CANN: Support Ascend310P to accelerate F32 and F16 Model (#10216)

* CANN Support Ascend310P to accelerate F32 and F16 Model

* Add compile option soc type macro ASCEND_310P to ggml-cann lib

* Remove unused code

* Remove the ascend soc_type hard code compile option in CMakelist.txt

.clang-format

@@ -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']
+...
+

.devops/full-cuda.Dockerfile

@@ -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
+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"]

.devops/full-musa.Dockerfile

@@ -0,0 +1,26 @@
+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
+
+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 . .
+
+RUN 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"]

.devops/full-rocm.Dockerfile

@@ -6,12 +6,12 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} as build
+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

.devops/full.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1

.devops/llama-cli-cann.Dockerfile

@@ -0,0 +1,44 @@
+ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
+
+FROM ascendai/cann:$ASCEND_VERSION AS build
+
+WORKDIR /app
+
+COPY . .
+
+RUN yum install -y gcc g++ cmake make
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
+ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
+ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
+ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
+
+# find libascend_hal.so, because the drive hasn`t been mounted.
+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_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake --build build --config Release --target llama-cli
+
+# TODO: use image with NNRT
+FROM ascendai/cann:$ASCEND_VERSION AS runtime
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENV LC_ALL=C.utf8
+
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
+ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
+ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
+ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
+
+ENTRYPOINT ["/llama-cli" ]

.devops/llama-cli-cuda.Dockerfile

@@ -1,35 +1,38 @@
 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
 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} as build
+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=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
+# Use the default CUDA archs if not specified
+RUN if [ "${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 \;
 
-RUN make -j$(nproc) llama-cli
-
-FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+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" ]

.devops/llama-cli-intel.Dockerfile

@@ -1,6 +1,6 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
@@ -14,10 +14,12 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         echo "GGML_SYCL_F16 is set" && \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
-    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    echo "Building with static libs" && \
+    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
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
 
 COPY --from=build /app/build/bin/llama-cli /llama-cli
 

.devops/llama-cli-musa.Dockerfile

@@ -0,0 +1,31 @@
+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
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+RUN 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" ]

.devops/llama-cli-rocm.Dockerfile

@@ -6,12 +6,12 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} as build
+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

.devops/llama-cli-vulkan.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=jammy
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget libgomp1
@@ -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

.devops/llama-cli.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
     apt-get install -y build-essential git
@@ -11,7 +11,7 @@ COPY . .
 
 RUN make -j$(nproc) llama-cli
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1

.devops/llama-server-cuda.Dockerfile

@@ -1,38 +1,42 @@
 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
 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} as build
+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=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
-# Enable cURL
-ENV LLAMA_CURL=1
+# Use the default CUDA archs if not specified
+RUN if [ "${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 \;
 
-RUN make -j$(nproc) llama-server
-
-FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+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" ]
 

.devops/llama-server-intel.Dockerfile

@@ -1,6 +1,6 @@
-ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
+ARG ONEAPI_VERSION=2025.0.0-0-devel-ubuntu22.04
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
@@ -14,10 +14,11 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         echo "GGML_SYCL_F16 is set" && \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
-    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
+    echo "Building with dynamic libs" && \
+    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:$ONEAPI_VERSION as runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev curl
@@ -25,6 +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" ]
 

.devops/llama-server-musa.Dockerfile

@@ -0,0 +1,36 @@
+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
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN 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" ]

.devops/llama-server-rocm.Dockerfile

@@ -6,12 +6,12 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} as build
+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

.devops/llama-server-vulkan.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=jammy
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget
@@ -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" ]
 

.devops/llama-server.Dockerfile

@@ -1,9 +1,9 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev curl
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -13,14 +13,16 @@ ENV LLAMA_CURL=1
 
 RUN make -j$(nproc) llama-server
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM ubuntu:$UBUNTU_VERSION AS runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
 COPY --from=build /app/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" ]
 

.devops/nix/apps.nix

@@ -10,7 +10,6 @@
             "llama-embedding"
             "llama-server"
             "llama-quantize"
-            "llama-train-text-from-scratch"
           ];
           mkApp = name: {
             type = "app";

.devops/nix/devshells.nix

@@ -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))
+        ];
     };
 }

.devops/nix/nixpkgs-instances.nix

@@ -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 {

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

@@ -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 ];
+  };
+}

.devops/nix/package.nix

@@ -3,33 +3,35 @@
   glibc,
   config,
   stdenv,
-  mkShell,
   runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
-  python3,
   mpi,
   blas,
   cudaPackages,
+  autoAddDriverRunpath,
   darwin,
   rocmPackages,
   vulkan-headers,
   vulkan-loader,
-  clblast,
-  useBlas ? builtins.all (x: !x) [
-    useCuda
-    useMetalKit
-    useOpenCL
-    useRocm
-    useVulkan
-  ] && blas.meta.available,
+  curl,
+  shaderc,
+  useBlas ?
+    builtins.all (x: !x) [
+      useCuda
+      useMetalKit
+      useRocm
+      useVulkan
+    ]
+    && blas.meta.available,
   useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
-  useMpi ? false, # Increases the runtime closure size by ~700M
-  useOpenCL ? false,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
+  # Increases the runtime closure size by ~700M
+  useMpi ? false,
   useRocm ? config.rocmSupport,
+  enableCurl ? true,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
 
@@ -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";
@@ -56,45 +57,17 @@ let
     ++ lib.optionals useCuda [ "CUDA" ]
     ++ lib.optionals useMetalKit [ "MetalKit" ]
     ++ lib.optionals useMpi [ "MPI" ]
-    ++ lib.optionals useOpenCL [ "OpenCL" ]
     ++ lib.optionals useRocm [ "ROCm" ]
     ++ lib.optionals useVulkan [ "Vulkan" ];
 
   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
-    ]
-  );
-
-  xcrunHost = runCommand "xcrunHost" {} ''
+  xcrunHost = runCommand "xcrunHost" { } ''
     mkdir -p $out/bin
     ln -s /usr/bin/xcrun $out/bin
   '';
@@ -111,16 +84,9 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # <nv/target>
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
+    cuda_cudart
+    cuda_cccl # <nv/target>
+    libcublas
   ];
 
   rocmBuildInputs = with rocmPackages; [
@@ -132,185 +98,149 @@ let
   vulkanBuildInputs = [
     vulkan-headers
     vulkan-loader
+    shaderc
   ];
 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 useOpenCL [ clblast ]
-      ++ optionals useRocm rocmBuildInputs
-      ++ optionals useBlas [ blas ]
-      ++ optionals useVulkan vulkanBuildInputs;
+  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 "GGML_NATIVE" false)
-        (cmakeBool "GGML_BLAS" useBlas)
-        (cmakeBool "GGML_CLBLAST" useOpenCL)
-        (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
-        useOpenCL
-        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 || useOpenCL) 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;
+  };
+})

.devops/nix/python-scripts.nix

@@ -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
+    behave
+    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;
+})

.devops/nix/scope.nix

@@ -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 { };
+})

.devops/tools.sh

@@ -8,13 +8,11 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert-hf-to-gguf.py "$@"
+    python3 ./convert_hf_to_gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     ./llama-cli "$@"
-elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
-    ./llama-finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -36,8 +34,6 @@ else
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
-    echo "  --finetune (-f): Run finetune command to create a lora finetune of the model"
-    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"

.dockerignore

@@ -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/

.ecrc

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

.editorconfig

@@ -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
 

.github/ISSUE_TEMPLATE/config.yml

@@ -9,5 +9,3 @@ contact_links:
   - name: Want to contribute?
     url: https://github.com/ggerganov/llama.cpp/wiki/contribute
     about: Head to the contribution guide page of the wiki for areas you can help with
-
-

.github/labeler.yml

@@ -16,7 +16,9 @@ SYCL:
         - any-glob-to-any-file:
             - ggml/include/ggml-sycl.h
             - ggml/src/ggml-sycl.cpp
-            - README-sycl.md
+            - ggml/src/ggml-sycl/**
+            - docs/backend/SYCL.md
+            - examples/sycl/**
 Nvidia GPU:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/bench.yml.disabled

@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark
 
@@ -24,10 +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 * * *'
 
@@ -129,6 +132,8 @@ jobs:
 
       - name: Server bench
         id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -137,7 +142,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -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 -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 -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
@@ -222,7 +241,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -355,8 +374,10 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential vulkan-sdk
 
       - name: Build
         id: cmake_build
@@ -373,7 +394,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -384,22 +405,43 @@ 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
@@ -440,7 +482,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -544,7 +586,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -559,6 +601,7 @@ 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 \
@@ -574,7 +617,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -589,6 +632,7 @@ 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 \
@@ -608,7 +652,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -694,22 +738,20 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
@@ -726,7 +768,7 @@ 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
@@ -857,7 +899,8 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
+          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
+          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
@@ -908,7 +951,7 @@ 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_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
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
@@ -952,6 +995,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.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/*
 
@@ -963,19 +1007,20 @@ 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"
@@ -990,8 +1035,72 @@ 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 -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
+
+      - 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"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - 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 -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
@@ -1056,6 +1165,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
+      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

.github/workflows/close-issue.yml

@@ -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

.github/workflows/docker.yml

@@ -15,11 +15,17 @@ on:
     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, useful for debugging
 
 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
@@ -37,15 +43,20 @@ 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 +71,34 @@ 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 +116,13 @@ 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)
+      - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
-        uses: docker/build-push-action@v4
+        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 }}

.github/workflows/nix-ci-aarch64.yml

@@ -21,6 +21,13 @@ 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:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -12,6 +12,13 @@ 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:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-eval:
     strategy:

.github/workflows/python-check-requirements.yml

@@ -6,15 +6,13 @@ on:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
+      - '**/requirements*.txt'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

.github/workflows/python-type-check.yml

@@ -0,0 +1,40 @@
+name: Python Type-Check
+
+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'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  python-type-check:
+    runs-on: ubuntu-latest
+    name: pyright type-check
+    steps:
+      - name: Check out source repository
+        uses: actions/checkout@v4
+      - name: Set up Python environment
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Python dependencies
+        # TODO: use a venv
+        run: pip install -r requirements/requirements-all.txt
+      - name: Type-check with Pyright
+        uses: jakebailey/pyright-action@v2
+        with:
+          version: 1.1.382
+          level: warning
+          warnings: true

.github/workflows/server.yml

@@ -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
@@ -173,6 +179,7 @@ jobs:
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
+          $env:PYTHONIOENCODING = ":replace"
           behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests

.gitignore

@@ -3,6 +3,7 @@
 *.a
 *.bat
 *.bin
+*.d
 *.dll
 *.dot
 *.etag
@@ -47,8 +48,10 @@ build*
 !build-info.cpp.in
 !build-info.sh
 !build.zig
+!docs/build.md
 /libllama.so
 /llama-*
+/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -59,6 +62,12 @@ llama-batched-swift
 /rpc-server
 out/
 tmp/
+autogen-*.md
+
+# Deprecated
+
+/main
+/server
 
 # CI
 
@@ -72,7 +81,6 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
-
 zig-out/
 zig-cache/
 
@@ -98,13 +106,14 @@ examples/server/*.mjs.hpp
 
 # Python
 
-__pycache__
-.venv
-/Pipfile
-dist
-poetry.lock
+/.venv
+__pycache__/
+*/poetry.lock
 poetry.toml
 
+# Nix
+/result
+
 # Test binaries
 /tests/test-backend-ops
 /tests/test-double-float
@@ -122,3 +131,10 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
+
+# Test models for lora adapters
+/lora-tests
+
+# Local scripts
+/run-vim.sh
+/run-chat.sh

.gitmodules

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

CMakeLists.txt

@@ -42,13 +42,21 @@ endif()
 
 option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 
+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
 #
 
-# general
-option(LLAMA_CCACHE "llama: use ccache if available" ON)
-
 # debug
 option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
 option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
@@ -61,6 +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"       ${LLAMA_STANDALONE})
@@ -73,7 +84,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
 # override ggml options
-set(GGML_CCACHE             ${LLAMA_CCACHE})
 set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
 set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
 set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
@@ -82,11 +92,15 @@ 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_AMX)
+    set(GGML_AMX ON)
+endif()
+
+if (NOT DEFINED GGML_CUDA_GRAPHS)
+    set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()
 
 # transition helpers
@@ -106,12 +120,16 @@ llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
 llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
 llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
+llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
 #
 # build the library
 #
 
-add_subdirectory(ggml)
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+    # ... otherwise assume ggml is added by a parent CMakeLists.txt
+endif()
 add_subdirectory(src)
 
 #
@@ -129,7 +147,22 @@ 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")
 
-get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+# 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)
+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)
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
@@ -152,7 +185,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
 
 install(
-    FILES convert-hf-to-gguf.py
+    FILES convert_hf_to_gguf.py
     PERMISSIONS
         OWNER_READ
         OWNER_WRITE
@@ -171,17 +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()

CMakePresets.json

@@ -19,19 +19,22 @@
         "cacheVariables": {
             "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
             "CMAKE_CXX_COMPILER": "icx",
+            "CMAKE_C_COMPILER": "cl",
             "GGML_SYCL": "ON",
             "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": "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"
         }
@@ -39,26 +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-release", "inherits": [ "sycl-base", "release" ] }
+    { "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-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
+    { "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
   ]
 }

CONTRIBUTING.md

@@ -1,14 +1,33 @@
-# Contributing Guidelines
+# Pull requests (for contributors)
 
-## Checklist
+- 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
+- Optionally 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
+- 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
 
-* Make sure your PR follows the [coding guidelines](https://github.com/ggerganov/llama.cpp/blob/master/README.md#coding-guidelines)
-* 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
+# Pull requests (for collaborators)
 
-## PR formatting
+- 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
 
-* 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 `[X]` for your conveience. Refer to [About task lists](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) for more information.
-* If the pull request only contains documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times.
-* When squashing multiple commits on merge, use the following format for your commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : Fix typo in utils.py (#1234)`
+# 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
+- 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)
+- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+
+![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

Package.swift

@@ -4,11 +4,21 @@ import PackageDescription
 
 var sources = [
     "src/llama.cpp",
+    "src/llama-vocab.cpp",
+    "src/llama-grammar.cpp",
+    "src/llama-sampling.cpp",
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
+    "ggml/src/ggml-aarch64.c",
     "ggml/src/ggml-alloc.c",
-    "ggml/src/ggml-backend.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",
 ]
 
@@ -17,6 +27,7 @@ var linkerSettings: [LinkerSetting] = []
 var cSettings: [CSetting] =  [
     .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
     .unsafeFlags(["-fno-objc-arc"]),
+    .headerSearchPath("ggml/src"),
     // 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)
@@ -25,8 +36,9 @@ var cSettings: [CSetting] =  [
 ]
 
 #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: [
@@ -56,13 +68,15 @@ 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,

README.md

@@ -3,73 +3,26 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![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?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![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
 
-### Recent API changes
+- [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)
 
-- [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
+## Hot topics
 
-### 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)
 
 ----
 
-<details>
-  <summary>Table of Contents</summary>
-  <ol>
-    <li>
-      <a href="#description">Description</a>
-    </li>
-    <li>
-      <a href="#usage">Usage</a>
-      <ul>
-        <li><a href="#get-the-code">Get the Code</a></li>
-        <li><a href="#build">Build</a></li>
-        <li><a href="#blas-build">BLAS Build</a></li>
-        <li><a href="#prepare-and-quantize">Prepare and Quantize</a></li>
-        <li><a href="#run-the-quantized-model">Run the quantized model</a></li>
-        <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
-        <li><a href="#quantization">Quantization</a></li>
-        <li><a href="#interactive-mode">Interactive mode</a></li>
-        <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
-        <li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
-        <li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
-        <li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
-        <li><a href="#android">Android</a></li>
-        <li><a href="#docker">Docker</a></li>
-      </ul>
-    </li>
-    <li><a href="#contributing">Contributing</a></li>
-    <li><a href="#coding-guidelines">Coding guidelines</a></li>
-    <li><a href="#docs">Docs</a></li>
-  </ol>
-</details>
-
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -77,9 +30,9 @@ variety 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
 
@@ -87,14 +40,6 @@ Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomm
 improved significantly thanks to many contributions. It is the main playground for developing new features for the
 [ggml](https://github.com/ggerganov/ggml) library.
 
-**Supported platforms:**
-
-- [X] Mac OS
-- [X] Linux
-- [X] Windows (via CMake)
-- [X] Docker
-- [X] FreeBSD
-
 **Supported models:**
 
 Typically finetunes of the base models below are supported as well.
@@ -108,6 +53,7 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
@@ -133,9 +79,24 @@ 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] [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)
+- [x] [Smaug](https://huggingface.co/models?search=Smaug)
+- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
+- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
+- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
+- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
+- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
+- [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/HOWTO-add-model.md))
+(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
 **Multimodal models:**
 
@@ -149,18 +110,13 @@ 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)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) 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.
-
-[simplechat](./examples/server/public_simplechat) is a simple chat client, which can be used to chat with the model exposed using above web server (use --path to point to simplechat), from a local web browser.
-
 **Bindings:**
 
 - 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)
@@ -168,24 +124,31 @@ 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:**
 
 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)
@@ -211,16 +174,33 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [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)
+- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 **Tools:**
 
 - [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 prebuild Mobile and Web platform wrappers and a model example)
 
----
+**Infrastructure:**
 
-Here is a typical run using LLaMA v2 13B on M2 Ultra:
+- [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
+
+**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
@@ -300,454 +280,85 @@ llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms
 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.
 
-### Get the Code
+### 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
-git clone https://github.com/ggerganov/llama.cpp
-cd llama.cpp
+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.
 ```
 
-### Build
+See [this page](./examples/main/README.md) for a full list of parameters.
 
-In order to build llama.cpp you have four different options.
+### Conversation mode
 
-- Using `make`:
-  - On Linux or MacOS:
-
-      ```bash
-      make
-      ```
-
-  - On Windows:
-
-    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:
-    - 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`
-
-- Using `CMake`:
-
-  ```bash
-  cmake -B build
-  cmake --build build --config Release
-  ```
-
-  **Notes**:
-
-    - 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:
-
-      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
-
-      ```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
-      ```
-
--   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
-        ```
-
-### Homebrew
-
-On Mac and Linux, the homebrew package manager can be used via
-```
-brew install llama.cpp
-```
-The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
-
-### Nix
-
-On Mac and Linux, the Nix package manager can be used via
-```
-nix profile install nixpkgs#llama-cpp
-```
-For flake enabled installs.
-
-Or
-```
-nix-env --file '<nixpkgs>' --install --attr llama-cpp
-```
-For non-flake enabled installs.
-
-This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-#### Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-```
-flox install llama-cpp
-```
-Flox follows the nixpkgs build of llama.cpp.
-
-### 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.
-
-### 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:
-
-- #### Accelerate Framework:
-
-  This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
-
-- #### 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
-      cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
-      cmake --build build --config Release
-      ```
-
-- #### BLIS
-
-  Check [BLIS.md](docs/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](README-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](./README-sycl.md).
-
-  - Using manual oneAPI installation:
-    By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
-      ```bash
-      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
-      cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
-      cmake --build build --config Release
-      ```
-
-  - Using oneAPI docker image:
-    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
-
-  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
-
-  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).
-
-  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.
-
-  - Using `make`:
-    ```bash
-    make GGML_CUDA=1
-    ```
-  - Using `CMake`:
-
-    ```bash
-    cmake -B build -DGGML_CUDA=ON
-    cmake --build build --config Release
-    ```
-
-  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. 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.                                                                                                  |
-
-- #### hipBLAS
-
-  This provides BLAS 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 --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`.
-    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
-    Note that if you get the following error:
-    ```
-    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
-    ```
-    Try searching for a directory under `HIP_PATH` that contains the file
-    `oclc_abi_version_400.bc`. Then, add the following to the start of the
-    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
-    like:
-    ```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 --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 --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)
-    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
-
-
-  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
-
-  **With docker**:
-
-  You don't need to install Vulkan SDK. It will be installed inside the container.
-
-  ```sh
-  # Build the image
-  docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
-
-  # Then, use it:
-  docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
-  ```
-
-  **Without docker**:
-
-  Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
-
-  For example, on Ubuntu 22.04 (jammy), use the command below:
-
-  ```bash
-  wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
-  wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-  apt update -y
-  apt-get install -y vulkan-sdk
-  # To verify the installation, use the command below:
-  vulkaninfo
-  ```
-
-  Alternatively your package manager might be able to provide the appropriate libraries.
-  For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-  For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
-
-  Then, build llama.cpp using the cmake command below:
-
-  ```bash
-  cmake -B build -DGGML_VULKAN=1
-  cmake --build build --config Release
-  # Test the output binary (with "-ngl 33" to offload all layers to GPU)
-  ./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
-
-  # You should see in the output, ggml_vulkan detected your GPU. For example:
-  # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
-  ```
-
-### Prepare and Quantize
-
-> [!NOTE]
-> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
-
-To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
-
-Note: `convert.py` 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.
+If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
 
 ```bash
-# obtain the official LLaMA model weights and place them in ./models
-ls ./models
-llama-2-7b tokenizer_checklist.chk tokenizer.model
-# [Optional] for models using BPE tokenizers
-ls ./models
-<folder containing weights and tokenizer json> vocab.json
-# [Optional] for PyTorch .bin models like Mistral-7B
-ls ./models
-<folder containing weights and tokenizer json>
+llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
 
-# install Python dependencies
-python3 -m pip install -r requirements.txt
-
-# convert the model to ggml FP16 format
-python3 convert-hf-to-gguf.py models/mymodel/
-
-# quantize the model to 4-bits (using Q4_K_M method)
-./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
-
-# update the gguf filetype to current version if older version is now unsupported
-./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+# 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!
 ```
 
-### Run the quantized model
+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
-# start inference on a gguf model
-./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
+./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
 ```
 
-When running the larger models, make sure you have enough disk space to store all the intermediate files.
+You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
 
-### Running on Windows with prebuilt binaries
-
-You will find prebuilt Windows binaries on the release page.
-
-Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
-
-From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
-
-```
-.\main -m llama-2-7b.Q4_0.gguf -n 128
+```bash
+./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
 ```
 
-### Memory/Disk Requirements
+### Web server
 
-As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
+[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.
 
-| Model | Original size | Quantized size (Q4_0) |
-|------:|--------------:|----------------------:|
-|    7B |         13 GB |                3.9 GB |
-|   13B |         24 GB |                7.8 GB |
-|   30B |         60 GB |               19.5 GB |
-|   65B |        120 GB |               38.5 GB |
+Example usage:
 
-### Quantization
+```bash
+./llama-server -m your_model.gguf --port 8080
 
-Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
-
-*(outdated)*
-
-| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
-|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
-|    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
-|    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |
-|    7B | ms/tok @ 4th |    127 |     55 |     54 |     76 |     83 |     72 |
-|    7B | ms/tok @ 8th |    122 |     43 |     45 |     52 |     56 |     67 |
-|    7B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
-|   13B | perplexity   | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
-|   13B | file size    |  25.0G |   6.8G |   7.6G |   8.3G |   9.1G |    13G |
-|   13B | ms/tok @ 4th |      - |    103 |    105 |    148 |    160 |    131 |
-|   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
-|   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
-
-- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
-- recent k-quants improvements and new i-quants
-  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
-  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
-  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
-  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
-  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
-  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
-  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
-  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
-  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
-  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
-  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
-  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
-  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
-  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
-
-### Perplexity (measuring model quality)
-
-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).
-
-The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
-The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
-
-#### How to run
-
-1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-2. Run `./llama-perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
-3. Output:
+# Basic web UI can be accessed via browser: http://localhost:8080
+# Chat completion endpoint: http://localhost:8080/v1/chat/completions
 ```
-perplexity : calculating perplexity over 655 chunks
-24.43 seconds per pass - ETA 4.45 hours
-[1]4.5970,[2]5.1807,[3]6.0382,...
-```
-And after 4.45 hours, you will have the final perplexity.
 
 ### Interactive mode
 
-If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
+> [!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
@@ -798,18 +409,72 @@ The `grammars/` folder contains a handful of sample grammars. To write your own,
 
 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.
 
-### Obtaining and using the Facebook LLaMA 2 model
+## Build
 
-- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
-- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
-  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
-  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
-  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
-  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
-  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
-  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
+Please refer to [Build llama.cpp locally](./docs/build.md)
 
-### Seminal papers and background on the models
+## 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 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
+
+### Prepare and Quantize
+
+> [!NOTE]
+> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
+
+To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
+
+Note: `convert.py` 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.
+
+To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
+
+### Perplexity (measuring model quality)
+
+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).
+
+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, 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 documentation
+
+- [main (cli)](./examples/main/README.md)
+- [server](./examples/server/README.md)
+- [jeopardy](./examples/jeopardy/README.md)
+- [GBNF grammars](./grammars/README.md)
+
+**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)
+- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+
+**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:
@@ -820,178 +485,3 @@ 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)
-
-### 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 ~/
-```
-
-[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
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
-```
-
-Here's a demo of an interactive session running on Pixel 5 phone:
-
-https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
-
-### Docker
-
-#### Prerequisites
-* Docker must be installed and running on your system.
-* Create a folder to store big models & intermediate files (ex. /llama/models)
-
-#### Images
-We have three Docker images available for this project:
-
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
-
-Additionally, there the following images, similar to the above:
-
-- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `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`)
-
-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).
-
-#### Usage
-
-The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
-
-Replace `/path/to/models` below with the actual path where you downloaded the models.
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
-```
-
-On completion, you are ready to play!
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a light image:
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a server image:
-
-```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
-```
-
-### Docker With CUDA
-
-Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
-
-#### Building Locally
-
-```bash
-docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
-docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
-```
-
-You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
-
-The defaults are:
-
-- `CUDA_VERSION` set to `11.7.1`
-- `CUDA_DOCKER_ARCH` set to `all`
-
-The resulting images, are essentially the same as the non-CUDA images:
-
-1. `local/llama.cpp:full-cuda`: 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-cuda`: This image only includes the main executable file.
-3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
-
-#### Usage
-
-After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
-
-```bash
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --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 --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
-```
-
-### 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!
-- 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)
-
-### 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
-- 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`
-- 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
-- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
-
-![matmul](media/matmul.png)
-
-### Docs
-
-- [main (cli)](./examples/main/README.md)
-- [server](./examples/server/README.md)
-- [jeopardy](./examples/jeopardy/README.md)
-- [BLIS](./docs/BLIS.md)
-- [Performance troubleshooting](./docs/token_generation_performance_tips.md)
-- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
-- [GBNF grammars](./grammars/README.md)

ci/run.sh

@@ -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
 
@@ -103,8 +110,11 @@ function gg_run_ctest_debug {
 
     set -e
 
+    # Check cmake, make and ctest are installed
+    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
 
@@ -131,8 +141,11 @@ function gg_run_ctest_release {
 
     set -e
 
+    # Check cmake, make and ctest are installed
+    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
@@ -260,7 +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}
@@ -284,10 +296,10 @@ 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
+    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -314,36 +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"
@@ -419,9 +431,9 @@ 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
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -448,34 +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"
@@ -529,7 +541,6 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
-# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -550,10 +561,10 @@ 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
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -580,36 +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"
@@ -686,17 +697,17 @@ 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
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
 
     ./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
 }
@@ -710,8 +721,92 @@ function gg_sum_embd_bge_small {
     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'
+    fi
+
+    if ! command -v make &> /dev/null; then
+        gg_printf 'make not found, please install'
+    fi
+
+    if ! command -v ctest &> /dev/null; then
+        gg_printf 'ctest not found, please install'
+    fi
+}
+
 ## 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
@@ -734,6 +829,7 @@ 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
@@ -741,7 +837,7 @@ 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

cmake/arm64-apple-clang.cmake

@@ -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}" )

cmake/llama-config.cmake.in

@@ -3,11 +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_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_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@
 
@@ -15,47 +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)
-endif()
+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()
 
-if (GGML_BLAS)
-    find_package(BLAS REQUIRED)
-endif()
+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")
 
-if (GGML_CUDA)
-    find_package(CUDAToolkit REQUIRED)
-endif()
+    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 (GGML_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK Metal REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
-endif()
+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_HIPBLAS)
-    find_package(hip REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
+    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()
 
 find_library(llama_LIBRARY llama
     REQUIRED
-    HINTS ${LLAMA_LIB_DIR})
-
-set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
-set(_llama_transient_defines "@LLAMA_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}"

common/CMakeLists.txt

@@ -51,21 +51,21 @@ 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
     )
 
 if (BUILD_SHARED_LIBS)

common/arg.h

@@ -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);

common/common.h

@@ -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 '\\'
@@ -33,40 +24,138 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
+struct common_lora_adapter_info {
+    std::string path;
+    float scale;
+};
+
+struct common_lora_adapter_container : common_lora_adapter_info {
+    struct llama_lora_adapter * adapter;
+};
+
 // 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
+// sampler parameters
+struct common_sampler_params {
+    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
+
+    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 {
     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
@@ -89,7 +178,12 @@ 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
+
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+    struct cpu_params draft_cpuparams;
+    struct cpu_params draft_cpuparams_batch;
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
@@ -99,36 +193,35 @@ struct gpt_params {
     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_sampler_params sparams;
 
-    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_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_draft          = ""; // draft model for speculative decoding                          // 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;
 
-    // TODO: avoid tuple, use struct
-    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
-    std::string lora_base  = "";                              // base model path for the lora adapter
+    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<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
@@ -163,15 +256,15 @@ 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
@@ -181,33 +274,37 @@ 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;
@@ -253,28 +350,49 @@ struct gpt_params {
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     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_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();
 
+void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
+
 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;
@@ -287,9 +405,30 @@ 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
 //
@@ -304,20 +443,29 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-// TODO: avoid tuplue, use struct
-std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct common_init_result {
+    struct llama_model   * model   = nullptr;
+    struct llama_context * context = nullptr;
+    std::vector<common_lora_adapter_container> lora_adapters;
+};
 
-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 common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, 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 struct llama_model_params & params);
+struct llama_model_params     common_model_params_to_llama  (const 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 * common_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 * common_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
+
+// clear LoRA adapters from context, then apply new list of 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,
@@ -330,13 +478,13 @@ void llama_batch_add(
 
 // 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,
@@ -344,61 +492,49 @@ 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);
 
-// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
-//       that takes into account the tokenizer type and decides how to handle the leading space
-//
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
-// removes the leading space from the first non-BOS token
-std::string llama_detokenize_spm(
+// optionally renders special/control tokens
+std::string common_detokenize(
                          llama_context * ctx,
-        const std::vector<llama_token> & tokens);
-
-// detokenizes a vector of tokens into a string
-// should work similar to Python's `tokenizer.decode`
-std::string llama_detokenize_bpe(
-                         llama_context * ctx,
-        const std::vector<llama_token> & tokens);
-
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
+        const std::vector<llama_token> & tokens,
+                                  bool   special = true);
 
 //
 // Chat template utils
 //
 
 // 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);
 
 //
@@ -406,31 +542,31 @@ 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;
@@ -438,7 +574,7 @@ 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
@@ -447,16 +583,3 @@ 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);
-

common/console.cpp

@@ -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) {

common/grammar-parser.cpp

@@ -1,536 +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) {
-                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;
-    }
-}

common/grammar-parser.h

@@ -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);
-}

common/json-schema-to-grammar.cpp

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

common/log.cpp

@@ -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);
+}

common/log.h

@@ -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__)

common/ngram-cache.cpp

@@ -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;

common/ngram-cache.h

@@ -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;
@@ -37,21 +37,28 @@ struct llama_ngram {
     }
 };
 
-struct llama_ngram_hash_function {
-    size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = 0;
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
+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 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 ^= 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.
@@ -63,8 +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.
@@ -74,21 +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);

common/sampling.cpp

@@ -1,459 +1,466 @@
-#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_sampler_params 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_sampler_params::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_sampler_params & 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;
+}
+
+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;
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
-
-    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) {
-        // 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_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
-
-        // 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 + llama_n_vocab(llama_get_model(ctx_main))};
-    }
-
-    // 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.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(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_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
+    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_main, ctx_sampling->grammar, id);
-    }
+    return samplers;
 }

common/sampling.h

@@ -2,159 +2,82 @@
 
 #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_sampler_params & 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);
+
+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);

common/train.h

@@ -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);

convert_hf_to_gguf_update.py

@@ -2,7 +2,7 @@
 # -*- coding: utf-8 -*-
 
 # This script downloads the tokenizer models of the specified models from Huggingface and
-# generates the get_vocab_base_pre() function for convert-hf-to-gguf.py
+# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
 #
 # This is necessary in order to analyze the type of pre-tokenizer used by the model and
 # provide the necessary information to llama.cpp via the GGUF header in order to implement
@@ -15,9 +15,9 @@
 # - Add a new model to the "models" list
 # - Run the script with your huggingface token:
 #
-#   python3 convert-hf-to-gguf-update.py <huggingface_token>
+#   python3 convert_hf_to_gguf_update.py <huggingface_token>
 #
-# - Copy-paste the generated get_vocab_base_pre() function into convert-hf-to-gguf.py
+# - Copy-paste the generated get_vocab_base_pre() function into convert_hf_to_gguf.py
 # - Update llama.cpp with the new pre-tokenizer if necessary
 #
 # TODO: generate tokenizer tests for llama.cpp
@@ -31,13 +31,14 @@ import re
 import requests
 import sys
 import json
+import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
 from transformers import AutoTokenizer
 
 logging.basicConfig(level=logging.DEBUG)
-logger = logging.getLogger("convert-hf-to-gguf-update")
+logger = logging.getLogger("convert_hf_to_gguf_update")
 sess = requests.Session()
 
 
@@ -45,20 +46,21 @@ class TOKENIZER_TYPE(IntEnum):
     SPM = auto()
     BPE = auto()
     WPM = auto()
+    UGM = auto()
 
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
-chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
+CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
 
 if len(sys.argv) == 2:
     token = sys.argv[1]
     if not token.startswith("hf_"):
         logger.info("Huggingface token seems invalid")
-        logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
+        logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
         sys.exit(1)
 else:
-    logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
+    logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
     sys.exit(1)
 
 # TODO: add models here, base models preferred
@@ -70,6 +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", },
@@ -79,6 +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", },
@@ -86,6 +90,18 @@ models = [
     {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
     {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
     {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
+    {"name": "gemma",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
+    {"name": "gemma-2",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
+    {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "t5",             "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
+    {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
+    {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
+    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'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", },
 ]
 
 
@@ -94,8 +110,8 @@ def download_file_with_auth(url, token, save_path):
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
-    with open(save_path, 'wb') as f:
-        f.write(response.content)
+    with open(save_path, 'wb') as downloaded_file:
+        downloaded_file.write(response.content)
     logger.info(f"File {save_path} downloaded successfully")
 
 
@@ -107,15 +123,34 @@ def download_model(model):
     os.makedirs(f"models/tokenizers/{name}", exist_ok=True)
 
     files = ["config.json", "tokenizer.json", "tokenizer_config.json"]
+
     if tokt == TOKENIZER_TYPE.SPM:
         files.append("tokenizer.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 tokt == TOKENIZER_TYPE.UGM:
+        files.append("spiece.model")
+
+    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:
@@ -125,14 +160,14 @@ for model in models:
         logger.error(f"Failed to download model {model['name']}. Error: {e}")
 
 
-# generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function:
+# generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
 for model in models:
     name = model["name"]
     tokt = model["tokt"]
 
-    if tokt == TOKENIZER_TYPE.SPM:
+    if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
 
     # Skip if the tokenizer folder does not exist or there are other download issues previously
@@ -142,12 +177,15 @@ for model in models:
 
     # create the tokenizer
     try:
-        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+        if name == "t5":
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
     except OSError as e:
         logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
         continue  # Skip to the next model if the tokenizer can't be loaded
 
-    chktok = tokenizer.encode(chktxt)
+    chktok = tokenizer.encode(CHK_TXT)
     chkhsh = sha256(str(chktok).encode()).hexdigest()
 
     logger.info(f"model: {name}")
@@ -179,7 +217,7 @@ src_func = f"""
         # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
         # use in llama.cpp to implement the same pre-tokenizer
 
-        chktxt = {repr(chktxt)}
+        chktxt = {repr(CHK_TXT)}
 
         chktok = tokenizer.encode(chktxt)
         chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -189,7 +227,7 @@ src_func = f"""
 
         res = None
 
-        # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script
+        # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
         #       or pull the latest version of the model from Huggingface
         #       don't edit the hashes manually!
 {src_ifs}
@@ -198,9 +236,9 @@ src_func = f"""
             logger.warning("**************************************************************************************")
             logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
             logger.warning("**          There are 2 possible reasons for this:")
-            logger.warning("**          - the model has not been added to convert-hf-to-gguf-update.py yet")
+            logger.warning("**          - the model has not been added to convert_hf_to_gguf_update.py yet")
             logger.warning("**          - the pre-tokenization config has changed upstream")
-            logger.warning("**          Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
+            logger.warning("**          Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
             logger.warning("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
             logger.warning("**")
             logger.warning(f"** chkhsh:  {{chkhsh}}")
@@ -214,7 +252,7 @@ src_func = f"""
         return res
 """
 
-convert_py_pth = pathlib.Path("convert-hf-to-gguf.py")
+convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
 convert_py = convert_py_pth.read_text(encoding="utf-8")
 convert_py = re.sub(
     r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
@@ -225,7 +263,7 @@ convert_py = re.sub(
 
 convert_py_pth.write_text(convert_py, encoding="utf-8")
 
-logger.info("+++ convert-hf-to-gguf.py was updated")
+logger.info("+++ convert_hf_to_gguf.py was updated")
 
 # generate tests for each tokenizer model
 
@@ -263,6 +301,7 @@ tests = [
     "\n =",
     "' era",
     "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
+    "!!!!!!",
     "3",
     "33",
     "333",
@@ -272,8 +311,9 @@ tests = [
     "3333333",
     "33333333",
     "333333333",
-    # "Cửa Việt", # llama-bpe fails on this
-    chktxt,
+    "Cửa Việt", # llama-bpe fails on this
+    " discards",
+    CHK_TXT,
 ]
 
 # write the tests to ./models/ggml-vocab-{name}.gguf.inp
@@ -300,7 +340,10 @@ for model in models:
 
     # create the tokenizer
     try:
-        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+        if name == "t5":
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
     except OSError as e:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
@@ -326,6 +369,6 @@ logger.info("\nRun the following commands to generate the vocab files for testin
 for model in models:
     name = model["name"]
 
-    print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
+    print(f"python3 convert_hf_to_gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
 
 logger.info("\n")

convert_llama_ggml_to_gguf.py

@@ -116,7 +116,7 @@ class Tensor:
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
-        self.dtype= dtype
+        self.dtype= gguf.GGMLQuantizationType(dtype)
         self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])
@@ -132,6 +132,10 @@ class Tensor:
 
 
 class GGMLModel:
+
+    file_format: GGMLFormat
+    format_version: int
+
     def __init__(self):
         self.hyperparameters = None
         self.vocab = None
@@ -290,7 +294,7 @@ class GGMLToGGUF:
         if self.vocab_override is not None:
             vo = self.vocab_override
             logger.info('* Adding vocab item(s)')
-            for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
+            for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
                 tokens.append(vbytes)
                 scores.append(score)
                 toktypes.append(ttype)
@@ -354,7 +358,8 @@ class GGMLToGGUF:
 
 
 def handle_metadata(cfg, hp):
-    import convert
+    import examples.convert_legacy_llama as convert
+
     assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
     hf_config_path   = cfg.model_metadata_dir / "config.json"
     orig_config_path = cfg.model_metadata_dir / "params.json"

convert_lora_to_gguf.py

@@ -0,0 +1,433 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+import logging
+import argparse
+import os
+import sys
+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
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+import gguf
+
+# reuse model definitions from convert_hf_to_gguf.py
+from convert_hf_to_gguf import LazyTorchTensor, Model
+
+logger = logging.getLogger("lora-to-gguf")
+
+
+@dataclass
+class PartialLoraTensor:
+    A: Tensor | None = None
+    B: Tensor | None = None
+
+
+# magic to support tensor shape modifications and splitting
+class LoraTorchTensor:
+    _lora_A: Tensor  # (n_rank, row_size)
+    _lora_B: Tensor  # (col_size, n_rank)
+    _rank: int
+
+    def __init__(self, A: Tensor, B: Tensor):
+        assert len(A.shape) == len(B.shape)
+        assert A.shape[-2] == B.shape[-1]
+        if A.dtype != B.dtype:
+            A = A.to(torch.float32)
+            B = B.to(torch.float32)
+        self._lora_A = A
+        self._lora_B = B
+        self._rank = B.shape[-1]
+
+    def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
+        return (self._lora_A, self._lora_B)
+
+    def __getitem__(
+        self,
+        indices: (
+            SupportsIndex
+            | slice
+            | tuple[SupportsIndex | slice | Tensor, ...]  # TODO: add ellipsis in the type signature
+        ),
+    ) -> LoraTorchTensor:
+        shape = self.shape
+        if isinstance(indices, SupportsIndex):
+            if len(shape) > 2:
+                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
+            else:
+                raise NotImplementedError  # can't return a vector
+        elif isinstance(indices, slice):
+            if len(shape) > 2:
+                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
+            else:
+                return LoraTorchTensor(self._lora_A, self._lora_B[indices])
+        elif isinstance(indices, tuple):
+            assert len(indices) > 0
+            if indices[-1] is Ellipsis:
+                return self[indices[:-1]]
+            # expand ellipsis
+            indices = tuple(
+                u
+                for v in (
+                    (
+                        (slice(None, None) for _ in range(len(indices) - 1))
+                        if i is Ellipsis
+                        else (i,)
+                    )
+                    for i in indices
+                )
+                for u in v
+            )
+
+            if len(indices) < len(shape):
+                indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
+
+            # TODO: make sure this is correct
+            indices_A = (
+                *(
+                    (
+                        j.__index__() % self._lora_A.shape[i]
+                        if isinstance(j, SupportsIndex)
+                        else slice(None, None)
+                    )
+                    for i, j in enumerate(indices[:-2])
+                ),
+                slice(None, None),
+                indices[-1],
+            )
+            indices_B = indices[:-1]
+            return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
+        else:
+            raise NotImplementedError  # unknown indice type
+
+    @property
+    def dtype(self) -> torch.dtype:
+        assert self._lora_A.dtype == self._lora_B.dtype
+        return self._lora_A.dtype
+
+    @property
+    def shape(self) -> tuple[int, ...]:
+        assert len(self._lora_A.shape) == len(self._lora_B.shape)
+        return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
+
+    def size(self, dim=None):
+        assert dim is None
+        return self.shape
+
+    def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
+        if isinstance(shape[0], tuple):
+            new_shape: tuple[int, ...] = shape[0]
+        else:
+            new_shape = cast(tuple[int, ...], shape)
+        orig_shape = self.shape
+        if len(new_shape) < 2:
+            raise NotImplementedError  # can't become a vector
+
+        # expand -1 in the shape
+        if any(dim == -1 for dim in new_shape):
+            n_elems = prod(orig_shape)
+            n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
+            assert n_elems % n_new_elems == 0
+            new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
+
+        if new_shape[-1] != orig_shape[-1]:
+            raise NotImplementedError  # can't reshape the row size trivially
+
+        shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
+        shape_B = (*new_shape[:-1], self._rank)
+        return LoraTorchTensor(
+            self._lora_A.reshape(shape_A),
+            self._lora_B.reshape(shape_B),
+        )
+
+    def reshape_as(self, other: Tensor) -> LoraTorchTensor:
+        return self.reshape(*other.shape)
+
+    def view(self, *size: int) -> LoraTorchTensor:
+        return self.reshape(*size)
+
+    def permute(self, *dims: int) -> LoraTorchTensor:
+        shape = self.shape
+        dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
+        if dims[-1] == -1:
+            # TODO: support higher dimensional A shapes bigger than 1
+            assert all(dim == 1 for dim in self._lora_A.shape[:-2])
+            return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
+        if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
+            return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
+        else:
+            # TODO: compose the above two
+            raise NotImplementedError
+
+    def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
+        shape = self.shape
+        dims = [i for i in range(len(shape))]
+        dims[dim0], dims[dim1] = dims[dim1], dims[dim0]
+        return self.permute(*dims)
+
+    def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor:
+        return self.transpose(axis0, axis1)
+
+    def to(self, *args, **kwargs):
+        return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
+
+    @classmethod
+    def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
+        del types  # unused
+
+        if kwargs is None:
+            kwargs = {}
+
+        if func is torch.permute:
+            return type(args[0]).permute(*args, **kwargs)
+        elif func is torch.reshape:
+            return type(args[0]).reshape(*args, **kwargs)
+        elif func is torch.stack:
+            assert isinstance(args[0], Sequence)
+            dim = kwargs.get("dim", 0)
+            assert dim == 0
+            return LoraTorchTensor(
+                torch.stack([a._lora_A for a in args[0]], dim),
+                torch.stack([b._lora_B for b in args[0]], dim),
+            )
+        elif func is torch.cat:
+            assert isinstance(args[0], Sequence)
+            dim = kwargs.get("dim", 0)
+            assert dim == 0
+            if len(args[0][0].shape) > 2:
+                return LoraTorchTensor(
+                    torch.cat([a._lora_A for a in args[0]], dim),
+                    torch.cat([b._lora_B for b in args[0]], dim),
+                )
+            elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
+                return LoraTorchTensor(
+                    args[0][0]._lora_A,
+                    torch.cat([b._lora_B for b in args[0]], dim),
+                )
+            else:
+                raise NotImplementedError
+        else:
+            raise NotImplementedError
+
+
+def get_base_tensor_name(lora_tensor_name: str) -> str:
+    base_name = lora_tensor_name.replace("base_model.model.", "")
+    base_name = base_name.replace(".lora_A.weight", ".weight")
+    base_name = base_name.replace(".lora_B.weight", ".weight")
+    return base_name
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        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.",
+    )
+    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",
+    )
+    parser.add_argument(
+        "--bigendian", action="store_true",
+        help="model is executed on big endian machine",
+    )
+    parser.add_argument(
+        "--no-lazy", action="store_true",
+        help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
+    )
+    parser.add_argument(
+        "--verbose", action="store_true",
+        help="increase output verbosity",
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="only print out what will be done, without writing any new files",
+    )
+    parser.add_argument(
+        "--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 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)
+
+    ftype_map: dict[str, gguf.LlamaFileType] = {
+        "f32": gguf.LlamaFileType.ALL_F32,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+        "auto": gguf.LlamaFileType.GUESSED,
+    }
+
+    ftype = ftype_map[args.outtype]
+
+    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"
+
+    if args.outfile is not None:
+        fname_out = args.outfile
+    else:
+        # output in the same directory as the model by default
+        fname_out = dir_lora
+
+    if os.path.exists(input_model):
+        # lazy import load_file only if lora is in safetensors format.
+        from safetensors.torch import load_file
+
+        lora_model = load_file(input_model, device="cpu")
+    else:
+        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
+    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])
+        except NotImplementedError:
+            logger.error(f"Model {hparams['architectures'][0]} is not supported")
+            sys.exit(1)
+
+        class LoraModel(model_class):
+            model_arch = model_class.model_arch
+
+            lora_alpha: float
+
+            def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
+
+                super().__init__(*args, **kwargs)
+
+                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)
+
+            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] = {}
+
+                for name, tensor in lora_model.items():
+                    if self.lazy:
+                        tensor = LazyTorchTensor.from_eager(tensor)
+                    base_name = get_base_tensor_name(name)
+                    is_lora_a = ".lora_A.weight" in name
+                    is_lora_b = ".lora_B.weight" in name
+                    if not is_lora_a and not is_lora_b:
+                        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:
+                        if is_lora_a:
+                            tensor_map[base_name].A = tensor
+                        else:
+                            tensor_map[base_name].B = tensor
+                    else:
+                        if is_lora_a:
+                            tensor_map[base_name] = PartialLoraTensor(A=tensor)
+                        else:
+                            tensor_map[base_name] = PartialLoraTensor(B=tensor)
+
+                for name, tensor in tensor_map.items():
+                    assert tensor.A is not None
+                    assert tensor.B is not None
+                    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 = 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()
+
+                    yield (dest_name + ".lora_a", lora_a)
+                    yield (dest_name + ".lora_b", lora_b)
+
+        alpha: float = lparams["lora_alpha"]
+
+        model_instance = LoraModel(
+            dir_base_model,
+            ftype,
+            fname_out,
+            is_big_endian=args.bigendian,
+            use_temp_file=False,
+            eager=args.no_lazy,
+            dry_run=args.dry_run,
+            dir_lora_model=dir_lora,
+            lora_alpha=alpha,
+            hparams=hparams,
+        )
+
+        logger.info("Exporting model...")
+        model_instance.write()
+        logger.info(f"Model successfully exported to {model_instance.fname_out}")

docs/android.md

@@ -0,0 +1,83 @@
+
+# Android
+
+## Build on Android using Termux
+
+[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.
+
+With Termux, you can install and run `llama.cpp` as if the environment were Linux. Once in the Termux shell:
+
+```
+$ apt update && apt upgrade -y
+$ apt install git cmake
+```
+
+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:
+
+```
+$ curl -L {model-url} -o ~/{model}.gguf
+```
+
+Then, if you are not already in the repo directory, `cd` into `llama.cpp` and:
+
+```
+$ ./build/bin/llama-simple -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
+```
+
+Here, we show `llama-simple`, 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`.

docs/backend/CANN.md

@@ -0,0 +1,259 @@
+# llama.cpp for CANN
+
+ - [Background](#background)
+ - [News](#news)
+ - [OS](#os)
+ - [Hardware](#hardware)
+ - [Model Supports](#model-supports)
+ - [DataType Supports](#datatype-supports)
+ - [Docker](#docker)
+ - [Linux](#linux)
+ - [TODO](#todo)
+
+
+## Background
+
+**Ascend NPU** is a range of AI processors using Neural Processing Unit. It will efficiently handle matrix-matrix multiplication, dot-product and scalars.
+
+**CANN** (Compute Architecture for Neural Networks) is a heterogeneous computing architecture for AI scenarios, providing support for multiple AI frameworks on the top and serving AI processors and programming at the bottom. It plays a crucial role in bridging the gap between upper and lower layers, and is a key platform for improving the computing efficiency of Ascend AI processors. Meanwhile, it offers a highly efficient and easy-to-use programming interface for diverse application scenarios, allowing users to rapidly build AI applications and services based on the Ascend platform.
+
+**Llama.cpp + CANN**
+
+The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
+
+## News
+
+- 2024.8
+  - Support `Q4_0` and `Q8_0` data type for Ascend NPU.
+- 2024.7
+  - Create CANN backend for Ascend NPU.
+
+## OS
+
+| OS      | Status  | Verified                                       |
+|:-------:|:-------:|:----------------------------------------------:|
+| Linux   | Support | Ubuntu 22.04, OpenEuler22.03                   |
+
+
+## Hardware
+
+### Ascend NPU
+
+**Verified devices**
+| Ascend NPU                    | Status  |
+|:-----------------------------:|:-------:|
+| Atlas 300T A2                 | Support |
+
+*Notes:*
+
+- If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
+- If you run successfully with your Ascend NPU device, please help update the upper table.
+
+
+## Model Supports
+
+| Model Name                  | FP16  | Q8_0 | Q4_0 |
+|:----------------------------|:-----:|:----:|:----:|
+| AquilaChat2-7B              |   √   |   √  |   √  |
+| Baichuan-7b                 |   √   |   √  |   √  |
+| Baichuan2-7B-Chat           |   √   |   √  |   √  |
+| bitnet_b1_58-large          |   √   |   √  |   √  |
+| bloom-560m                  |   √   |   x  |   √  |
+| bloomz-alpaca-560m          |   √   |   x  |   √  |
+| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
+| chatglm3-6B                 |   x   |   x  |   x  |
+| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
+| CodeShell-7B                |   √   |   √  |   √  |
+| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
+| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
+| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
+| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
+| falcon-7b-instruct          |   √   |   √  |   √  |
+| flan-t5-large               |   √   |   √  |   √  |
+| gemma-2-9b-it               |   √   |   √  |   √  |
+| glm-4-9B                    |   x   |   x  |   x  |
+| gpt2                        |   √   |   √  |   √  |
+| Gpt2-163M                   |   √   |   √  |   √  |
+| granite-3B-code-instruct    |   √   |   √  |   √  |
+| GritLM-7B                   |   √   |   √  |   √  |
+| internlm2_5-7b-chat         |   √   |   √  |   √  |
+| koala-7B-HF                 |   √   |   √  |   √  |
+| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
+| Llama-3-Smaug-8B            |   √   |   √  |   √  |
+| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
+| Llama3-8B                   |   √   |   √  |   √  |
+| Llama3-8b-chinese           |   √   |   √  |   √  |
+| mamba-130m-hf               |   √   |   √  |   √  |
+| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
+| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
+| mpt-7B                      |   √   |   √  |   √  |
+| OLMo-1B-hf                  |   √   |   √  |   √  |
+| OpenELM-3B-Instruct         |   √   |   √  |   √  |
+| Orion-14b-base              |   √   |   √  |   √  |
+| phi1                        |   x   |   x  |   x  |
+| phi2                        |   x   |   x  |   x  |
+| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
+| plamo-13b                   |   √   |   √  |   √  |
+| pythia-70M                  |   x   |   x  |   x  |
+| Qwen-7B                     |   √   |   √  |   √  |
+| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
+| Refact-1_6B-fim             |   √   |   √  |   √  |
+| SmolLM-135M                 |   √   |   √  |   √  |
+| stablelm-zephyr             |   x   |   x  |   x  |
+| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
+| starcoderbase-1b            |   √   |   √  |   √  |
+| starcoder2-3b               |   √   |   √  |   √  |
+| vigogne-7b-chat             |   √   |   √  |   √  |
+| xverse-7b-chat              |   √   |   √  |   √  |
+| Yi-6b-Chat                  |   √   |   √  |   √  |
+
+
+
+## DataType Supports
+
+| DataType               | Status  |
+|:----------------------:|:-------:|
+| FP16                   | Support |
+| Q8_0                   | Support |
+| Q4_0                   | Support |
+
+## Docker
+
+### Build Images
+You can get a image with llama.cpp in one command.
+```sh
+docker build -t llama-cpp-cann -f .devops/llama-cli-cann.Dockerfile .
+```
+
+### Run container
+
+```sh
+# Find all cards.
+npu-smi info
+
+# Select the cards that you want to use, make sure these cards are not used by someone.
+# Following using cards of device0.
+docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /PATH_TO_YOUR_MODELS/:/app/models -it llama-cpp-cann -m /app/models/MODEL_PATH -ngl 32 -p "Building a website can be done in 10 simple steps:"
+```
+
+*Notes:*
+
+- You may need to install Ascend Driver and firmware on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
+
+## Linux
+
+### I. Setup Environment
+
+1. **Install Ascend Driver and firmware**
+
+    ```sh
+    # create driver running user.
+    sudo groupadd -g HwHiAiUser
+    sudo useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash
+    sudo usermod -aG HwHiAiUser $USER
+
+    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
+    # and install driver.
+    sudo sh Ascend-hdk-910b-npu-driver_x.x.x_linux-{arch}.run --full --install-for-all
+    ```
+
+    Once installed, run `npu-smi info` to check whether driver is installed successfully.
+    ```sh
+    +-------------------------------------------------------------------------------------------+
+    | npu-smi 24.1.rc2               Version: 24.1.rc2                                          |
+    +----------------------+---------------+----------------------------------------------------+
+    | NPU   Name           | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
+    | Chip                 | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+    +======================+===============+====================================================+
+    | 2     xxx            | OK            | 64.4        51                15   / 15            |
+    | 0                    | 0000:01:00.0  | 0           1873 / 15077      0    / 32768         |
+    +======================+===============+====================================================+
+    | 5     xxx            | OK            | 64.0        52                15   / 15            |
+    | 0                    | 0000:81:00.0  | 0           1874 / 15077      0    / 32768         |
+    +======================+===============+====================================================+
+    | No running processes found in NPU 2                                                       |
+    +======================+===============+====================================================+
+    | No running processes found in NPU 5                                                       |
+    +======================+===============+====================================================+
+    ```
+
+2. **Install Ascend Firmware**
+    ```sh
+    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
+    # and install driver.
+    sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
+    ```
+    If the following messaage appers, firmware is installed successfully.
+    ```sh
+    Firmware package installed successfully!
+    ```
+
+
+3. **Install CANN toolkit and kernels**
+
+    CANN toolkit and kernels can be obtained from the official [CANN Toolkit](https://www.hiascend.com/zh/developer/download/community/result?module=cann) page.
+
+    Please download the corresponding version that satified your system. The minimum version required is 8.0.RC2.alpha002 and here is the install command.
+    ```sh
+    pip3 install attrs numpy decorator sympy cffi pyyaml pathlib2 psutil protobuf scipy requests absl-py wheel typing_extensions
+    sh Ascend-cann-toolkit_8.0.RC2.alpha002_linux-aarch64.run --install
+    sh Ascend-cann-kernels-910b_8.0.RC2.alpha002_linux.run --install
+    ```
+
+    Set Ascend Variables:
+    ```sh
+    echo "source ~/Ascend/ascend-toolkit/set_env.sh" >> ~/.bashrc
+    source ~/.bashrc
+    ```
+
+Upon a successful installation, CANN is enabled for the available ascend devices.
+
+### II. Build llama.cpp
+
+```sh
+cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
+cmake --build build --config release
+```
+
+### III. Run the inference
+
+1. **Retrieve and prepare model**
+
+    You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration.
+
+    **Notes**:
+
+      - CANN backend only supports FP16/Q4_0/Q8_0 models currently.
+
+2. **Launch inference**
+
+    There are two device selection modes:
+
+    - Single device: Use one device target specified by the user.
+    - Multiple devices: Automatically choose the devices with the same backend.
+
+    | Device selection | Parameter                              |
+    |:----------------:|:--------------------------------------:|
+    | Single device    | --split-mode none --main-gpu DEVICE_ID |
+    | Multiple devices | --split-mode layer (default)           |
+
+    Examples:
+
+    - Use device 0:
+
+    ```sh
+    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+    ```
+
+    - Use multiple devices:
+
+    ```sh
+    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
+    ```
+
+### **GitHub contribution**:
+Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay.
+
+
+## TODO
+- Support more models and data types.

docs/backend/SYCL.md

@@ -20,17 +20,13 @@
 **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
-- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
 ### Llama.cpp + SYCL
 
-The llama.cpp SYCL backend is 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*).
-
-When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend.
-
-It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
+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
 
@@ -38,13 +34,20 @@ 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.
+
 - 2024.5
   - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
   - Arch Linux is verified successfully.
@@ -80,7 +83,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +98,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -104,10 +114,18 @@ The following release is verified with good quality:
 
 **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.
@@ -179,6 +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**
@@ -189,7 +211,7 @@ Please follow the instructions for downloading and installing the Toolkit for Li
 
 Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
 
-Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs.
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
 
 - **Adding support to Nvidia GPUs**
 
@@ -205,6 +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**
 
@@ -216,33 +251,48 @@ 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
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
-# Build LLAMA with MKL BLAS acceleration for intel GPU
-
 # Option 1: Use FP32 (recommended for better performance in most cases)
 cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
@@ -254,6 +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:$LD_LIBRARY_PATH
@@ -262,62 +313,106 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$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
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
-A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
+
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
-found 6 SYCL devices:
+found 2 SYCL devices:
+
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
 |ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
 |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
 | 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
-| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
-| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
-| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
-| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
-| Attribute              | Note                                                        |
-|------------------------|-------------------------------------------------------------|
-| compute capability 1.3 | Level-zero driver/runtime, recommended                      |
-| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases |
+#### Choose level-zero devices
 
-4. Launch inference
+|Chosen Device ID|Setting|
+|-|-|
+|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"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run-llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run-llama2.sh
+```
+
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
-- Multiple devices: Automatically select the devices with the same largest Max compute-units.
+- Single device: Use one device assigned by user. Default device id is 0.
+- Multiple devices: Automatically choose the devices with the same backend.
+
+In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
 
 | Device selection | Parameter                              |
 |------------------|----------------------------------------|
@@ -331,11 +426,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -343,12 +433,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -395,7 +479,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -416,6 +500,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -430,12 +526,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -447,7 +539,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -455,52 +549,65 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
-The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
+This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
-found 6 SYCL devices:
+found 2 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
 |ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
 |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
 | 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
-| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
-| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
-| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
-| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 
 ```
+#### Choose level-zero devices
 
-| Attribute              | Note                                                      |
-|------------------------|-----------------------------------------------------------|
-| compute capability 1.3 | Level-zero running time, recommended                      |
-| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases |
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
+#### Execute
 
-4. Launch inference
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
+
+2. Command line
+
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user.
-- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+- Single device: Use one device assigned by user. Default device id is 0.
+- Multiple devices: Automatically choose the devices with the same backend.
+
+In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
 
 | Device selection | Parameter                              |
 |------------------|----------------------------------------|
@@ -520,11 +627,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -538,17 +641,19 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
-| Name               | Value                             | Function                                    |
-|--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
-| 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.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for 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. |
 
 #### Runtime
 
@@ -584,9 +689,26 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  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.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA

docs/build.md

@@ -0,0 +1,408 @@
+# Build llama.cpp locally
+
+**To get the Code:**
+
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+In order to build llama.cpp you have four different options.
+
+- Using `make`:
+  - On Linux or MacOS:
+
+      ```bash
+      make
+      ```
+
+  - On Windows (x86/x64 only, arm64 requires cmake):
+
+    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:
+    - 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`
+
+- Using `CMake`:
+
+  ```bash
+  cmake -B build
+  cmake --build build --config Release
+  ```
+
+  **Notes**:
+
+    - 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:
+
+      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+
+      ```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.
+
+## 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:
+
+### Accelerate Framework:
+
+This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
+
+### 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
+    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+    cmake --build build --config Release
+    ```
+
+### BLIS
+
+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).
+
+- Using manual oneAPI installation:
+  By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
+    ```bash
+    source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
+    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
+    cmake --build build --config Release
+    ```
+
+- Using oneAPI docker image:
+  If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
+
+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
+
+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).
+
+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.
+
+- Using `make`:
+  ```bash
+  make GGML_CUDA=1
+  ```
+- Using `CMake`:
+
+  ```bash
+  cmake -B build -DGGML_CUDA=ON
+  cmake --build build --config Release
+  ```
+
+The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
+
+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. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
+
+The following compilation options are also available to tweak performance:
+
+| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
+|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| 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_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
+
+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
+  cmake -B build -DGGML_MUSA=ON
+  cmake --build build --config Release
+  ```
+
+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.
+
+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.
+
+### hipBLAS
+
+This provides BLAS 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_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`.
+  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
+  Note that if you get the following error:
+  ```
+  clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+  ```
+  Try searching for a directory under `HIP_PATH` that contains the file
+  `oclc_abi_version_400.bc`. Then, add the following to the start of the
+  command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
+  like:
+  ```bash
+  HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+  HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
+      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_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)
+  Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
+
+
+The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
+If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+
+### Vulkan
+
+**Windows**
+
+#### w64devkit
+
+Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
+
+Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
+
+Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
+```sh
+SDK_VERSION=1.3.283.0
+cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/
+cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/
+cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/
+cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc <<EOF
+Name: Vulkan-Loader
+Description: Vulkan Loader
+Version: $SDK_VERSION
+Libs: -lvulkan-1
+EOF
+
+```
+Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
+
+#### Git Bash MINGW64
+
+Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
+
+Download and install [`Visual Studio Community Edition`](https://visualstudio.microsoft.com/) and make sure you select `C++`
+
+Download and install [`CMake`](https://cmake.org/download/) with the default settings
+
+Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
+
+Go into your `llama.cpp` directory and right click, select `Open Git Bash Here` and then run the following commands
+
+```
+cmake -B build -DGGML_VULKAN=ON
+cmake --build build --config Release
+```
+
+Now you can load the model in conversation mode using `Vulkan`
+
+```
+build/bin/release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
+```
+
+#### MSYS2
+Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
+  ```sh
+  pacman -S git \
+      mingw-w64-ucrt-x86_64-gcc \
+      mingw-w64-ucrt-x86_64-cmake \
+      mingw-w64-ucrt-x86_64-vulkan-devel \
+      mingw-w64-ucrt-x86_64-shaderc
+  ```
+Switch into `llama.cpp` directory and build using CMake.
+```sh
+cmake -B build -DGGML_VULKAN=ON
+cmake --build build --config Release
+```
+
+**With docker**:
+
+You don't need to install Vulkan SDK. It will be installed inside the container.
+
+```sh
+# Build the image
+docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
+
+# Then, use it:
+docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+
+**Without docker**:
+
+Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+
+For example, on Ubuntu 22.04 (jammy), use the command below:
+
+```bash
+wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
+wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+apt update -y
+apt-get install -y vulkan-sdk
+# To verify the installation, use the command below:
+vulkaninfo
+```
+
+Alternatively your package manager might be able to provide the appropriate libraries.
+For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
+For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
+
+Then, build llama.cpp using the cmake command below:
+
+```bash
+cmake -B build -DGGML_VULKAN=1
+cmake --build build --config Release
+# Test the output binary (with "-ngl 33" to offload all layers to GPU)
+./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+
+# You should see in the output, ggml_vulkan detected your GPU. For example:
+# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
+```
+
+### 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/).
+
+Make sure to have the CANN toolkit installed. You can download it from here: [CANN Toolkit](https://www.hiascend.com/developer/download/community/result?module=cann)
+
+Go to `llama.cpp` directory and build using CMake.
+```bash
+cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
+cmake --build build --config release
+```
+
+You can test with:
+
+`./build/bin/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
+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
+
+To read documentation for how to build on Android, [click here](./android.md)
+
+### Arm CPU optimized mulmat kernels
+
+Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
+
+To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).

docs/development/HOWTO-add-model.md

@@ -1,4 +1,4 @@
-## Add a new model architecture to `llama.cpp`
+# Add a new model architecture to `llama.cpp`
 
 Adding a model requires few steps:
 
@@ -9,15 +9,15 @@ Adding a model requires few steps:
 After following these steps, you can open PR.
 
 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
-- [main](../examples/main)
-- [imatrix](../examples/imatrix)
-- [quantize](../examples/quantize)
-- [server](../examples/server)
+- [main](/examples/main/)
+- [imatrix](/examples/imatrix/)
+- [quantize](/examples/quantize/)
+- [server](/examples/server/)
 
 ### 1. Convert the model to GGUF
 
 This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
-Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format).
+Depending on the model architecture, you can use either [convert_hf_to_gguf.py](/convert_hf_to_gguf.py) or [examples/convert_legacy_llama.py](/examples/convert_legacy_llama.py) (for `llama/llama2` models in `.pth` format).
 
 The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
 
@@ -31,7 +31,7 @@ class MyModel(Model):
     model_arch = gguf.MODEL_ARCH.GROK
 ```
 
-2. Define the layout of the GGUF tensors in [constants.py](../gguf-py/gguf/constants.py)
+2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)
 
 Add an enum entry in `MODEL_ARCH`, the model human friendly name in `MODEL_ARCH_NAMES` and the GGUF tensor names in `MODEL_TENSORS`.
 
@@ -54,7 +54,7 @@ Example for `falcon` model:
 
 As a general rule, before adding a new tensor name to GGUF, be sure the equivalent naming does not already exist.
 
-Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](../gguf-py/gguf/tensor_mapping.py) file.
+Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](/gguf-py/gguf/tensor_mapping.py) file.
 
 If the tensor name is part of a repetitive layer/block, the key word `bid` substitutes it.
 
@@ -100,7 +100,7 @@ Have a look at existing implementation like `build_llama`, `build_dbrx` or `buil
 
 When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
 
-Note: to debug the inference graph: you can use [llama-eval-callback](../examples/eval-callback).
+Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
 
 ## GGUF specification
 

docs/development/token_generation_performance_tips.md

@@ -1,7 +1,7 @@
 # Token generation performance troubleshooting
 
 ## Verifying that the model is running on the GPU with CUDA
-Make sure you compiled llama with the correct env variables according to [this guide](../README.md#CUDA), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+Make sure you compiled llama with the correct env variables according to [this guide](/docs/build.md#cuda), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
 ```shell
 ./llama-cli -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
 ```

docs/docker.md

@@ -0,0 +1,123 @@
+# Docker
+
+## Prerequisites
+* Docker must be installed and running on your system.
+* Create a folder to store big models & intermediate files (ex. /llama/models)
+
+## Images
+We have three Docker images available for this project:
+
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `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, ROCm or MUSA library, you'll need to build the images locally for now).
+
+## Usage
+
+The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
+
+Replace `/path/to/models` below with the actual path where you downloaded the models.
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
+```
+
+On completion, you are ready to play!
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a light image:
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
+## Docker With CUDA
+
+Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
+
+## Building Docker locally
+
+```bash
+docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `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:
+
+1. `local/llama.cpp:full-cuda`: 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-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
+
+## Usage
+
+After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --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 --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
+```

docs/install.md

@@ -0,0 +1,39 @@
+# Install pre-built version of llama.cpp
+
+## Homebrew
+
+On Mac and Linux, the homebrew package manager can be used via
+
+```sh
+brew install llama.cpp
+```
+The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+
+## Nix
+
+On Mac and Linux, the Nix package manager can be used via
+
+```sh
+nix profile install nixpkgs#llama-cpp
+```
+For flake enabled installs.
+
+Or
+
+```sh
+nix-env --file '<nixpkgs>' --install --attr llama-cpp
+```
+
+For non-flake enabled installs.
+
+This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+## Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+
+```sh
+flox install llama-cpp
+```
+
+Flox follows the nixpkgs build of llama.cpp.

examples/CMakeLists.txt

@@ -13,16 +13,14 @@ 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(finetune)
     add_subdirectory(gbnf-validator)
+    add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
     add_subdirectory(gritlm)
@@ -50,7 +48,7 @@ else()
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(simple-chat)
     add_subdirectory(speculative)
     add_subdirectory(tokenize)
-    add_subdirectory(train-text-from-scratch)
 endif()

examples/baby-llama/CMakeLists.txt

@@ -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)

examples/batched-bench/README.md

@@ -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}
+```

examples/batched-bench/batched-bench.cpp

@@ -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);
 
@@ -66,10 +45,10 @@ 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 = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -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;
 }

examples/batched.swift/Sources/main.swift

@@ -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
@@ -229,7 +218,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
 
 private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
     var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), 0, false)
     if nTokens < 0 {
         let actualTokensCount = -Int(nTokens)
         result = .init(repeating: 0, count: actualTokensCount)
@@ -238,6 +227,7 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
             token,
             &result,
             Int32(result.count),
+            0,
             false
         )
         assert(check == actualTokensCount)

examples/batched/batched.cpp

@@ -1,37 +1,36 @@
+#include "arg.h"
 #include "common.h"
+#include "log.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-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 -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
-    LOG_TEE("\n");
+static void print_usage(int, char ** argv) {
+    LOG("\nexample usage:\n");
+    LOG("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
+    LOG("\n");
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
+    common_init();
 
     // number of parallel batches
     int n_parallel = params.n_parallel;
 
     // total length of the sequences including the prompt
-    int n_predict = 32;
+    int n_predict = params.n_predict;
 
     // init LLM
 
@@ -40,83 +39,110 @@ 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);
 
     if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        LOG_ERR("%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
-    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    tokens_list = common_tokenize(model, params.prompt, true);
 
     const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_predict, n_parallel);
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
+    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
+
     if (ctx == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
         return 1;
     }
 
     const int n_ctx = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
-        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
-        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_ERR("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
     // print the prompt token-by-token
 
-    fprintf(stderr, "\n");
+    LOG("\n");
 
     for (auto id : tokens_list) {
-        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+        LOG("%s", common_token_to_piece(ctx, id).c_str());
     }
 
-    fflush(stderr);
-
     // create a llama_batch
     // we use this object to submit token data for decoding
-    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
+    llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
+
+    std::vector<llama_seq_id> seq_ids(n_parallel, 0);
+    for (int32_t i = 0; i < n_parallel; ++i) {
+        seq_ids[i] = i;
+    }
 
     // evaluate the initial prompt
     for (size_t i = 0; i < tokens_list.size(); ++i) {
-        llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
+        common_batch_add(batch, tokens_list[i], i, seq_ids, false);
     }
     GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
+    if (llama_model_has_encoder(model)) {
+        if (llama_encode(ctx, batch)) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == -1) {
+            decoder_start_token_id = llama_token_bos(model);
+        }
+
+        common_batch_clear(batch);
+        common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
+    }
+
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
 
     if (llama_decode(ctx, batch) != 0) {
-        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        LOG_ERR("%s: llama_decode() failed\n", __func__);
         return 1;
     }
 
-    // assign the system KV cache to all parallel sequences
-    // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
-    for (int32_t i = 1; i < n_parallel; ++i) {
-        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
-    }
+    //// assign the system KV cache to all parallel sequences
+    //// this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
+    //for (int32_t i = 1; i < n_parallel; ++i) {
+    //    llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+    //}
 
     if (n_parallel > 1) {
-        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+        LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
     }
 
     // main loop
@@ -135,7 +161,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_predict) {
         // prepare the next batch
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -144,36 +170,14 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
-
-            std::vector<llama_token_data> candidates;
-            candidates.reserve(n_vocab);
-
-            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-            }
-
-            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-            const int   top_k = 40;
-            const float top_p = 0.9f;
-            const float temp  = 0.4f;
-
-            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
-            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-            llama_sample_temp (ctx, &candidates_p, temp);
-
-            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
-
-            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
 
             // is it an end of generation? -> mark the stream as finished
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
                 i_batch[i] = -1;
-                LOG_TEE("\n");
+                LOG("\n");
                 if (n_parallel > 1) {
-                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                    LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
                 }
 
                 continue;
@@ -181,16 +185,15 @@ int main(int argc, char ** argv) {
 
             // if there is only one stream, we print immediately to stdout
             if (n_parallel == 1) {
-                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-                fflush(stdout);
+                LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
             }
 
-            streams[i] += llama_token_to_piece(ctx, new_token_id);
+            streams[i] += common_token_to_piece(ctx, new_token_id);
 
             i_batch[i] = batch.n_tokens;
 
             // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
+            common_batch_add(batch, new_token_id, n_cur, { i }, true);
 
             n_decode += 1;
         }
@@ -204,32 +207,33 @@ int main(int argc, char ** argv) {
 
         // evaluate the current batch with the transformer model
         if (llama_decode(ctx, batch)) {
-            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
     }
 
-    LOG_TEE("\n");
-
     if (n_parallel > 1) {
-        LOG_TEE("\n");
+        LOG("\n");
 
         for (int32_t i = 0; i < n_parallel; ++i) {
-            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+            LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
         }
     }
 
     const auto t_main_end = ggml_time_us();
 
-    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+    LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG("\n");
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
 
+    llama_sampler_free(smpl);
     llama_free(ctx);
     llama_free_model(model);
 

examples/benchmark/CMakeLists.txt

@@ -1,6 +0,0 @@
-set(TARGET llama-bench-matmult)
-add_executable(${TARGET} benchmark-matmult.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/benchmark/benchmark-matmult.cpp

@@ -1,275 +0,0 @@
-#include "common.h"
-#include "ggml.h"
-
-#include <locale.h>
-#include <assert.h>
-#include <math.h>
-#include <cstring>
-#include <cstdio>
-#include <cinttypes>
-#include <unordered_map>
-#include <queue>
-#include <string.h>
-#include <cassert>
-#include <fstream>
-#include <string>
-#include <iterator>
-#include <algorithm>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
-
-    if (plan.work_size > 0) {
-        buf.resize(plan.work_size);
-        plan.work_data = buf.data();
-    }
-
-    ggml_graph_compute(graph, &plan);
-}
-
-static float tensor_sum_elements(const ggml_tensor * tensor) {
-    double sum = 0;
-    if (tensor->type == GGML_TYPE_F32) {
-        for (int j = 0; j < tensor->ne[1]; j++) {
-            for (int k = 0; k < tensor->ne[0]; k++) {
-                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
-            }
-        }
-    }
-    return sum;
-}
-
-static void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
-        tensor->type, ggml_type_name(tensor->type),
-        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
-    float sum = tensor_sum_elements(tensor);
-    printf("Sum of tensor %s is %6.2f\n", name, sum);
-}
-
-#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
-
-struct benchmark_params_struct {
-    int32_t n_threads     = 1;
-    int32_t n_iterations  = 10;
-};
-
-static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -i N, --iter N     number of iterations to use during computation (default: %d)\n", params.n_iterations);
-    fprintf(stderr, "\n");
-}
-
-int main(int argc, char ** argv)  {
-    struct benchmark_params_struct benchmark_params;
-
-    bool invalid_param = false;
-    std::string arg;
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
-
-        if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_threads = std::stoi(argv[i]);
-        } else if (arg == "-i" || arg == "--iter") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_iterations = std::stoi(argv[i]);
-        }  else if (arg == "-h" || arg == "--help") {
-            print_usage(argc, argv, benchmark_params);
-            exit(0);
-        }
-    }
-    if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        print_usage(argc, argv, benchmark_params);
-        exit(1);
-    }
-
-    print_build_info();
-    printf("Starting Test\n");
-
-    // create the ggml context
-    struct ggml_context * ctx;
-    //const int sizex = 4096;
-    //const int sizey = 11008;
-
-#undef VERBOSE_DEBUGGING
-#ifndef VERBOSE_DEBUGGING
-    const int sizey = 4096;
-    const int sizex = 11008;
-    const int sizez = 128;
-#else
-    /* Working - let's increase size */
-    const int sizey = 1;
-    const int sizex = (8*32);
-    const int sizez = 1;
-
-    /*const int sizey = 1;
-    const int sizex = 3*(8*32);
-    const int sizez = 1;*/
-#endif
-
-    //printf("Memsize required = %i\n", sizex*sizex);
-
-    // TODO: perform the bench for all types or for a user specified type
-    const ggml_type qtype = GGML_TYPE_Q4_1;
-
-    size_t ctx_size = 0;
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += 1024*1024*16;
-
-    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
-
-    struct ggml_init_params params = {
-        /*.mem_size   =*/ ctx_size,
-        /*.mem_buffer =*/ NULL,
-        /* no_alloc   =*/ 0
-    };
-
-    ctx = ggml_init(params);
-    if (!ctx) {
-        fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-        return 1;
-    }
-
-
-    printf("Creating new tensors\n");
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m11, 1.0f);
-
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m12, 1.5f);
-
-    // printf("Creating new tensor m2\n");
-    struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
-    ggml_set_f32(m2, 2.0f);
-
-    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
-    // printf("Creating new tensor m11xm2\n");
-    struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf, m11xm2);
-
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    TENSOR_DUMP(m11);
-    TENSOR_DUMP(m2);
-
-    std::vector<uint8_t> work_buffer;
-
-    ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
-
-    TENSOR_DUMP(gf->nodes[0]);
-
-    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
-
-    int32_t nelements = sizex*sizey;
-
-    // Set up a the benchmark matrices
-    // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
-
-    // Set up a the compute graph
-    // printf("Creating new tensor q31\n");
-    struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf31, q31);
-
-    // Set up a second graph computation to make sure we override the CPU cache lines
-    // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
-
-    // printf("Creating new tensor q32\n");
-    struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
-
-    //printf("Creating compute graph\n");
-    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf32, q32);
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    const int dimx = sizex;
-    const int dimy = sizey;
-    const int dimz = sizez;
-    long long int flops_per_dot_product = dimy + dimy;
-    long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
-    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
-
-
-    // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
-
-    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
-    printf("=====================================================================================\n");
-
-    double  gflops_sum = 0;
-    for (int i=0;i<benchmark_params.n_iterations ;i++) {
-
-        long long int start = ggml_time_us();
-        //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
-
-        long long int stop = ggml_time_us();
-        long long int usec = stop-start;
-        double gflops = (double)(flops_per_matrix)/usec/1000.0;
-        gflops_sum += gflops;
-        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
-            i,
-            benchmark_params.n_threads,
-            sizex, sizey, sizez, flops_per_matrix,
-            usec,gflops);
-
-#ifdef VERBOSE_DEBUGGING
-        TENSOR_DUMP("res",gf31.nodes[0])
-#endif
-
-        // Check that the matrix multiplication result is in the right ballpark
-        // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
-        float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
-        float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
-
-        if (delta > allowed_delta)  {
-            printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
-                sum_of_F32_reference,
-                sum_of_Q4_result,
-                delta,
-                allowed_delta
-            );
-            exit(0);
-        }
-
-        // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
-    }
-    printf("\n");
-    printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
-    printf("=====================================================================================\n");
-}

examples/chat-persistent.sh

@@ -23,8 +23,9 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
 NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
 NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
 
-SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
-SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
+SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
+'|'\
+'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
 CTX_SIZE=2048
@@ -129,15 +130,12 @@ while read -e line; do
 
     printf ' '
 
-    # HACK get num tokens from debug message
-    # TODO get both messages in one go
-    if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
-        ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+    if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
         echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
         exit 1
     fi
 
-    n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
+    n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
 
     if ((n_tokens > CTX_ROTATE_POINT)); then
         tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"

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

@@ -9,6 +9,7 @@
 #include <climits>
 #include <cstring>
 #include <cstdarg>
+#include <cinttypes>
 #include <ctime>
 #include <random>
 #include <stdexcept>
@@ -105,43 +106,43 @@ static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_
     const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
     try {
         w->token_embedding_table.resize(p->vocab_size * p->dim);
-        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
         w->rms_att_weight.resize(p->n_layers * p->dim);
-        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
         w->rms_ffn_weight.resize(p->n_layers * p->dim);
-        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
         w->wq.resize(p->n_layers * p->dim * p->dim);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
         w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
         w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
         w->wo.resize(p->n_layers * p->dim * p->dim);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
         w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->rms_final_weight.resize(p->dim);
-        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        LOG_INF("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
         if (shared_weights) {
             w->wcls = {};
         } else {
             w->wcls.resize(p->vocab_size * p->dim);
-            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+            LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
         }
     }
     catch (std::length_error &) {
@@ -173,7 +174,7 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
+        LOG_ERR("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -181,26 +182,26 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    LOG("----- Quick print of first of the weight vales of all the variables\n");
-    LOG("%f\n", w->token_embedding_table[0]);
-    LOG("%f\n", w->rms_att_weight[0]);
-    LOG("%f\n", w->rms_ffn_weight[0]);
+    LOG_INF("----- Quick print of first of the weight vales of all the variables\n");
+    LOG_INF("%f\n", w->token_embedding_table[0]);
+    LOG_INF("%f\n", w->rms_att_weight[0]);
+    LOG_INF("%f\n", w->rms_ffn_weight[0]);
 
-    LOG("%f\n", w->wq[0]);
-    LOG("%f\n", w->wk[0]);
-    LOG("%f\n", w->wv[0]);
-    LOG("%f\n", w->wo[0]);
-    LOG("%f\n", w->w1[0]);
-    LOG("%f\n", w->w2[0]);
-    LOG("%f\n", w->w3[0]);
-    LOG("%f\n", w->rms_att_weight[0]);
-    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
+    LOG_INF("%f\n", w->wq[0]);
+    LOG_INF("%f\n", w->wk[0]);
+    LOG_INF("%f\n", w->wv[0]);
+    LOG_INF("%f\n", w->wo[0]);
+    LOG_INF("%f\n", w->w1[0]);
+    LOG_INF("%f\n", w->w2[0]);
+    LOG_INF("%f\n", w->w3[0]);
+    LOG_INF("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG_INF("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
 //////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
 
-struct llama_vocab {
+struct my_llama_vocab {
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -318,20 +319,20 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
-    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
-    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
-    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
-    LOG("%s: n_head:    %u\n", __func__, params->n_head);
-    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
-    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
-    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
-    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+    LOG_INF("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG_INF("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG_INF("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG_INF("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG_INF("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG_INF("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG_INF("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG_INF("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG_INF("%s: n_rot:     %u\n", __func__, params->n_rot);
 }
 
 static void print_tensor_info(const struct ggml_context * ctx) {
     for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        LOG("%s: Allocating ", __func__);
+        LOG_INF("%s: Allocating ", __func__);
         int64_t total = 1;
         int i = 0;
         for (; i < ggml_n_dims(t); ++i) {
@@ -524,9 +525,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
+static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
-        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
+        LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -574,7 +575,7 @@ static void load_vocab(const char * filename, const Config * config, struct llam
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
+        LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -582,13 +583,13 @@ static void load_vocab(const char * filename, const Config * config, struct llam
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (llama_vocab::id id=0; id<n_vocab; ++id) {
+        for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
 
             unsigned char byte_val;
-            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
             if (id == UNKNOWN_TOKEN_ID) {
                 text = "<unk>";
                 type = LLAMA_TOKEN_TYPE_UNKNOWN;
@@ -630,7 +631,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
 }
 
 static void save_as_llama_model(
-    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+    struct my_llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
 ) {
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
@@ -670,7 +671,7 @@ static void save_as_llama_model(
     std::vector<const char*> tokens;
     std::vector<float> scores;
     std::vector<llama_token_type> token_types;
-    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+    for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
         tokens.push_back(token_data.text.c_str());
         scores.push_back(token_data.score);
         token_types.push_back(token_data.type);
@@ -871,23 +872,25 @@ static std::string basename(const std::string &path) {
 }
 
 int main(int argc, char ** argv) {
+    common_init();
+
     struct train_params params = get_default_train_params();
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
-    log_set_target(stdout);
+
     Config config;
     TransformerWeights weights = {};
     {
-        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        LOG_INF("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
         FILE * file = fopen(params.fn_llama2c_model, "rb");
         if (!file) {
-            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            LOG_ERR("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
             return 1;
         }
         // read in the config header
         if (fread(&config, sizeof(Config), 1, file) != 1) {
-            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            LOG_ERR("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
             return 1;
         }
         auto shared_weights = config.vocab_size > 0;
@@ -896,13 +899,13 @@ int main(int argc, char ** argv) {
         // read in the Transformer weights
         alloc_weights(&weights, &config, shared_weights);
         if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
-            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            LOG_ERR("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
             return 1;
         }
         fclose(file);
     }
 
-    struct llama_vocab vocab;
+    struct my_llama_vocab vocab;
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
@@ -929,7 +932,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG_INF("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/convert_legacy_llama.py

@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
+from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
 
 import numpy as np
 
@@ -346,42 +346,6 @@ class Params:
         return params
 
 
-@dataclass
-class Metadata:
-    name: Optional[str] = None
-    author: Optional[str] = None
-    version: Optional[str] = None
-    url: Optional[str] = None
-    description: Optional[str] = None
-    licence: Optional[str] = None
-    source_url: Optional[str] = None
-    source_hf_repo: Optional[str] = None
-
-    @staticmethod
-    def load(metadata_path: Path) -> Metadata:
-        if metadata_path is None or not metadata_path.exists():
-            return Metadata()
-
-        with open(metadata_path, 'r') as file:
-            data = json.load(file)
-
-        # Create a new Metadata instance
-        metadata = Metadata()
-
-        # Assigning values to Metadata attributes if they exist in the JSON file
-        # This is based on LLM_KV_NAMES mapping in llama.cpp
-        metadata.name = data.get("general.name")
-        metadata.author = data.get("general.author")
-        metadata.version = data.get("general.version")
-        metadata.url = data.get("general.url")
-        metadata.description = data.get("general.description")
-        metadata.license = data.get("general.license")
-        metadata.source_url = data.get("general.source.url")
-        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
-
-        return metadata
-
-
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)
@@ -492,12 +456,13 @@ class LazyTensor:
 
 LazyModel: TypeAlias = 'dict[str, LazyTensor]'
 
+ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none']
 
 @dataclass
 class ModelPlus:
     model: LazyModel
     paths: list[Path]  # Where this was read from.
-    format: Literal['ggml', 'torch', 'safetensors', 'none']
+    format: ModelFormat
     vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
@@ -536,7 +501,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
 
 
 def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
-    formats = set(mp.format for mp in models_plus)
+    formats: set[ModelFormat] = set(mp.format for mp in models_plus)
     assert len(formats) == 1, "different formats?"
     format = formats.pop()
     paths = [path for mp in models_plus for path in mp.paths]
@@ -555,7 +520,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     else:
         model = merge_sharded([mp.model for mp in models_plus])
 
-    return ModelPlus(model, paths, format, vocab)  # pytype: disable=wrong-arg-types
+    return ModelPlus(model, paths, format, vocab)
 
 
 def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
@@ -805,7 +770,7 @@ class OutputFile:
     def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
-    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
+    def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
         # Metadata About The Model And Its Provenence
         name = "LLaMA"
         if metadata is not None and metadata.name is not None:
@@ -823,16 +788,95 @@ class OutputFile:
                 self.gguf.add_author(metadata.author)
             if metadata.version is not None:
                 self.gguf.add_version(metadata.version)
-            if metadata.url is not None:
-                self.gguf.add_url(metadata.url)
+            if metadata.organization is not None:
+                self.gguf.add_organization(metadata.organization)
+
+            if metadata.finetune is not None:
+                self.gguf.add_finetune(metadata.finetune)
+            if metadata.basename is not None:
+                self.gguf.add_basename(metadata.basename)
+
             if metadata.description is not None:
                 self.gguf.add_description(metadata.description)
-            if metadata.licence is not None:
-                self.gguf.add_licence(metadata.licence)
+            if metadata.quantized_by is not None:
+                self.gguf.add_quantized_by(metadata.quantized_by)
+
+            if metadata.size_label is not None:
+                self.gguf.add_size_label(metadata.size_label)
+
+            if metadata.license is not None:
+                self.gguf.add_license(metadata.license)
+            if metadata.license_name is not None:
+                self.gguf.add_license_name(metadata.license_name)
+            if metadata.license_link is not None:
+                self.gguf.add_license_link(metadata.license_link)
+
+            if metadata.url is not None:
+                self.gguf.add_url(metadata.url)
+            if metadata.doi is not None:
+                self.gguf.add_doi(metadata.doi)
+            if metadata.uuid is not None:
+                self.gguf.add_uuid(metadata.uuid)
+            if metadata.repo_url is not None:
+                self.gguf.add_repo_url(metadata.repo_url)
+
             if metadata.source_url is not None:
                 self.gguf.add_source_url(metadata.source_url)
-            if metadata.source_hf_repo is not None:
-                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
+            if metadata.source_doi is not None:
+                self.gguf.add_source_doi(metadata.source_doi)
+            if metadata.source_uuid is not None:
+                self.gguf.add_source_uuid(metadata.source_uuid)
+            if metadata.source_repo_url is not None:
+                self.gguf.add_source_repo_url(metadata.source_repo_url)
+
+            if metadata.base_models is not None:
+                self.gguf.add_base_model_count(len(metadata.base_models))
+                for key, base_model_entry in enumerate(metadata.base_models):
+                    if "name" in base_model_entry:
+                        self.gguf.add_base_model_name(key, base_model_entry["name"])
+                    if "author" in base_model_entry:
+                        self.gguf.add_base_model_author(key, base_model_entry["author"])
+                    if "version" in base_model_entry:
+                        self.gguf.add_base_model_version(key, base_model_entry["version"])
+                    if "organization" in base_model_entry:
+                        self.gguf.add_base_model_organization(key, base_model_entry["organization"])
+                    if "description" in base_model_entry:
+                        self.gguf.add_base_model_description(key, base_model_entry["description"])
+                    if "url" in base_model_entry:
+                        self.gguf.add_base_model_url(key, base_model_entry["url"])
+                    if "doi" in base_model_entry:
+                        self.gguf.add_base_model_doi(key, base_model_entry["doi"])
+                    if "uuid" in base_model_entry:
+                        self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
+                    if "repo_url" in base_model_entry:
+                        self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
+
+            if metadata.datasets is not None:
+                self.gguf.add_dataset_count(len(metadata.datasets))
+                for key, dataset_entry in enumerate(metadata.datasets):
+                    if "name" in dataset_entry:
+                        self.gguf.add_dataset_name(key, dataset_entry["name"])
+                    if "author" in dataset_entry:
+                        self.gguf.add_dataset_author(key, dataset_entry["author"])
+                    if "version" in dataset_entry:
+                        self.gguf.add_dataset_version(key, dataset_entry["version"])
+                    if "organization" in dataset_entry:
+                        self.gguf.add_dataset_organization(key, dataset_entry["organization"])
+                    if "description" in dataset_entry:
+                        self.gguf.add_dataset_description(key, dataset_entry["description"])
+                    if "url" in dataset_entry:
+                        self.gguf.add_dataset_url(key, dataset_entry["url"])
+                    if "doi" in dataset_entry:
+                        self.gguf.add_dataset_doi(key, dataset_entry["doi"])
+                    if "uuid" in dataset_entry:
+                        self.gguf.add_dataset_uuid(key, dataset_entry["uuid"])
+                    if "repo_url" in dataset_entry:
+                        self.gguf.add_dataset_repo_url(key, dataset_entry["repo_url"])
+
+            if metadata.tags is not None:
+                self.gguf.add_tags(metadata.tags)
+            if metadata.languages is not None:
+                self.gguf.add_languages(metadata.languages)
 
     def add_meta_arch(self, params: Params) -> None:
         # Metadata About The Neural Architecture Itself
@@ -943,7 +987,7 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -977,7 +1021,7 @@ class OutputFile:
         fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
-        metadata: Metadata = None,
+        metadata: gguf.Metadata | None = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -1020,35 +1064,32 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
     raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
-def model_parameter_count(model: LazyModel) -> int:
-    total_model_parameters = 0
-    for i, (name, lazy_tensor) in enumerate(model.items()):
-        sum_weights_in_tensor = 1
+def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
+    total_params = 0
+    shared_params = 0
+    expert_params = 0
+
+    for name, lazy_tensor in tensors:
+        # We don't need these
+        if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
+            continue
+
+        # Got A Tensor
+        sum_weights_in_tensor: int = 1
+
+        # Tensor Volume
         for dim in lazy_tensor.shape:
             sum_weights_in_tensor *= dim
-        total_model_parameters += sum_weights_in_tensor
-    return total_model_parameters
 
+        if ".experts." in name:
+            if ".experts.0." in name:
+                expert_params += sum_weights_in_tensor
+        else:
+            shared_params += sum_weights_in_tensor
 
-def model_parameter_count_rounded_notation(model_params_count: int) -> str:
-    if model_params_count > 1e12 :
-        # Trillions Of Parameters
-        scaled_model_params = model_params_count * 1e-12
-        scale_suffix = "T"
-    elif model_params_count > 1e9 :
-        # Billions Of Parameters
-        scaled_model_params = model_params_count * 1e-9
-        scale_suffix = "B"
-    elif model_params_count > 1e6 :
-        # Millions Of Parameters
-        scaled_model_params = model_params_count * 1e-6
-        scale_suffix = "M"
-    else:
-        # Thousands Of Parameters
-        scaled_model_params = model_params_count * 1e-3
-        scale_suffix = "K"
+        total_params += sum_weights_in_tensor
 
-    return f"{round(scaled_model_params)}{scale_suffix}"
+    return total_params, shared_params, expert_params
 
 
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1230,34 +1271,24 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
-    quantization = {
+def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
+    name = metadata.name if metadata.name is not None else None
+    basename = metadata.basename if metadata.basename is not None else None
+    finetune = metadata.finetune if metadata.finetune is not None else None
+    version = metadata.version if metadata.version is not None else None
+    size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
+
+    output_type = {
         GGMLFileType.AllF32:    "F32",
         GGMLFileType.MostlyF16: "F16",
         GGMLFileType.MostlyQ8_0: "Q8_0",
     }[file_type]
 
-    parameters = model_parameter_count_rounded_notation(model_params_count)
-
-    expert_count = ""
-    if params.n_experts is not None:
-        expert_count = f"{params.n_experts}x"
-
-    version = ""
-    if metadata is not None and metadata.version is not None:
-        version = f"-{metadata.version}"
-
-    name = "ggml-model"
-    if metadata is not None and metadata.name is not None:
-        name = metadata.name
-    elif params.path_model is not None:
-        name = params.path_model.name
-
-    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
+    return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
 
 
-def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
-    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
     ret = model_paths[0].parent / f"{default_filename}.gguf"
     if ret in model_paths:
         logger.error(
@@ -1296,8 +1327,9 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
     parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
-    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for an authorship metadata override file")
     parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
+    parser.add_argument("--model-name",   type=str, default=None, help="name of the model")
 
     args = parser.parse_args(args_in)
 
@@ -1309,32 +1341,36 @@ def main(args_in: list[str] | None = None) -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
-    metadata = Metadata.load(args.metadata)
+    model_name = args.model_name
+    dir_model = args.model
+
+    metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
 
     if args.get_outfile:
-        model_plus = load_some_model(args.model)
+        model_plus = load_some_model(dir_model)
         params = Params.load(model_plus)
-        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
-        model_params_count = model_parameter_count(model_plus.model)
-        ftype   = pick_output_type(model, args.outtype)
-        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
+        model = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = per_model_weight_count_estimation(model_plus.model.items())
+        ftype = pick_output_type(model, args.outtype)
+
+        if (metadata is None or metadata.name is None) and params.path_model is not None:
+            metadata.name = params.path_model.name
+
+        print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
         return
 
     if args.no_vocab and args.vocab_only:
         raise ValueError("--vocab-only does not make sense with --no-vocab")
 
     if args.dump_single:
-        model_plus = lazy_load_file(args.model)
+        model_plus = lazy_load_file(dir_model)
         do_dump_model(model_plus)
         return
 
     if not args.vocab_only:
-        model_plus = load_some_model(args.model)
+        model_plus = load_some_model(dir_model)
     else:
-        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
-
-    model_params_count = model_parameter_count(model_plus.model)
-    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
+        model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
 
     if args.dump:
         do_dump_model(model_plus)
@@ -1367,7 +1403,7 @@ def main(args_in: list[str] | None = None) -> None:
         logger.info(f"params = {params}")
 
     model_parent_path = model_plus.paths[0].parent
-    vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
+    vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
     vocab_types = None if args.no_vocab else args.vocab_type.split(",")
     vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
@@ -1396,13 +1432,23 @@ def main(args_in: list[str] | None = None) -> None:
     if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
         vocab = model_plus.vocab
 
+    assert params is not None
+
+    if metadata.name is None and params.path_model is not None:
+        metadata.name = params.path_model.name
+
+    model_params_count = per_model_weight_count_estimation(model_plus.model.items())
+    logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
+
     logger.info(f"Vocab info: {vocab}")
     logger.info(f"Special vocab info: {special_vocab}")
     model   = model_plus.model
     model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
+
+    metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")

examples/cvector-generator/cvector-generator.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -12,14 +13,15 @@
 #include "ggml-metal.h"
 #endif
 
+#include <algorithm>
+#include <climits>
 #include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <iostream>
 #include <string>
 #include <tuple>
 #include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
-#include <climits>
 
 
 //////////////////////////////////////////////////
@@ -29,15 +31,13 @@ template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     std::string ret;
     for (; begin != end; ++begin) {
-        ret += llama_token_to_piece(ctx, *begin);
+        ret += common_token_to_piece(ctx, *begin);
     }
 
     return ret;
 }
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
     printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
     printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
@@ -271,9 +271,9 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
-        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+        tokens_pos = common_tokenize(ctx, pos, add_bos, true);
+        tokens_neg = common_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
         padding_seq(ctx, tokens_pos, max_seq_len);
         padding_seq(ctx, tokens_neg, max_seq_len);
@@ -281,7 +281,7 @@ struct tokenized_prompt {
 
     void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
         // TODO: customize padding token
-        std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
+        std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
         llama_token pad_tok = pad_tokens.back();
         while (tokens.size() < len) {
             tokens.push_back(pad_tok);
@@ -339,7 +339,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
     llama_kv_cache_clear(ctx);
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }
@@ -370,7 +370,7 @@ static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const
  * Load prompt files and completion file.
  * Then format each pair of prompt + completion to make an entry.
  */
-static int prepare_entries(gpt_params & params, train_context & ctx_train) {
+static int prepare_entries(common_params & params, train_context & ctx_train) {
     // load prompts
     std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
     std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
@@ -388,10 +388,9 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
 }
 
 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_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
 
@@ -414,9 +413,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_model * model;
-    llama_context * ctx;
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
+
+    llama_model * model = llama_init.model;
+    llama_context * ctx = llama_init.context;
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);
@@ -485,8 +485,8 @@ int main(int argc, char ** argv) {
     if (use_pca) {
         // run PCA
         PCA::pca_params pca_params;
-        pca_params.n_threads = params.n_threads;
-        pca_params.n_batch = params.n_pca_batch;
+        pca_params.n_threads    = params.cpuparams.n_threads;
+        pca_params.n_batch      = params.n_pca_batch;
         pca_params.n_iterations = params.n_pca_iterations;
         PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
     } else {

examples/cvector-generator/pca.hpp

@@ -12,12 +12,9 @@
 
 #include <cstdio>
 #include <ctime>
+#include <random>
 #include <string>
-#include <tuple>
 #include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
 
 #define DEBUG_POS 5
 
@@ -207,13 +204,6 @@ static ggml_status compute_piter(
         ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
     }
 
-// TODO: enable GPU support when support for GGML_OP_SQRT is added
-//#ifdef GGML_USE_METAL
-//    if (ggml_backend_is_metal(model.backend)) {
-//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
-//    }
-//#endif
-
     ggml_status res = ggml_backend_graph_compute(model.backend, gf);
     if (res == GGML_STATUS_SUCCESS) {
         auto extract_i = [](std::string prefix, std::string str) -> int {
@@ -229,8 +219,8 @@ static ggml_status compute_piter(
         result.eigenvectors.resize(params.n_batch);
         result.distances.resize(params.n_batch);
         // get output nodes
-        for (int i = 0; i < gf->n_nodes; ++i) {
-            auto node = gf->nodes[i];
+        for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
+            auto node = ggml_graph_node(gf, i);
             int iter = -1;
             // find b_tensor (without copying data from device)
             if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {

examples/deprecation-warning/README.md

@@ -0,0 +1,49 @@
+# Migration notice for binary filenames
+
+> [!IMPORTANT]
+[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
+
+This migration was important, but it is a breaking change that may not always be immediately obvious to users.
+
+Please update all scripts and workflows to use the new binary names.
+
+| Old Filename | New Filename |
+| ---- | ---- |
+| main | llama-cli |
+| server | llama-server |
+| llama-bench | llama-bench |
+| embedding | llama-embedding |
+| quantize | llama-quantize |
+| tokenize | llama-tokenize |
+| export-lora | llama-export-lora |
+| libllava.a | libllava.a |
+| baby-llama | llama-baby-llama |
+| batched | llama-batched |
+| batched-bench | llama-batched-bench |
+| benchmark-matmult | llama-benchmark-matmult |
+| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml |
+| eval-callback | llama-eval-callback |
+| gbnf-validator | llama-gbnf-validator |
+| gguf | llama-gguf |
+| gguf-split | llama-gguf-split |
+| gritlm | llama-gritlm |
+| imatrix | llama-imatrix |
+| infill | llama-infill |
+| llava-cli | llama-llava-cli |
+| lookahead | llama-lookahead |
+| lookup | llama-lookup |
+| lookup-create | llama-lookup-create |
+| lookup-merge | llama-lookup-merge |
+| lookup-stats | llama-lookup-stats |
+| parallel | llama-parallel |
+| passkey | llama-passkey |
+| perplexity | llama-perplexity |
+| q8dot | llama-q8dot |
+| quantize-stats | llama-quantize-stats |
+| retrieval | llama-retrieval |
+| save-load-state | llama-save-load-state |
+| simple | llama-simple |
+| speculative | llama-speculative |
+| vdot | llama-vdot |
+| tests/test-c.o | tests/test-c.o |
+