cpp
View file

File diff suppressed because it is too large Load diff

233
common/train.h
View file

@ -1,233 +0,0 @@
// Various helper functions and utilities for training
#pragma once
#include <string>
#include <random>
#include <vector>
#include "ggml.h"
#include "llama.h"
#define LLAMA_TRAIN_MAX_NODES 16384
typedef std::string mt19937_state;
struct train_state {
struct ggml_opt_context * opt;
uint64_t train_its;
uint64_t train_samples;
uint64_t train_tokens;
uint64_t train_epochs;
size_t shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
mt19937_state shuffle_rng_state_current;
mt19937_state shuffle_rng_state_next;
size_t shuffle_sample_count;
size_t shuffle_next_sample;
};
struct train_params_common {
const char * fn_train_data;
const char * fn_checkpoint_in;
const char * fn_checkpoint_out;
const char * pattern_fn_it;
const char * fn_latest;
bool print_usage;
int save_every;
uint32_t seed;
int n_ctx;
int n_threads;
int n_batch;
int n_gradient_accumulation;
int n_epochs;
int n_gpu_layers;
bool custom_n_ctx;
bool use_flash;
bool use_checkpointing;
std::string sample_start;
bool include_sample_start;
bool escape;
bool overlapping_samples;
bool fill_with_next_samples;
bool separate_with_eos;
bool separate_with_bos;
bool sample_random_offsets;
bool force_reshuffle;
int warmup;
int cos_decay_steps;
float cos_decay_restart;
float cos_decay_min;
bool enable_restart;
int opt_past;
float opt_delta;
int opt_max_no_improvement;
int adam_n_iter;
float adam_alpha;
float adam_min_alpha;
float adam_decay;
int adam_decay_min_ndim;
float adam_beta1;
float adam_beta2;
float adam_gclip;
float adam_eps_f;
};
typedef void (*save_train_files_callback)(void * data, struct train_state * train);
struct train_opt_callback_data {
struct train_params_common * params;
struct train_state * train;
save_train_files_callback save_cb;
void * save_data;
struct llama_context * lctx;
int last_save_iter;
llama_token * tokens_data;
size_t tokens_size;
size_t * samples_begin;
size_t * samples_size;
size_t * shuffled_samples_offs;
size_t * shuffled_samples_begin;
size_t * shuffled_samples_size;
size_t samples_count;
struct ggml_tensor * tokens_input;
struct ggml_tensor * target_probs;
int first_iter;
int first_epoch;
int iter_at_last_epoch;
int64_t last_time;
double millis_per_iter;
};
struct train_state * init_train_state();
void free_train_state(struct train_state * state);
struct train_params_common get_default_train_params_common();
void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
void finish_processing_train_args(struct train_params_common * params);
struct random_normal_distribution;
struct random_uniform_distribution;
struct random_normal_distribution * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
void free_random_normal_distribution (struct random_normal_distribution * rnd);
void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
// generate random float in interval [0,1)
float frand();
float frand_normal (struct random_normal_distribution * rnd);
float frand_uniform(struct random_uniform_distribution * rnd);
int clamp (const int v, const int min, const int max);
float fclamp(const float v, const float min, const float max);
void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
size_t tokenize_file(
struct llama_context * lctx,
const char * filename,
const std::string & sample_start,
bool include_sample_start,
bool overlapping_samples,
unsigned context_length,
std::vector<llama_token> & out_tokens,
std::vector<size_t> & out_samples_begin,
std::vector<size_t> & out_samples_size);
int64_t get_example_targets_batch(
struct llama_context * lctx,
struct ggml_tensor * tokens_input,
struct ggml_tensor * target_probs,
int64_t example_id,
const size_t * samples_offs,
const size_t * samples_begin,
const size_t * samples_size,
size_t samples_count,
const llama_token * train_data,
size_t n_train_data,
bool separate_with_eos,
bool separate_with_bos,
bool fill_with_next_samples,
bool sample_random_offsets);
void mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
mt19937_state mt19937_get_state(const std::mt19937& rng);
mt19937_state mt19937_seed_to_state(unsigned seed);
mt19937_state shuffle_samples(
const mt19937_state & rng_state,
size_t * shuffled_offs,
size_t * shuffled_begins,
size_t * shuffled_sizes,
const size_t * begins,
const size_t * sizes,
size_t count);
size_t hash_combine(size_t h1, size_t h2);
size_t compute_samples_hash(
const char* fn,
const size_t* samples_begin,
const size_t* samples_size,
size_t sample_count);
std::string replace_str(const char * s, const char * needle, const char * replacement);
void print_duration(double milliseconds);
float cosine_decay(
int64_t step,
int64_t decay_steps,
float minimum);
float cosine_decay_restart(
int64_t step,
int64_t decay_steps,
float minimum,
float restart_step_mult);
float learning_schedule(
int64_t step,
int64_t warmup_steps,
int64_t decay_steps,
float learning_rate,
float overall_minimum,
float cos_decay_minimum,
float cos_decay_restart_step_mult,
bool enable_restart);
void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

491
convert_hf_to_gguf.py
View file

@ -3,6 +3,7 @@
from __future__ import annotations
import ast
import logging
import argparse
import contextlib
@ -14,6 +15,7 @@ from enum import IntEnum
from pathlib import Path
from hashlib import sha256
from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
from itertools import chain
import math
import numpy as np
@ -63,7 +65,6 @@ class Model:
model_name: str | None
metadata_override: Path | None
dir_model_card: Path
is_lora: bool
# subclasses should define this!
model_arch: gguf.MODEL_ARCH
@ -71,7 +72,8 @@ class Model:
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
use_temp_file: bool = False, eager: bool = False,
metadata_override: Path | None = None, model_name: str | None = None,
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False):
split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
if type(self) is Model:
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
@ -86,14 +88,13 @@ class Model:
self.is_safetensors = len(self.part_names) > 0
if not self.is_safetensors:
self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
self.hparams = Model.load_hparams(self.dir_model)
self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
self.tensor_names = None
self.metadata_override = metadata_override
self.model_name = model_name
self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py
self.is_lora = is_lora # true if model is used inside convert_lora_to_gguf.py
# Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
if self.ftype == gguf.LlamaFileType.GUESSED:
@ -131,12 +132,14 @@ class Model:
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_names_from_parts: set[str] = set()
if len(self.part_names) > 1:
index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
index_name += ".index.json"
index_file = self.dir_model / index_name
if index_file.is_file():
self.tensor_names = set()
index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
index_name += ".index.json"
logger.info(f"gguf: loading model weight map from '{index_name}'")
with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
with open(index_file, "r", encoding="utf-8") as f:
index: dict[str, Any] = json.load(f)
weight_map = index.get("weight_map")
if weight_map is None or not isinstance(weight_map, dict):
@ -144,6 +147,7 @@ class Model:
self.tensor_names.update(weight_map.keys())
else:
self.tensor_names = tensor_names_from_parts
weight_map = {}
for part_name in self.part_names:
logger.info(f"gguf: loading model part '{part_name}'")
@ -170,9 +174,17 @@ class Model:
data = LazyTorchTensor.from_eager(data)
yield name, data
# only verify tensor name presence; it doesn't matter if they are not in the right files
if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
# verify tensor name presence and identify potentially missing files
if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
if len(extra) == 0 and len(missing_files) > 0:
raise ValueError(f"Missing or incomplete model files: {missing_files}")
else:
raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
f"Missing tensors: {missing}\n"
f"Extra tensors: {extra}")
def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
if key not in gguf.MODEL_TENSORS[self.model_arch]:
@ -258,10 +270,14 @@ class Model:
return False
# some models need extra generated tensors (like rope_freqs)
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
return ()
def prepare_tensors(self):
max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
for name, data_torch in self.get_tensors():
for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()):
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
@ -279,8 +295,13 @@ class Model:
bid = int(part)
break
for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
data: np.ndarray # type hint
for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
data = data_torch.squeeze().numpy()
# if data ends up empty, it means data_torch was a scalar tensor -> restore
if len(data.shape) == 0:
data = data_torch.numpy()
n_dims = len(data.shape)
data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
@ -298,12 +319,31 @@ class Model:
gguf.MODEL_TENSOR.POS_EMBD,
gguf.MODEL_TENSOR.TOKEN_TYPES,
gguf.MODEL_TENSOR.SSM_CONV1D,
gguf.MODEL_TENSOR.TIME_MIX_FIRST,
gguf.MODEL_TENSOR.TIME_MIX_W1,
gguf.MODEL_TENSOR.TIME_MIX_W2,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
)
)
or not name.endswith(".weight")
or not new_name.endswith(".weight")
):
data_qtype = gguf.GGMLQuantizationType.F32
if data_qtype is False and any(
self.match_model_tensor_name(new_name, key, bid)
for key in (
gguf.MODEL_TENSOR.TOKEN_EMBD,
gguf.MODEL_TENSOR.OUTPUT,
)
):
if self.ftype in (
gguf.LlamaFileType.MOSTLY_TQ1_0,
gguf.LlamaFileType.MOSTLY_TQ2_0,
):
# TODO: use Q4_K and Q6_K
data_qtype = gguf.GGMLQuantizationType.F16
# No override (data_qtype is False), or wants to be quantized (data_qtype is True)
if isinstance(data_qtype, bool):
if self.ftype == gguf.LlamaFileType.ALL_F32:
@ -314,6 +354,10 @@ class Model:
data_qtype = gguf.GGMLQuantizationType.BF16
elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
data_qtype = gguf.GGMLQuantizationType.Q8_0
elif self.ftype == gguf.LlamaFileType.MOSTLY_TQ1_0:
data_qtype = gguf.GGMLQuantizationType.TQ1_0
elif self.ftype == gguf.LlamaFileType.MOSTLY_TQ2_0:
data_qtype = gguf.GGMLQuantizationType.TQ2_0
else:
raise ValueError(f"Unknown file type: {self.ftype.name}")
@ -530,6 +574,9 @@ class Model:
if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
# ref: https://huggingface.co/BAAI/bge-small-en-v1.5
res = "bert-bge"
if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
# ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
res = "bert-bge-large"
if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
# ref: https://huggingface.co/mosaicml/mpt-7b
res = "mpt"
@ -557,6 +604,9 @@ class Model:
if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
# ref: https://huggingface.co/databricks/dbrx-base
res = "dbrx"
if chkhsh == "c7699093ba4255a91e702aa38a596aa81669f3525dae06c2953267dde580f448":
# ref: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
res = "jina-v1-en"
if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
# ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
res = "jina-v2-en"
@ -602,6 +652,12 @@ class Model:
if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
# ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
res = "exaone"
if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
# ref: https://huggingface.co/microsoft/phi-2
res = "phi-2"
if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
# ref: https://huggingface.co/facebook/chameleon-7b
res = "chameleon"
if res is None:
logger.warning("\n")
@ -1460,7 +1516,7 @@ class StableLMModel(Model):
raise ValueError(f"Unprocessed norms: {norms}")
@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
@Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
class LlamaModel(Model):
model_arch = gguf.MODEL_ARCH.LLAMA
@ -1486,6 +1542,17 @@ class LlamaModel(Model):
special_vocab._set_special_token("eot", 32010)
special_vocab.add_to_gguf(self.gguf_writer)
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
if "add_prefix_space" in tokenizer_config_json:
self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
# Apply to granite small models only
if self.hparams.get("vocab_size", 32000) == 49152:
self.gguf_writer.add_add_bos_token(False)
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
@ -1502,17 +1569,6 @@ class LlamaModel(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
if tokenizer_config_file.is_file():
with open(tokenizer_config_file, "r", encoding="utf-8") as f:
tokenizer_config_json = json.load(f)
if "add_prefix_space" in tokenizer_config_json:
self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
# Apply to granite small models only
if self.hparams.get("vocab_size", 32000) == 49152:
self.gguf_writer.add_add_bos_token(False)
@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:
@ -1568,11 +1624,11 @@ class LlamaModel(Model):
return [(self.map_tensor_name(name), data_torch)]
def prepare_tensors(self):
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
if rope_scaling.get("rope_type", '').lower() == "llama3":
base = self.hparams.get("rope_theta", 10000.0)
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
factor = rope_scaling.get("factor", 8.0)
@ -1595,9 +1651,9 @@ class LlamaModel(Model):
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
rope_factors.append(1 / ((1 - smooth) / factor + smooth))
if not self.is_lora:
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
def prepare_tensors(self):
super().prepare_tensors()
if self._experts is not None:
@ -1619,15 +1675,16 @@ class BitnetModel(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(1.0)
def weight_quant(self, weight):
def weight_quant(self, weight: Tensor) -> Tensor:
dtype = weight.dtype
weight = weight.float()
s = 1 / weight.abs().mean().clamp(min=1e-5)
weight = (weight * s).round().clamp(-1, 1) / s
scale = weight.abs().max().unsqueeze(0)
weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
weight = torch.sign(weight).type(dtype)
return weight.type(dtype), scale.type(torch.float32)
scale = weight.abs().mean().clamp(min=1e-5)
iscale = 1 / scale
# TODO: multiply by the scale directly instead of inverting it twice
# (this is also unnecessarily doubly inverted upstream)
# ref: https://huggingface.co/1bitLLM/bitnet_b1_58-3B/blob/af89e318d78a70802061246bf037199d2fb97020/utils_quant.py#L10
result = (weight * iscale).round().clamp(-1, 1) / iscale
return result.type(dtype)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
new_name = self.map_tensor_name(name)
@ -1642,11 +1699,9 @@ class BitnetModel(Model):
gguf.MODEL_TENSOR.FFN_GATE,
]):
# transform weight into 1/0/-1 (in fp32)
weight_torch, scale_torch = self.weight_quant(data_torch)
yield (new_name, weight_torch)
yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
else:
yield (new_name, data_torch)
data_torch = self.weight_quant(data_torch)
yield (new_name, data_torch)
@Model.register("GrokForCausalLM")
@ -1815,6 +1870,59 @@ class MiniCPMModel(Model):
return [(self.map_tensor_name(name), data_torch)]
@Model.register("MiniCPM3ForCausalLM")
class MiniCPM3Model(Model):
model_arch = gguf.MODEL_ARCH.MINICPM3
def set_gguf_parameters(self):
hparams = self.hparams
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
rope_scaling = self.find_hparam(['rope_scaling'], True)
if rope_scaling is not None:
rope_dims = self.hparams["qk_rope_head_dim"]
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_sentencepiece()
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)
)
@Model.register("QWenLMHeadModel")
class QwenModel(Model):
model_arch = gguf.MODEL_ARCH.QWEN
@ -2114,6 +2222,13 @@ class Phi3MiniModel(Model):
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
rope_dims = n_embd // n_head
# write rope scaling for long context (128k) model
rope_scaling = self.find_hparam(['rope_scaling'], True)
if rope_scaling is None:
@ -2143,9 +2258,8 @@ class Phi3MiniModel(Model):
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}')
if not self.is_lora:
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
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))
@Model.register("PlamoForCausalLM")
@ -2507,7 +2621,7 @@ class NomicBertModel(BertModel):
self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
@Model.register("XLMRobertaModel")
@Model.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
class XLMRobertaModel(BertModel):
model_arch = gguf.MODEL_ARCH.BERT
@ -2593,7 +2707,7 @@ class XLMRobertaModel(BertModel):
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
self.gguf_writer.add_add_space_prefix(add_prefix)
self.gguf_writer.add_token_type_count(1)
self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
if precompiled_charsmap:
self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
@ -2605,6 +2719,11 @@ class XLMRobertaModel(BertModel):
self.gguf_writer.add_add_eos_token(True)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# if name starts with "roberta.", remove the prefix
# e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
if name.startswith("roberta."):
name = name[8:]
# position embeddings start at pad_token_id + 1, so just chop down the weight tensor
if name == "embeddings.position_embeddings.weight":
if self._position_offset is not None:
@ -2716,6 +2835,89 @@ class StarCoder2Model(Model):
model_arch = gguf.MODEL_ARCH.STARCODER2
@Model.register("Rwkv6ForCausalLM")
class Rwkv6Model(Model):
model_arch = gguf.MODEL_ARCH.RWKV6
def set_vocab(self):
assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
vocab_size = self.hparams.get("vocab_size", 65536)
tokens: list[bytes] = ['<s>'.encode("utf-8")]
toktypes: list[int] = [gguf.TokenType.CONTROL]
with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
lines = f.readlines()
for line in lines:
parts = line.split(' ')
assert len(parts) >= 3
token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
token = token.encode("utf-8") if isinstance(token, str) else token
assert isinstance(token, bytes)
assert len(token) == token_len
token_text: str = repr(token)[2:-1] # "b'\xff'" -> "\xff"
tokens.append(token_text.encode("utf-8"))
toktypes.append(gguf.TokenType.NORMAL)
remainder = vocab_size - len(tokens)
assert remainder >= 0
for i in range(len(tokens), vocab_size):
tokens.append(f"[PAD{i}]".encode("utf-8"))
toktypes.append(gguf.TokenType.UNUSED)
self.gguf_writer.add_tokenizer_model("rwkv")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
special_vocab.chat_template = "rwkv-world"
# hack: Add '\n\n' as the EOT token to make it chat normally
special_vocab._set_special_token("eot", 261)
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
head_size = self.hparams["head_size"]
hidden_size = self.hparams["hidden_size"]
layer_norm_eps = self.hparams["layer_norm_epsilon"]
rescale_every_n_layers = self.hparams["rescale_every"]
intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else int((hidden_size * 3.5) // 32 * 32)
time_mix_extra_dim = 64 if hidden_size == 4096 else 32
time_decay_extra_dim = 128 if hidden_size == 4096 else 64
# RWKV isn't context limited
self.gguf_writer.add_context_length(1048576)
self.gguf_writer.add_embedding_length(hidden_size)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
self.gguf_writer.add_rescale_every_n_layers(rescale_every_n_layers)
self.gguf_writer.add_wkv_head_size(head_size)
self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
self.gguf_writer.add_feed_forward_length(intermediate_size)
self.gguf_writer.add_file_type(self.ftype)
# required by llama.cpp, unused
self.gguf_writer.add_head_count(0)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
new_name = self.map_tensor_name(name)
if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
new_name += ".weight"
if new_name.endswith("time_mix_w1.weight") or new_name.endswith("time_mix_decay_w1.weight") or new_name.endswith("time_mix_decay_w2.weight"):
data_torch = data_torch.transpose(0, 1)
if new_name.endswith("time_mix_w2.weight"):
data_torch = data_torch.permute(0, 2, 1)
rescale_every_n_layers = self.hparams["rescale_every"]
if rescale_every_n_layers > 0:
if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
yield (new_name, data_torch)
@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
class MambaModel(Model):
model_arch = gguf.MODEL_ARCH.MAMBA
@ -2838,6 +3040,71 @@ class OlmoModel(Model):
return [(self.map_tensor_name(name), data_torch)]
@Model.register("Olmo2ForCausalLM")
class Olmo2Model(Model):
model_arch = gguf.MODEL_ARCH.OLMO2
@Model.register("OlmoeForCausalLM")
class OlmoeModel(Model):
model_arch = gguf.MODEL_ARCH.OLMOE
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_layer_norm_rms_eps(1e-5)
if (n_experts := self.hparams.get("num_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
_experts: list[dict[str, Tensor]] | None = None
# Copied from: Qwen2MoeModel
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# process the experts separately
if name.find("experts") != -1:
n_experts = self.hparams["num_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
# Copied from: Qwen2MoeModel
def prepare_tensors(self):
super().prepare_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("JinaBertModel", "JinaBertForMaskedLM")
class JinaBertV2Model(BertModel):
model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
@ -2876,6 +3143,14 @@ class JinaBertV2Model(BertModel):
self.gguf_writer.add_add_bos_token(True)
self.gguf_writer.add_add_eos_token(True)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# if name starts with "bert.", remove the prefix
# e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
if name.startswith("bert."):
name = name[5:]
return super().modify_tensors(data_torch, name, bid)
@Model.register("OpenELMForCausalLM")
class OpenELMModel(Model):
@ -3478,10 +3753,7 @@ class JaisModel(Model):
# Embeddings scale
self.embeddings_scale = 1.0
# note: For some JAIS flavors, output is tied to (same as) wte in original model
self.output_is_wte = False
if 'mup_embeddings_scale' in self.hparams:
self.output_is_wte = True # Hack (?)
self.embeddings_scale = self.hparams['mup_embeddings_scale']
elif 'embeddings_scale' in self.hparams:
self.embeddings_scale = self.hparams['embeddings_scale']
@ -3538,10 +3810,7 @@ class JaisModel(Model):
if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
tensors.append((new_name, data_torch * self.embeddings_scale))
if self.output_is_wte:
tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
assert not self.output_is_wte
tensors.append((new_name, data_torch * self.width_scale))
else:
tensors.append((new_name, data_torch))
@ -3816,11 +4085,11 @@ class ExaoneModel(Model):
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
def prepare_tensors(self):
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
if rope_scaling.get("rope_type", '').lower() == "llama3":
base = self.hparams.get("rope_theta", 10000.0)
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
factor = rope_scaling.get("factor", 8.0)
@ -3843,14 +4112,112 @@ class ExaoneModel(Model):
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
rope_factors.append(1 / ((1 - smooth) / factor + smooth))
if not self.is_lora:
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
super().prepare_tensors()
@Model.register("GraniteForCausalLM")
class GraniteModel(LlamaModel):
"""Conversion for IBM's GraniteForCausalLM"""
model_arch = gguf.MODEL_ARCH.GRANITE
def set_gguf_parameters(self):
"""Granite uses standard llama parameters with the following differences:
- No head_dim support
- New multiplier params:
- attention_scale
- embedding_scale
- residual_scale
- logits_scaling
"""
if head_dim := self.hparams.pop("head_dim", None):
logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
super().set_gguf_parameters()
# NOTE: Convert _multiplier params to _scale params for naming
# consistency
if attention_scale := self.hparams.get("attention_multiplier"):
self.gguf_writer.add_attention_scale(attention_scale)
logger.info("gguf: (granite) attention_scale = %s", attention_scale)
if embedding_scale := self.hparams.get("embedding_multiplier"):
self.gguf_writer.add_embedding_scale(embedding_scale)
logger.info("gguf: (granite) embedding_scale = %s", embedding_scale)
if residual_scale := self.hparams.get("residual_multiplier"):
self.gguf_writer.add_residual_scale(residual_scale)
logger.info("gguf: (granite) residual_scale = %s", residual_scale)
if logits_scale := self.hparams.get("logits_scaling"):
self.gguf_writer.add_logit_scale(logits_scale)
logger.info("gguf: (granite) logits_scale = %s", logits_scale)
@Model.register("GraniteMoeForCausalLM")
class GraniteMoeModel(GraniteModel):
"""Conversion for IBM's GraniteMoeForCausalLM"""
model_arch = gguf.MODEL_ARCH.GRANITE_MOE
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
"""In modeling_granitemoe, the JetMoe implementation of parallel experts
is used. This essentially merges w1 and w3 into a single tensor with 2x
the hidden size that is then split during forward. To keep compatibility
with existing mixtral support, we pull them apart here.
"""
if name.endswith("block_sparse_moe.input_linear.weight"):
ffn_dim = self.hparams["intermediate_size"]
assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
return [
(self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
(self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
]
return super().modify_tensors(data_torch, name, bid)
@Model.register("ChameleonForConditionalGeneration")
@Model.register("ChameleonForCausalLM") # obsolete
class ChameleonModel(Model):
model_arch = gguf.MODEL_ARCH.CHAMELEON
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
def set_vocab(self):
self._set_vocab_gpt2()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# ignore image tokenizer for now
# TODO: remove this once image support is implemented for Chameleon
if name.startswith("model.vqmodel"):
return []
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
hidden_dim = self.hparams.get("hidden_size")
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
if name.endswith(("q_norm.weight", "q_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
if name.endswith(("k_norm.weight", "k_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
return [(self.map_tensor_name(name), data_torch)]
# see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
@staticmethod
def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
head_dim = hidden_dim // n_heads
data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
data_torch = data_torch.repeat_interleave(n_heads, 0)
return data_torch
###### CONVERSION LOGIC ######
# tree of lazy tensors
class LazyTorchTensor(gguf.LazyBase):
_tensor_type = torch.Tensor
@ -3929,8 +4296,8 @@ def parse_args() -> argparse.Namespace:
help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
)
parser.add_argument(
"--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
"--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="f16",
help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
)
parser.add_argument(
"--bigendian", action="store_true",
@ -4017,6 +4384,8 @@ def main() -> None:
"f16": gguf.LlamaFileType.MOSTLY_F16,
"bf16": gguf.LlamaFileType.MOSTLY_BF16,
"q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
"tq1_0": gguf.LlamaFileType.MOSTLY_TQ1_0,
"tq2_0": gguf.LlamaFileType.MOSTLY_TQ2_0,
"auto": gguf.LlamaFileType.GUESSED,
}

32
convert_hf_to_gguf_update.py
View file

@ -31,6 +31,7 @@ import re
import requests
import sys
import json
import shutil
from hashlib import sha256
from enum import IntEnum, auto
@ -71,6 +72,7 @@ models = [
{"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
{"name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
{"name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
{"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
{"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
@ -80,6 +82,7 @@ models = [
{"name": "qwen2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
{"name": "olmo", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
{"name": "dbrx", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
{"name": "jina-v1-en", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
{"name": "jina-v2-en", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
{"name": "jina-v2-es", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
{"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
@ -97,6 +100,8 @@ models = [
{'name': "bloom", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
{'name': "gpt3-finnish", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
{"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", },
]
@ -125,12 +130,27 @@ def download_model(model):
if tokt == TOKENIZER_TYPE.UGM:
files.append("spiece.model")
for file in files:
save_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(save_path):
logger.info(f"{name}: File {save_path} already exists - skipping")
continue
download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
if os.path.isdir(repo):
# If repo is a path on the file system, copy the directory
for file in files:
src_path = os.path.join(repo, file)
dst_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(dst_path):
logger.info(f"{name}: File {dst_path} already exists - skipping")
continue
if os.path.isfile(src_path):
shutil.copy2(src_path, dst_path)
logger.info(f"{name}: Copied {src_path} to {dst_path}")
else:
logger.warning(f"{name}: Source file {src_path} does not exist")
else:
# If repo is a URL, download the files
for file in files:
save_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(save_path):
logger.info(f"{name}: File {save_path} already exists - skipping")
continue
download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
for model in models:

65
convert_lora_to_gguf.py
View file

@ -12,6 +12,7 @@ import json
from math import prod
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
from transformers import AutoConfig
import torch
@ -230,7 +231,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
parser.add_argument(
"--outfile", type=Path,
help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
@ -256,17 +257,23 @@ def parse_args() -> argparse.Namespace:
help="only print out what will be done, without writing any new files",
)
parser.add_argument(
"--base", type=Path, required=True,
help="directory containing base model file",
"--base", type=Path,
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
)
parser.add_argument(
"lora_path", type=Path,
help="directory containing LoRA adapter file",
help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
)
return parser.parse_args()
def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
# normally, adapter does not come with base model config, we need to load it from AutoConfig
config = AutoConfig.from_pretrained(hf_model_id)
return config.to_dict()
if __name__ == '__main__':
args = parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
@ -281,7 +288,7 @@ if __name__ == '__main__':
ftype = ftype_map[args.outtype]
dir_base_model: Path = args.base
dir_base_model: Path | None = args.base
dir_lora: Path = args.lora_path
lora_config = dir_lora / "adapter_config.json"
input_model = dir_lora / "adapter_model.safetensors"
@ -301,9 +308,29 @@ if __name__ == '__main__':
input_model = os.path.join(dir_lora, "adapter_model.bin")
lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
# load LoRA config
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
# load base model
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = Model.load_hparams(dir_base_model)
if dir_base_model is None:
if "base_model_name_or_path" in lparams:
model_id = lparams["base_model_name_or_path"]
logger.info(f"Loading base model from Hugging Face: {model_id}")
try:
hparams = load_hparams_from_hf(model_id)
except OSError as e:
logger.error(f"Failed to load base model config: {e}")
logger.error("Please try downloading the base model and add its path to --base")
sys.exit(1)
else:
logger.error("'base_model_name_or_path' is not found in adapter_config.json")
logger.error("Base model config is required. Please download the base model and add its path to --base")
sys.exit(1)
else:
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = Model.load_hparams(dir_base_model)
with torch.inference_mode():
try:
model_class = Model.from_model_architecture(hparams["architectures"][0])
@ -323,13 +350,19 @@ if __name__ == '__main__':
self.dir_model_card = dir_lora_model
self.lora_alpha = float(lora_alpha)
def set_vocab(self):
pass
def set_type(self):
self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
def set_gguf_parameters(self):
self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
super().set_gguf_parameters()
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
# Never add extra tensors (e.g. rope_freqs) for LoRA adapters
return ()
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_map: dict[str, PartialLoraTensor] = {}
@ -344,6 +377,9 @@ if __name__ == '__main__':
if ".base_layer.weight" in name:
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
sys.exit(1)
if base_name in tensor_map:
@ -363,7 +399,13 @@ if __name__ == '__main__':
yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
dest = super().modify_tensors(data_torch, name, bid)
dest = list(super().modify_tensors(data_torch, name, bid))
# some archs may have the same tensor for lm_head and output (tie word embeddings)
# in this case, adapters targeting lm_head will fail when using llama-export-lora
# therefore, we ignore them for now
# see: https://github.com/ggerganov/llama.cpp/issues/9065
if name == "lm_head.weight" and len(dest) == 0:
raise ValueError("lm_head is present in adapter, but is ignored in base model")
for dest_name, dest_data in dest:
assert isinstance(dest_data, LoraTorchTensor)
lora_a, lora_b = dest_data.get_lora_A_B()
@ -371,9 +413,6 @@ if __name__ == '__main__':
yield (dest_name + ".lora_a", lora_a)
yield (dest_name + ".lora_b", lora_b)
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
alpha: float = lparams["lora_alpha"]
model_instance = LoraModel(
@ -386,7 +425,7 @@ if __name__ == '__main__':
dry_run=args.dry_run,
dir_lora_model=dir_lora,
lora_alpha=alpha,
is_lora=True,
hparams=hparams,
)
logger.info("Exporting model...")

109
docs/android.md
View file

@ -2,55 +2,82 @@
# Android
## Build on Android using Termux
[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
```
apt update && apt upgrade -y
apt install git make cmake
```
It's recommended to move your model inside the `~/` directory for best performance:
```
cd storage/downloads
mv model.gguf ~/
```
[Termux](https://termux.dev/en/) is an Android terminal emulator and Linux environment app (no root required). As of writing, Termux is available experimentally in the Google Play Store; otherwise, it may be obtained directly from the project repo or on F-Droid.
[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
## Building the Project using Android NDK
Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
```
$ mkdir build-android
$ cd build-android
$ export NDK=<your_ndk_directory>
$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
$ make
```
Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
```
$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
$cd /data/data/com.termux/files/home/bin
$chmod +x ./*
```
Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
With Termux, you can install and run `llama.cpp` as if the environment were Linux. Once in the Termux shell:
```
$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
$ apt update && apt upgrade -y
$ apt install git cmake
```
Now, you can start chatting:
Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake.
Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance:
```
$cd /data/data/com.termux/files/home/bin
$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
$ curl -L {model-url} -o ~/{model}.gguf
```
Here's a demo of an interactive session running on Pixel 5 phone:
Then, if you are not already in the repo directory, `cd` into `llama.cpp` and:
```
$ ./build/bin/llama-cli -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
```
Here, we show `llama-cli`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
To see what it might look like visually, here's an old demo of an interactive session running on a Pixel 5 phone:
https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
## Cross-compile using Android NDK
It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
Once you're ready and have cloned `llama.cpp`, invoke the following in the project directory:
```
$ cmake \
-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
-DANDROID_ABI=arm64-v8a \
-DANDROID_PLATFORM=android-28 \
-DCMAKE_C_FLAGS="-march=armv8.7a" \
-DCMAKE_CXX_FLAGS="-march=armv8.7a" \
-DGGML_OPENMP=OFF \
-DGGML_LLAMAFILE=OFF \
-B build-android
```
Notes:
- While later versions of Android NDK ship with OpenMP, it must still be installed by CMake as a dependency, which is not supported at this time
- `llamafile` does not appear to support Android devices (see: https://github.com/Mozilla-Ocho/llamafile/issues/325)
The above command should configure `llama.cpp` with the most performant options for modern devices. Even if your device is not running `armv8.7a`, `llama.cpp` includes runtime checks for available CPU features it can use.
Feel free to adjust the Android ABI for your target. Once the project is configured:
```
$ cmake --build build-android --config Release -j{n}
$ cmake --install build-android --prefix {install-dir} --config Release
```
After installing, go ahead and download the model of your choice to your host system. Then:
```
$ adb shell "mkdir /data/local/tmp/llama.cpp"
$ adb push {install-dir} /data/local/tmp/llama.cpp/
$ adb push {model}.gguf /data/local/tmp/llama.cpp/
$ adb shell
```
In the `adb shell`:
```
$ cd /data/local/tmp/llama.cpp
$ LD_LIBRARY_PATH=lib ./bin/llama-simple -m {model}.gguf -c {context-size} -p "{your-prompt}"
```
That's it!
Be aware that Android will not find the library path `lib` on its own, so we must specify `LD_LIBRARY_PATH` in order to run the installed executables. Android does support `RPATH` in later API levels, so this could change in the future. Refer to the previous section for information about `context-size` (very important!) and running other `examples`.

7
docs/backend/BLIS.md
View file

@ -27,13 +27,6 @@ We recommend using openmp since it's easier to modify the cores being used.
### llama.cpp compilation
Makefile:
```bash
make GGML_BLIS=1 -j
# make GGML_BLIS=1 llama-benchmark-matmult
```
CMake:
```bash

4
docs/backend/CANN.md
View file

@ -23,6 +23,8 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
## News
- 2024.11
- Support F16 and F32 data type model for Ascend 310P NPU.
- 2024.8
- Support `Q4_0` and `Q8_0` data type for Ascend NPU.
- 2024.7
@ -40,9 +42,11 @@ The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the abi
### Ascend NPU
**Verified devices**
| Ascend NPU | Status |
|:-----------------------------:|:-------:|
| Atlas 300T A2 | Support |
| Atlas 300I Duo | Support |
*Notes:*

123
docs/backend/SYCL.md
View file

@ -26,7 +26,7 @@
### Llama.cpp + SYCL
The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*).
The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
## Recommended Release
@ -34,13 +34,16 @@ The SYCL backend would be broken by some PRs due to no online CI.
The following release is verified with good quality:
|Commit ID|Tag|Release|Verified Platform|
|-|-|-|-|
|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
|Commit ID|Tag|Release|Verified Platform| Update date|
|-|-|-|-|-|
|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
## News
- 2024.11
- Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer.
- 2024.8
- Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
@ -111,10 +114,18 @@ SYCL backend supports Intel GPU Family:
**Verified devices**
| Nvidia GPU | Status | Verified Model |
|--------------------------|---------|----------------|
| Ampere Series | Support | A100, A4000 |
| Ampere Series *(Mobile)* | Support | RTX 40 Series |
| Nvidia GPU | Status | Verified Model |
|--------------------------|-----------|----------------|
| Ampere Series | Supported | A100, A4000 |
| Ampere Series *(Mobile)* | Supported | RTX 40 Series |
| AMD GPU | Status | Verified Model |
|--------------------------|--------------|----------------|
| Radeon Pro | Experimental | W6800 |
| Radeon RX | Experimental | 6700 XT |
Note: AMD GPU support is highly experimental and is incompatible with F16.
Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
## Docker
The docker build option is currently limited to *intel GPU* targets.
@ -186,6 +197,10 @@ Platform #0: Intel(R) OpenCL HD Graphics
In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
- **AMD GPU**
To target AMD GPUs with SYCL, the ROCm stack must be installed first.
2. **Install Intel® oneAPI Base toolkit**
- **For Intel GPU**
@ -212,6 +227,19 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
cmake --build buildWithCublas --config Release
```
- **Adding support to AMD GPUs**
**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
```sh
git clone https://github.com/oneapi-src/oneMKL
cd oneMKL
# Find your HIPTARGET with rocminfo, under the key 'Name:'
cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
cmake --build buildWithrocBLAS --config Release
```
3. **Verify installation and environment**
@ -223,22 +251,32 @@ sycl-ls
- **Intel GPU**
When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below:
When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
```
[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000]
[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50]
[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000]
[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50]
[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
```
- **Nvidia GPU**
Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
```
[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.12.0.12_195853.xmain-hotfix]
[opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
[ext_oneapi_cuda:gpu:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.2]
[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.12.0.12_195853.xmain-hotfix]
[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
```
- **AMD GPU**
For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
```
[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
```
### II. Build llama.cpp
@ -266,6 +304,7 @@ cmake --build build --config Release -j -v
```
#### Nvidia GPU
```sh
# Export relevant ENV variables
export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
@ -274,16 +313,37 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_
export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
# Build LLAMA with Nvidia BLAS acceleration through SYCL
# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
# Option 1: Use FP32 (recommended for better performance in most cases)
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# Option 2: Use FP16
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
# build all binary
cmake --build build --config Release -j -v
```
#### AMD GPU
```sh
# Export relevant ENV variables
export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
# Build LLAMA with rocBLAS acceleration through SYCL
## AMD
# Use FP32, FP16 is not supported
# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
# build all binary
cmake --build build --config Release -j -v
```
### III. Run the inference
@ -323,7 +383,7 @@ found 2 SYCL devices:
|Chosen Device ID|Setting|
|-|-|
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action|
|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
@ -336,12 +396,12 @@ Choose one of following methods to run.
- Use device 0:
```sh
./examples/sycl/run_llama2.sh 0
./examples/sycl/run-llama2.sh 0
```
- Use multiple devices:
```sh
./examples/sycl/run_llama2.sh
./examples/sycl/run-llama2.sh
```
2. Command line
@ -586,11 +646,12 @@ use 1 SYCL GPUs: [0] with Max compute units:512
#### Build
| Name | Value | Function |
|--------------------|-----------------------------------|---------------------------------------------|
| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
| GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA | Set the SYCL target device type. |
| GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. |
| Name | Value | Function |
|--------------------|---------------------------------------|---------------------------------------------|
| GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
| GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA \| AMD | Set the SYCL target device type. |
| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD) | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
| GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. |
| CMAKE_C_COMPILER | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path. |
| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)* | Set `icpx/icx` compiler for SYCL code path. |
@ -636,6 +697,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512
It's same for other projects including llama.cpp SYCL backend.
- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
Device Memory is not enough.
|Reason|Solution|
|-|-|
|Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
|Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
### **GitHub contribution**:
Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.

291
docs/build.md
View file

@ -7,124 +7,63 @@ git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
```
In order to build llama.cpp you have four different options.
The following sections describe how to build with different backends and options.
- Using `make`:
- On Linux or MacOS:
## CPU Build
```bash
make
```
Build llama.cpp using `CMake`:
- On Windows (x86/x64 only, arm64 requires cmake):
```bash
cmake -B build
cmake --build build --config Release
```
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Extract `w64devkit` on your pc.
3. Run `w64devkit.exe`.
4. Use the `cd` command to reach the `llama.cpp` folder.
5. From here you can run:
```bash
make
```
**Notes**:
- Notes:
- For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, run `make LLAMA_DEBUG=1`
- For faster compilation, add the `-j` argument to run multiple jobs in parallel, or use a generator that does this automatically such as Ninja. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/)
- For debug builds, there are two cases:
- Using `CMake`:
1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
```bash
cmake -B build
cmake --build build --config Release
```
```bash
cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build
```
**Notes**:
2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
- For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
- For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
- For faster repeated compilation, install [ccache](https://ccache.dev/).
- For debug builds, there are two cases:
```bash
cmake -B build -G "Xcode"
cmake --build build --config Debug
```
1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
For more details and a list of supported generators, see the [CMake documentation](https://cmake.org/cmake/help/latest/manual/cmake-generators.7.html).
```bash
cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build
```
2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
```bash
cmake -B build -G "Xcode"
cmake --build build --config Debug
```
- Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
- Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
- Tab Workload: Desktop-development with C++
- Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
- Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
- For Windows on ARM (arm64, WoA) build with:
```bash
cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
cmake --build build-arm64-windows-llvm-release
```
Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
- Using `gmake` (FreeBSD):
1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
2. Add your user to **video** group
3. Install compilation dependencies.
```bash
sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
```
## Metal Build
On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
argument.
- Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
- Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
- Tab Workload: Desktop-development with C++
- Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
- Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
- For Windows on ARM (arm64, WoA) build with:
```bash
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.
## BLAS Build
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Using BLAS doesn't affect the generation performance. There are currently several different BLAS implementations available for build and use:
### Accelerate Framework:
### Accelerate Framework
This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
### OpenBLAS:
### OpenBLAS
This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
- Using `make`:
- On Linux:
```bash
make GGML_OPENBLAS=1
```
- On Windows:
1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
3. Extract `w64devkit` on your pc.
4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
6. Run `w64devkit.exe`.
7. Use the `cd` command to reach the `llama.cpp` folder.
8. From here you can run:
```bash
make GGML_OPENBLAS=1
```
- Using `CMake` on Linux:
```bash
@ -136,14 +75,6 @@ This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS i
Check [BLIS.md](./backend/BLIS.md) for more information.
### SYCL
SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
### Intel oneMKL
Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
@ -161,16 +92,29 @@ Building through oneAPI compilers will make avx_vnni instruction set available f
Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
### CUDA
### Other BLAS libraries
This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
Any other BLAS library can be used by setting the `GGML_BLAS_VENDOR` option. See the [CMake documentation](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) for a list of supported vendors.
For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
## Metal Build
On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
To disable the Metal build at compile time use the `-DGGML_METAL=OFF` cmake option.
When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers 0` command-line argument.
## SYCL
SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
## CUDA
This provides GPU acceleration using an NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from the [NVIDIA developer site](https://developer.nvidia.com/cuda-downloads).
- Using `make`:
```bash
make GGML_CUDA=1
```
- Using `CMake`:
```bash
@ -186,22 +130,16 @@ The following compilation options are also available to tweak performance:
| Option | Legal values | Default | Description |
|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_FORCE_DMMV | Boolean | false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
| GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models |
| GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
| GGML_CUDA_FA_ALL_QUANTS | Boolean | false | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer. |
### MUSA
## MUSA
This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
- Using `make`:
```bash
make GGML_MUSA=1
```
- Using `CMake`:
```bash
@ -209,20 +147,22 @@ The following compilation options are also available to tweak performance:
cmake --build build --config Release
```
### hipBLAS
The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
This provides BLAS acceleration on HIP-supported AMD GPUs.
The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
## HIP
This provides GPU acceleration on HIP-supported AMD GPUs.
Make sure to have ROCm installed.
You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
- Using `make`:
```bash
make GGML_HIPBLAS=1
```
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build --config Release -- -j 16
```
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
@ -239,19 +179,14 @@ You can download it from your Linux distro's package manager or from here: [ROCm
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
- Using `make` (example for target gfx1030, build with 16 CPU threads):
```bash
make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
```
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake --build build
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
@ -260,23 +195,16 @@ You can download it from your Linux distro's package manager or from here: [ROCm
The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
| Option | Legal values | Default | Description |
|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| GGML_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
| GGML_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
| GGML_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
### Vulkan
## Vulkan
**Windows**
#### w64devkit
### w64devkit
Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
```sh
@ -292,18 +220,47 @@ Libs: -lvulkan-1
EOF
```
Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
#### MSYS2
Switch into the `llama.cpp` directory and build using CMake.
```sh
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
```
### Git Bash MINGW64
Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
Download and install [`Visual Studio Community Edition`](https://visualstudio.microsoft.com/) and make sure you select `C++`
Download and install [`CMake`](https://cmake.org/download/) with the default settings
Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
Go into your `llama.cpp` directory and right click, select `Open Git Bash Here` and then run the following commands
```
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
```
Now you can load the model in conversation mode using `Vulkan`
```sh
build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
```
### MSYS2
Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
```sh
pacman -S git \
mingw-w64-ucrt-x86_64-gcc \
mingw-w64-ucrt-x86_64-cmake \
mingw-w64-ucrt-x86_64-vulkan-devel \
mingw-w64-ucrt-x86_64-shaderc
```
Switch into `llama.cpp` directory and build using CMake.
```sh
pacman -S git \
mingw-w64-ucrt-x86_64-gcc \
mingw-w64-ucrt-x86_64-cmake \
mingw-w64-ucrt-x86_64-vulkan-devel \
mingw-w64-ucrt-x86_64-shaderc
```
Switch into the `llama.cpp` directory and build using CMake.
```sh
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
@ -352,7 +309,7 @@ cmake --build build --config Release
# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
```
### CANN
## CANN
This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
For more information about Ascend NPU in [Ascend Community](https://www.hiascend.com/en/).
@ -367,16 +324,26 @@ cmake --build build --config release
You can test with:
`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
```bash
llm_load_tensors: CANN buffer size = 13313.00 MiB
./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32
```
If the following info is output on screen, you are using `llama.cpp` with the CANN backend:
```bash
llm_load_tensors: CANN model buffer size = 13313.00 MiB
llama_new_context_with_model: CANN compute buffer size = 1260.81 MiB
```
For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
### Android
## Android
To read documentation for how to build on Android, [click here](./android.md)
## Notes about GPU-accelerated backends
The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.
In most cases, it is possible to build and use multiple backends at the same time. For example, you can build llama.cpp with both CUDA and Vulkan support by using the `-DGGML_CUDA=ON -DGGML_VULKAN=ON` options with CMake. At runtime, you can specify which backend devices to use with the `--device` option. To see a list of available devices, use the `--list-devices` option.
Backends can be built as dynamic libraries that can be loaded dynamically at runtime. This allows you to use the same llama.cpp binary on different machines with different GPUs. To enable this feature, use the `GGML_BACKEND_DL` option when building.

43
docs/docker.md
View file

@ -19,8 +19,11 @@ Additionally, there the following images, similar to the above:
- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
## Usage
@ -66,8 +69,8 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
The defaults are:
- `CUDA_VERSION` set to `11.7.1`
- `CUDA_DOCKER_ARCH` set to `all`
- `CUDA_VERSION` set to `12.6.0`
- `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
The resulting images, are essentially the same as the non-CUDA images:
@ -84,3 +87,37 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run
docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
```
## Docker With MUSA
Assuming one has the [mt-container-toolkit](https://developer.mthreads.com/musa/native) properly installed on Linux, `muBLAS` should be accessible inside the container.
## Building Docker locally
```bash
docker build -t local/llama.cpp:full-musa -f .devops/full-musa.Dockerfile .
docker build -t local/llama.cpp:light-musa -f .devops/llama-cli-musa.Dockerfile .
docker build -t local/llama.cpp:server-musa -f .devops/llama-server-musa.Dockerfile .
```
You may want to pass in some different `ARGS`, depending on the MUSA environment supported by your container host, as well as the GPU architecture.
The defaults are:
- `MUSA_VERSION` set to `rc3.1.0`
The resulting images, are essentially the same as the non-MUSA images:
1. `local/llama.cpp:full-musa`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
2. `local/llama.cpp:light-musa`: This image only includes the main executable file.
3. `local/llama.cpp:server-musa`: This image only includes the server executable file.
## Usage
After building locally, Usage is similar to the non-MUSA examples, but you'll need to set `mthreads` as default Docker runtime. This can be done by executing `(cd /usr/bin/musa && sudo ./docker setup $PWD)` and verifying the changes by executing `docker info | grep mthreads` on the host machine. You will also want to use the `--n-gpu-layers` flag.
```bash
docker run -v /path/to/models:/models local/llama.cpp:full-musa --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run -v /path/to/models:/models local/llama.cpp:light-musa -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
docker run -v /path/to/models:/models local/llama.cpp:server-musa -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
```

36
examples/CMakeLists.txt
View file

@ -6,21 +6,20 @@ find_package(Threads REQUIRED)
# ...
# flags
llama_add_compile_flags()
# examples
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN)
else()
add_subdirectory(cvector-generator)
add_subdirectory(baby-llama)
add_subdirectory(batched-bench)
add_subdirectory(batched)
add_subdirectory(benchmark)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(embedding)
add_subdirectory(eval-callback)
add_subdirectory(export-lora)
add_subdirectory(gbnf-validator)
add_subdirectory(gguf-hash)
add_subdirectory(gguf-split)
@ -29,27 +28,36 @@ else()
add_subdirectory(imatrix)
add_subdirectory(infill)
add_subdirectory(llama-bench)
add_subdirectory(llava)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(main)
add_subdirectory(parallel)
add_subdirectory(passkey)
add_subdirectory(perplexity)
add_subdirectory(quantize-stats)
add_subdirectory(quantize)
add_subdirectory(retrieval)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (LLAMA_BUILD_SERVER)
add_subdirectory(server)
endif()
if (GGML_SYCL)
add_subdirectory(sycl)
add_subdirectory(server)
endif()
add_subdirectory(save-load-state)
add_subdirectory(run)
add_subdirectory(simple)
add_subdirectory(simple-chat)
add_subdirectory(speculative)
add_subdirectory(speculative-simple)
add_subdirectory(tokenize)
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)
add_subdirectory(llava)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (GGML_SYCL)
add_subdirectory(sycl)
endif()
endif()
endif()

5
examples/baby-llama/CMakeLists.txt
View file

@ -1,5 +0,0 @@
set(TARGET llama-baby-llama)
add_executable(${TARGET} baby-llama.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

1639
examples/baby-llama/baby-llama.cpp
View file

File diff suppressed because it is too large Load diff

61
examples/base-translate.sh
View file

@ -1,61 +0,0 @@
#!/bin/bash
#
# Few-shot translation example.
# Requires a base model (i.e. no fine-tuned or instruct models).
#
# Usage:
#
# cd llama.cpp
# make -j
#
# ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
#
if [ $# -lt 2 ]; then
echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
exit 1
fi
eargs=""
if [ $# -gt 2 ]; then
eargs="${@:3}"
fi
ftmp="__llama.cpp_example_tmp__.txt"
trap "rm -f $ftmp" EXIT
echo "Translate from English to French:
===
sea otter, peppermint, plush girafe:
sea otter => loutre de mer
peppermint => menthe poivrée
plush girafe => girafe peluche
===
violin
violin => violon
===
phone, computer, mouse, keyboard:
phone => téléphone
computer => ordinateur
mouse => souris
keyboard => clavier
===
" > $ftmp
echo "$2
" >> $ftmp
model=$1
# generate the most likely continuation until the string "===" is found
./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

2
examples/batched-bench/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-batched-bench)
add_executable(${TARGET} batched-bench.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

9
examples/batched-bench/README.md
View file

@ -49,3 +49,12 @@ There are 2 modes of operation:
| 128 | 256 | 8 | 3072 | 0.751 | 1363.92 | 15.110 | 135.54 | 15.861 | 193.69 |
| 128 | 256 | 16 | 6144 | 1.569 | 1304.93 | 18.073 | 226.64 | 19.642 | 312.80 |
| 128 | 256 | 32 | 12288 | 3.409 | 1201.35 | 19.223 | 426.15 | 22.633 | 542.93 |
### JSONL output
Pass `--output-format jsonl` to output JSONL instead of Markdown, á la
```json lines
{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 1, "n_kv": 256, "t_pp": 0.233810, "speed_pp": 547.453064, "t_tg": 3.503684, "speed_tg": 36.532974, "t": 3.737494, "speed": 68.495094}
{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 2, "n_kv": 512, "t_pp": 0.422602, "speed_pp": 605.770935, "t_tg": 11.106112, "speed_tg": 23.050371, "t": 11.528713, "speed": 44.410854}
```

93
examples/batched-bench/batched-bench.cpp
View file

@ -1,49 +1,28 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
// mutates the input string
static std::vector<int> parse_list(char * p) {
std::vector<int> ret;
char * q = p;
while (*p) {
if (*p == ',') {
*p = '\0';
ret.push_back(std::atoi(q));
q = p + 1;
}
++p;
}
ret.push_back(std::atoi(q));
return ret;
}
static void print_usage(int argc, char ** argv, const gpt_params & params) {
gpt_params_print_usage(argc, argv, params);
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG_TEE("\n");
static void print_usage(int, char ** argv) {
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
if (!gpt_params_parse(argc, argv, params)) {
print_usage(argc, argv, params);
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
return 1;
}
common_init();
int is_pp_shared = params.is_pp_shared;
std::vector<int> n_pp = params.n_pp;
@ -57,7 +36,7 @@ int main(int argc, char ** argv) {
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
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);
@ -66,7 +45,7 @@ int main(int argc, char ** argv) {
return 1;
}
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
llama_context_params ctx_params = common_context_params_to_llama(params);
// ensure enough sequences are available
ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
@ -95,12 +74,11 @@ int main(int argc, char ** argv) {
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
@ -113,21 +91,22 @@ int main(int argc, char ** argv) {
// warm up
{
for (int i = 0; i < 16; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
common_batch_add(batch, 0, i, { 0 }, false);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
if (!params.batched_bench_output_jsonl) {
LOG("\n");
LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG("\n");
LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
}
for ( int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
for ( int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
@ -142,11 +121,11 @@ int main(int argc, char ** argv) {
continue;
}
llama_batch_clear(batch);
common_batch_clear(batch);
for (int i = 0; i < pp; ++i) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
llama_batch_add(batch, 0, i, { j }, false);
common_batch_add(batch, 0, i, { j }, false);
}
}
batch.logits[batch.n_tokens - 1] = true;
@ -156,7 +135,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_clear(ctx);
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -171,14 +150,14 @@ int main(int argc, char ** argv) {
const auto t_tg_start = ggml_time_us();
for (int i = 0; i < tg; ++i) {
llama_batch_clear(batch);
common_batch_clear(batch);
for (int j = 0; j < pl; ++j) {
llama_batch_add(batch, 0, pp + i, { j }, true);
common_batch_add(batch, 0, pp + i, { j }, true);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
@ -195,12 +174,22 @@ int main(int argc, char ** argv) {
const float speed_tg = pl*tg / t_tg;
const float speed = n_kv / t;
LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
if(params.batched_bench_output_jsonl) {
LOG(
"{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
"\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
);
} else {
LOG("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
}
}
}
}
llama_print_timings(ctx);
LOG("\n");
llama_perf_context_print(ctx);
llama_batch_free(batch);
@ -209,7 +198,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

51
examples/batched.swift/Sources/main.swift
View file

@ -27,7 +27,6 @@ guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), mo
print("Failed to load model")
exit(1)
}
defer {
llama_free_model(model)
}
@ -37,7 +36,6 @@ var tokens = tokenize(text: prompt, add_bos: true)
let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel)
var context_params = llama_context_default_params()
context_params.seed = 1234
context_params.n_ctx = n_kv_req
context_params.n_batch = UInt32(max(n_len, n_parallel))
context_params.n_threads = 8
@ -48,11 +46,26 @@ guard context != nil else {
print("Failed to initialize context")
exit(1)
}
defer {
llama_free(context)
}
var sparams = llama_sampler_chain_default_params()
let smpl = llama_sampler_chain_init(sparams)
guard smpl != nil else {
print("Failed to initialize sampling")
exit(1)
}
defer {
llama_sampler_free(smpl)
}
llama_sampler_chain_add(smpl, llama_sampler_init_top_k(40));
llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1));
llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.4));
llama_sampler_chain_add(smpl, llama_sampler_init_dist (1234));
let n_ctx = llama_n_ctx(context)
print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
@ -125,32 +138,7 @@ while n_cur <= n_len {
continue
}
var n_vocab = llama_n_vocab(model)
var logits = llama_get_logits_ith(context, i_batch[i])
var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab))
for token_id in 0 ..< n_vocab {
candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
}
var candidates_p: llama_token_data_array = .init(
data: &candidates,
size: candidates.count,
sorted: false
)
let top_k: Int32 = 40
let top_p: Float = 0.9
let temp: Float = 0.4
llama_sample_top_k(context, &candidates_p, top_k, 1)
llama_sample_top_p(context, &candidates_p, top_p, 1)
llama_sample_temp(context, &candidates_p, temp)
let new_token_id = llama_sample_token(context, &candidates_p)
// const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
// is it an end of stream? -> mark the stream as finished
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
@ -210,9 +198,10 @@ if n_parallel > 1 {
let t_main_end = ggml_time_us()
print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n")
print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
llama_print_timings(context)
llama_perf_sampler_print(smpl)
llama_perf_context_print(context)
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let utf8Count = text.utf8.count

Some files were not shown because too many files have changed in this diff Show more