llama : minor llama_grammar refactoring

ggml-ci

.clang-format

@@ -1,161 +0,0 @@
----
-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

@@ -26,7 +26,7 @@ COPY . .
 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 -B build -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/* .
 

.devops/full-musa.Dockerfile

@@ -6,9 +6,6 @@ ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_V
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
-# MUSA architecture to build for (defaults to all supported archs)
-ARG MUSA_DOCKER_ARCH=default
-
 RUN apt-get update && \
     apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
 
@@ -22,11 +19,7 @@ WORKDIR /app
 
 COPY . .
 
-# Use the default MUSA archs if not specified
-RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+RUN cmake -B build -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/* .
 

.devops/llama-cli-cann.Dockerfile

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

.devops/llama-cli-cuda.Dockerfile

@@ -22,17 +22,16 @@ COPY . .
 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 \;
+    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc)
 
 FROM ${BASE_CUDA_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 /
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-cli /llama-cli
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-intel.Dockerfile

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

.devops/llama-cli-musa.Dockerfile

@@ -8,9 +8,6 @@ ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
-# MUSA architecture to build for (defaults to all supported archs)
-ARG MUSA_DOCKER_ARCH=default
-
 RUN apt-get update && \
     apt-get install -y build-essential git cmake
 
@@ -18,21 +15,16 @@ WORKDIR /app
 
 COPY . .
 
-# Use the default MUSA archs if not specified
-RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release --target llama-cli -j$(nproc) && \
-    mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc)
 
 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/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
 COPY --from=build /app/build/bin/llama-cli /llama-cli
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-vulkan.Dockerfile

@@ -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_NATIVE=OFF -DGGML_VULKAN=1 && \
+RUN cmake -B build -DGGML_VULKAN=1 && \
     cmake --build build --config Release --target llama-cli
 
 # Clean up

.devops/llama-server-cuda.Dockerfile

@@ -22,17 +22,16 @@ COPY . .
 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 \;
+    cmake -B build -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)
 
 FROM ${BASE_CUDA_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/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
 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

.devops/llama-server-intel.Dockerfile

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

.devops/llama-server-musa.Dockerfile

@@ -8,9 +8,6 @@ ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
-# MUSA architecture to build for (defaults to all supported archs)
-ARG MUSA_DOCKER_ARCH=default
-
 RUN apt-get update && \
     apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
@@ -18,21 +15,16 @@ WORKDIR /app
 
 COPY . .
 
-# Use the default MUSA archs if not specified
-RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release --target llama-server -j$(nproc) && \
-    mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
+RUN cmake -B build -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)
 
 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/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
 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

.devops/llama-server-vulkan.Dockerfile

@@ -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_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
     cmake --build build --config Release --target llama-server
 
 # Clean up

.devops/nix/package.nix

@@ -126,9 +126,9 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   };
 
   postPatch = ''
-    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+    substituteInPlace ./ggml/src/ggml-metal.m \
       --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-    substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
+    substituteInPlace ./ggml/src/ggml-metal.m \
       --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
   '';
 
@@ -173,7 +173,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       (cmakeBool "GGML_NATIVE" false)
       (cmakeBool "GGML_BLAS" useBlas)
       (cmakeBool "GGML_CUDA" useCuda)
-      (cmakeBool "GGML_HIP" useRocm)
+      (cmakeBool "GGML_HIPBLAS" useRocm)
       (cmakeBool "GGML_METAL" useMetalKit)
       (cmakeBool "GGML_VULKAN" useVulkan)
       (cmakeBool "GGML_STATIC" enableStatic)

.devops/nix/python-scripts.nix

@@ -34,7 +34,7 @@ let
 
     # server tests
     openai
-    pytest
+    behave
     prometheus-client
   ];
 in

.editorconfig

@@ -24,16 +24,6 @@ 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/01-bug-low.yml

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

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

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

.github/ISSUE_TEMPLATE/011-bug-results.yml

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

.github/ISSUE_TEMPLATE/019-bug-misc.yml

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

.github/ISSUE_TEMPLATE/02-bug-medium.yml

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

.github/ISSUE_TEMPLATE/03-bug-high.yml

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

.github/ISSUE_TEMPLATE/04-bug-critical.yml

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

.github/ISSUE_TEMPLATE/05-enhancement.yml

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

.github/ISSUE_TEMPLATE/06-research.yml

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

.github/ISSUE_TEMPLATE/07-refactor.yml

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

.github/labeler.yml

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

.github/workflows/build.yml

@@ -55,13 +55,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          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 -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -98,7 +92,7 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
-    runs-on: macos-13
+    runs-on: macos-12
 
     steps:
       - name: Clone
@@ -119,12 +113,7 @@ 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 \
-            -DLLAMA_CURL=ON \
-            -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
-            -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -405,36 +394,15 @@ jobs:
       - name: Build with native CMake HIP support
         id: cmake_build
         run: |
-          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=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_HIP=ON
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=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
 
@@ -601,7 +569,6 @@ 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 \
@@ -632,7 +599,6 @@ 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 \
@@ -728,7 +694,7 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     env:
       OPENBLAS_VERSION: 0.3.23
@@ -768,7 +734,7 @@ jobs:
         id: clone_kompute
         if: ${{ matrix.build == 'kompute-x64' }}
         run: |
-          git submodule update --init ggml/src/ggml-kompute/kompute
+          git submodule update --init ggml/src/kompute
 
       - name: Download OpenBLAS
         id: get_openblas
@@ -871,33 +837,12 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
           name: llama-bin-win-${{ matrix.build }}.zip
 
-  ubuntu-latest-cmake-cuda:
-    runs-on: ubuntu-latest
-    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
-
-    steps:
-        - name: Clone
-          id: checkout
-          uses: actions/checkout@v4
-
-        - name: Install dependencies
-          env:
-            DEBIAN_FRONTEND: noninteractive
-          run: |
-              apt update
-              apt install -y cmake build-essential ninja-build libgomp1 git
-
-        - name: Build with CMake
-          run: |
-            cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
-            cmake --build build
-
-  windows-2019-cmake-cuda:
+  windows-latest-cmake-cuda:
     runs-on: windows-2019
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.2.0', '11.7.1']
         build: ['cuda']
 
     steps:
@@ -905,83 +850,24 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
         with:
-            fetch-depth: 0
+          fetch-depth: 0
 
-      - name: Install Cuda Toolkit 11.7
-        if: ${{ matrix.cuda == '11.7' }}
-        run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-      - name: Install Cuda Toolkit 12.4
-        if: ${{ matrix.cuda == '12.4' }}
-        run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-
-      - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2
+      - name: Install CUDA toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.15
         with:
-          key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }}
-
-      - name: Install Ninja
-        id: install_ninja
-        run: |
-          choco install ninja
+          cuda: ${{ matrix.cuda }}
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 
       - name: Build
         id: cmake_build
-        shell: cmd
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
-          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
-          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
+          mkdir build
+          cd build
+          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
         id: tag
@@ -1010,12 +896,10 @@ jobs:
           name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
         run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
           $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
@@ -1033,8 +917,8 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
@@ -1044,8 +928,7 @@ jobs:
           fetch-depth: 0
 
       - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
       - name: Build
         id: cmake_build
@@ -1064,33 +947,25 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
-      - name: Build the release package
+      - name: Pack artifacts
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
-
-          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
-
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
-      - name: Upload the release package
+      - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
@@ -1126,7 +1001,7 @@ 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_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
+          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 -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
   windows-latest-cmake-hip-release:
@@ -1141,8 +1016,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
 
       - name: Install
         id: depends
@@ -1164,7 +1037,7 @@ 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_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ 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\"
@@ -1257,7 +1130,7 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
-      - windows-2019-cmake-cuda
+      - windows-latest-cmake-cuda
       - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64

.github/workflows/docker.yml

@@ -10,10 +10,12 @@
 name: Publish Docker image
 
 on:
-  workflow_dispatch: # allows manual triggering
-  schedule:
-    # Rebuild daily rather than on every push because it is expensive
-    - cron: '12 4 * * *'
+  #pull_request:
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  workflow_dispatch: # allows manual triggering, useful for debugging
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -27,6 +29,7 @@ permissions:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
+    #if: github.event.pull_request.draft == false
 
     runs-on: ubuntu-latest
     env:
@@ -114,7 +117,7 @@ jobs:
           swap-storage: true
 
       - name: Build and push Docker image (tagged + versioned)
-        if: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
+        if: github.event_name == 'push'
         uses: docker/build-push-action@v6
         with:
           context: .

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

@@ -0,0 +1,72 @@
+name: Nix aarch64 builds
+
+on:
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    # Rebuild daily rather than on every push because QEMU is expensive (e.g.
+    # 1.5h instead of minutes with the cold cache).
+    #
+    # randint(0, 59), randint(0, 23)
+    - cron: '26 12 * * *'
+  # But also rebuild if we touched any of the Nix expressions:
+  push:
+    branches:
+      - master
+    paths: ['**/*.nix', 'flake.lock']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/*.nix', 'flake.lock']
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+# 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
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install QEMU
+      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y qemu-user-static qemu-system-aarch64
+        sudo usermod -a -G kvm $USER
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-platforms = aarch64-linux
+          extra-system-features = nixos-test kvm
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: llama-cpp
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.aarch64-linux"
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --systems aarch64-linux
+          --flake
+          ".#checks.aarch64-linux"

.github/workflows/nix-ci.yml

@@ -0,0 +1,79 @@
+name: Nix CI
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+# 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:
+      fail-fast: false
+      matrix:
+        os: [ ubuntu-latest, macos-latest ]
+    runs-on: ${{ matrix.os }}
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: List all flake outputs
+      run: nix flake show --all-systems
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
+  nix-build:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ ubuntu-latest, macos-latest ]
+    runs-on: ${{ matrix.os }}
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: llama-cpp
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --flake
+          ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"

.github/workflows/nix-flake-update.yml

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

.github/workflows/nix-publish-flake.yml

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

.github/workflows/python-lint.yml

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

.github/workflows/server.yml

@@ -122,14 +122,14 @@ jobs:
         id: server_integration_tests
         run: |
           cd examples/server/tests
-          ./tests.sh
+          PORT=8888 ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
-          SLOW_TESTS=1 ./tests.sh
+          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
 
 
   server-windows:
@@ -180,12 +180,11 @@ jobs:
         run: |
           cd examples/server/tests
           $env:PYTHONIOENCODING = ":replace"
-          pytest -v -x
+          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
-          $env:SLOW_TESTS = "1"
-          pytest -v -x
+          behave.exe --stop --no-skipped --no-capture --tags slow

.gitignore

@@ -3,7 +3,6 @@
 *.a
 *.bat
 *.bin
-*.d
 *.dll
 *.dot
 *.etag
@@ -134,7 +133,3 @@ poetry.toml
 
 # 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/ggml-kompute/kompute
+	path = ggml/src/kompute
 	url = https://github.com/nomic-ai/kompute.git

AUTHORS

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

CMakeLists.txt

@@ -46,13 +46,6 @@ 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
 #
@@ -82,7 +75,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
-include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
 # override ggml options
 set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
@@ -96,10 +88,6 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_AMX)
-    set(GGML_AMX ON)
-endif()
-
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()
@@ -148,6 +136,7 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
+
 # At the moment some compile definitions are placed within the ggml/src
 # directory but not exported on the `ggml` target. This could be improved by
 # determining _precisely_ which defines are necessary for the llama-config
@@ -164,11 +153,8 @@ if (GGML_TARGET_DEFINES)
     list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
 endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
-# all public headers
-set(LLAMA_PUBLIC_HEADERS
-    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
-    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
-set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
+
+set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(

CMakePresets.json

@@ -24,12 +24,11 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-    { "name": "debug",    "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
-    { "name": "reldbg",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",   "hidden": true, "cacheVariables": { "GGML_STATIC":      "ON" } },
-    { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
-    { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "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": "arm64-windows-msvc", "hidden": true,
@@ -49,37 +48,21 @@
         }
     },
 
-    {
-        "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-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-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-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
-    { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
+    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
     { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
-    { "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
+    { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
     { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
     { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
-    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
-
-    { "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
-    { "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
+    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
   ]
 }

Makefile

@@ -1,6 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	libllava.a \
+	llama-baby-llama \
 	llama-batched \
 	llama-batched-bench \
 	llama-bench \
@@ -33,8 +34,6 @@ BUILD_TARGETS = \
 	llama-save-load-state \
 	llama-server \
 	llama-simple \
-	llama-simple-chat \
-	llama-run \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -49,12 +48,14 @@ TEST_TARGETS = \
 	tests/test-backend-ops \
 	tests/test-chat-template \
 	tests/test-double-float \
+	tests/test-grad0 \
 	tests/test-grammar-integration \
 	tests/test-grammar-parser \
 	tests/test-json-schema-to-grammar \
 	tests/test-llama-grammar \
 	tests/test-log \
 	tests/test-model-load-cancel \
+	tests/test-opt \
 	tests/test-quantize-fns \
 	tests/test-quantize-perf \
 	tests/test-rope \
@@ -62,7 +63,6 @@ TEST_TARGETS = \
 	tests/test-tokenizer-0 \
 	tests/test-tokenizer-1-bpe \
 	tests/test-tokenizer-1-spm
-#	tests/test-opt \
 
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
 LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
@@ -93,6 +93,11 @@ GGML_METAL := 1
 DEPRECATE_WARNING := 1
 endif
 
+ifdef LLAMA_OPENMP
+GGML_OPENMP := 1
+DEPRECATE_WARNING := 1
+endif
+
 ifdef LLAMA_RPC
 GGML_RPC := 1
 DEPRECATE_WARNING := 1
@@ -252,7 +257,7 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -DGGML_USE_CPU
+MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon
 MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
@@ -291,7 +296,6 @@ endif
 # some memory allocation are available on Linux through GNU extensions in libc
 ifeq ($(UNAME_S),Linux)
 	MK_CPPFLAGS += -D_GNU_SOURCE
-	MK_LDFLAGS  += -ldl
 endif
 
 # RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
@@ -360,10 +364,6 @@ ifdef LLAMA_SERVER_SSL
 	MK_LDFLAGS += -lssl -lcrypto
 endif
 
-ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
-endif
-
 # warnings
 WARN_FLAGS = \
 	-Wall \
@@ -528,59 +528,65 @@ ifndef GGML_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS  += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
-		MK_CPPFLAGS  += -DACCELERATE_NEW_LAPACK
-		MK_CPPFLAGS  += -DACCELERATE_LAPACK_ILP64
-		MK_LDFLAGS   += -framework Accelerate
-		OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
+		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
+		MK_LDFLAGS  += -framework Accelerate
+		OBJ_GGML    += ggml/src/ggml-blas.o
 	endif
 endif # GGML_NO_ACCELERATE
 
+ifdef GGML_MUSA
+	CC := clang
+	CXX := clang++
+	GGML_CUDA := 1
+	MK_CPPFLAGS += -DGGML_USE_MUSA
+endif
+
 ifndef GGML_NO_OPENMP
 	MK_CPPFLAGS += -DGGML_USE_OPENMP
 	MK_CFLAGS   += -fopenmp
 	MK_CXXFLAGS += -fopenmp
+	ifdef GGML_MUSA
+		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
+		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
+	endif # GGML_MUSA
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
-	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
-	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas)
-	MK_LDFLAGS   += $(shell pkg-config --libs openblas)
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
+	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_OPENBLAS
 
 ifdef GGML_OPENBLAS64
-	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
-	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas64)
-	MK_LDFLAGS   += $(shell pkg-config --libs openblas64)
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
+	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_OPENBLAS64
 
 ifdef GGML_BLIS
-	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS   += -lblis -L/usr/local/lib
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
+	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_BLIS
 
 ifdef GGML_NVPL
-	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
-	MK_LDFLAGS   += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
+	MK_LDFLAGS  += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
+	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_NVPL
 
 ifndef GGML_NO_LLAMAFILE
-	MK_CPPFLAGS  += -DGGML_USE_LLAMAFILE
-	OBJ_GGML_EXT += ggml/src/ggml-cpu/llamafile/sgemm.o
-endif
-
-ifndef GGML_NO_AMX
-	MK_CPPFLAGS += -DGGML_USE_AMX
-	OBJ_GGML_EXT += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
+	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
+	OBJ_GGML    += ggml/src/llamafile/sgemm.o
 endif
 
 ifdef GGML_RPC
-	MK_CPPFLAGS  += -DGGML_USE_RPC
-	OBJ_GGML_EXT += ggml/src/ggml-rpc.o
+	MK_CPPFLAGS += -DGGML_USE_RPC
+	OBJ_GGML    += ggml/src/ggml-rpc.o
 endif # GGML_RPC
 
 OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@@ -595,27 +601,41 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
+	ifdef GGML_MUSA
+		ifneq ('', '$(wildcard /opt/musa)')
+			CUDA_PATH ?= /opt/musa
+		else
+			CUDA_PATH ?= /usr/local/musa
+		endif
+
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
+		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
 	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
+		ifneq ('', '$(wildcard /opt/cuda)')
+			CUDA_PATH ?= /opt/cuda
+		else
+			CUDA_PATH ?= /usr/local/cuda
+		endif
 
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+		MK_NVCCFLAGS += -use_fast_math
+	endif # GGML_MUSA
 
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
 
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
 
+ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
+endif # GGML_MUSA
 
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
@@ -628,7 +648,11 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	NVCC = $(CCACHE) nvcc
+	ifdef GGML_MUSA
+		NVCC = $(CCACHE) mcc
+	else
+		NVCC = $(CCACHE) nvcc
+	endif # GGML_MUSA
 endif # GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
@@ -637,6 +661,10 @@ else ifndef CUDA_POWER_ARCH
 	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 
+ifdef GGML_CUDA_FORCE_DMMV
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # GGML_CUDA_FORCE_DMMV
+
 ifdef GGML_CUDA_FORCE_MMQ
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # GGML_CUDA_FORCE_MMQ
@@ -645,6 +673,20 @@ ifdef GGML_CUDA_FORCE_CUBLAS
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
 endif # GGML_CUDA_FORCE_CUBLAS
 
+ifdef GGML_CUDA_DMMV_X
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
+else
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+endif # GGML_CUDA_DMMV_X
+
+ifdef GGML_CUDA_MMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+else ifdef GGML_CUDA_DMMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
+else
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+endif # GGML_CUDA_MMV_Y
+
 ifdef GGML_CUDA_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_F16
@@ -653,6 +695,12 @@ ifdef GGML_CUDA_DMMV_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_DMMV_F16
 
+ifdef GGML_CUDA_KQUANTS_ITER
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
+else
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+endif
+
 ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
@@ -676,9 +724,15 @@ define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
+	ifdef GGML_MUSA
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
+endef # NVCC_COMPILE
+	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
+	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
 ggml/src/ggml-cuda/%.o: \
@@ -688,8 +742,8 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
 	ggml/include/ggml-cuda.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h \
@@ -700,9 +754,9 @@ ggml/src/ggml-cuda/ggml-cuda.o: \
 endif # GGML_CUDA
 
 ifdef GGML_VULKAN
-	MK_CPPFLAGS  += -DGGML_USE_VULKAN
-	MK_LDFLAGS   += $(shell pkg-config --libs vulkan)
-	OBJ_GGML_EXT += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
+	MK_CPPFLAGS += -DGGML_USE_VULKAN
+	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
+	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
 
 ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
@@ -732,10 +786,10 @@ GLSLC_CMD  = glslc
 _ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
 _ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
 _ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
-_ggml_vk_input_dir = ggml/src/ggml-vulkan/vulkan-shaders
+_ggml_vk_input_dir = ggml/src/vulkan-shaders
 _ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
 
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
+ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
 	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
 
 $(_ggml_vk_header): $(_ggml_vk_source)
@@ -747,12 +801,12 @@ $(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
 		--target-hpp $(_ggml_vk_header) \
 		--target-cpp $(_ggml_vk_source)
 
-vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
 
 endif # GGML_VULKAN
 
-ifdef GGML_HIP
+ifdef GGML_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH      ?= /usr
 		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
@@ -761,7 +815,11 @@ ifdef GGML_HIP
 		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 
-	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
+	GGML_CUDA_DMMV_X       ?= 32
+	GGML_CUDA_MMV_Y        ?= 1
+	GGML_CUDA_KQUANTS_ITER ?= 2
+
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 
 ifdef GGML_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
@@ -774,6 +832,13 @@ endif # GGML_HIP_UMA
 	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 
 	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
+	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
+	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
+
+ifdef GGML_CUDA_FORCE_DMMV
+	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # GGML_CUDA_FORCE_DMMV
 
 ifdef GGML_CUDA_FORCE_MMQ
 	HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
@@ -787,12 +852,12 @@ ifdef GGML_CUDA_NO_PEER_COPY
 	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY
 
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
 
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
 	ggml/include/ggml-cuda.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h \
@@ -807,171 +872,72 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-common.h \
 	ggml/src/ggml-cuda/common.cuh
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-endif # GGML_HIP
-
-ifdef GGML_MUSA
-	ifeq ($(wildcard /opt/musa),)
-		MUSA_PATH ?= /usr/local/musa
-	else
-		MUSA_PATH ?= /opt/musa
-	endif
-	MUSA_ARCHITECTURES ?= 21;22
-
-	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
-	MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib
-	MK_LDFLAGS += -lmusa -lmusart -lmublas
-
-	ifndef GGML_NO_OPENMP
-		# For Ubuntu Focal
-		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
-		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
-		# For Ubuntu Jammy
-		MK_CPPFLAGS += -I/usr/lib/llvm-14/lib/clang/14.0.0/include
-		MK_LDFLAGS  += -L/usr/lib/llvm-14/lib
-	endif # GGML_NO_OPENMP
-
-	CC  := $(MUSA_PATH)/bin/clang
-	CXX := $(MUSA_PATH)/bin/clang++
-	MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
-
-	MUSAFLAGS  = -x musa -mtgpu
-	MUSAFLAGS += $(foreach arch,$(subst ;, ,$(MUSA_ARCHITECTURES)),--cuda-gpu-arch=mp_$(arch))
-
-ifdef GGML_CUDA_FORCE_MMQ
-	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_FORCE_CUBLAS
-	MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # GGML_CUDA_FORCE_CUBLAS
-
-ifdef GGML_CUDA_F16
-	MUSAFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_F16
-
-ifdef GGML_CUDA_DMMV_F16
-	MUSAFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_DMMV_F16
-
-ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
-	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
-else
-	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
-endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
-
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
-endif # GGML_MUSA
+endif # GGML_HIPBLAS
 
 ifdef GGML_METAL
-	MK_CPPFLAGS  += -DGGML_USE_METAL
-	MK_LDFLAGS   += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML_EXT += ggml/src/ggml-metal/ggml-metal.o
-
-ifdef GGML_METAL_USE_BF16
-	MK_CPPFLAGS += -DGGML_METAL_USE_BF16
-endif # GGML_METAL_USE_BF16
+	MK_CPPFLAGS += -DGGML_USE_METAL
+	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	OBJ_GGML	+= ggml/src/ggml-metal.o
 ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
 ifdef GGML_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS  += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML_EXT += ggml/src/ggml-metal-embed.o
+	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
+	OBJ_GGML   += ggml/src/ggml-metal-embed.o
 endif
 endif # GGML_METAL
 
 ifdef GGML_METAL
-ggml/src/ggml-metal/ggml-metal.o: \
-	ggml/src/ggml-metal/ggml-metal.m \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
+ggml/src/ggml-metal.o: \
+	ggml/src/ggml-metal.m \
 	ggml/include/ggml-metal.h \
 	ggml/include/ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 ifdef GGML_METAL_EMBED_LIBRARY
 ggml/src/ggml-metal-embed.o: \
-	ggml/src/ggml-metal/ggml-metal.metal \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
+	ggml/src/ggml-metal.metal \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
-	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
+	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
-	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
 	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
 	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
 	@rmdir ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
-DIR_GGML = ggml
-DIR_LLAMA = src
-DIR_COMMON = common
-
-OBJ_GGML = \
-	$(DIR_GGML)/src/ggml.o \
-	$(DIR_GGML)/src/ggml-aarch64.o \
-	$(DIR_GGML)/src/ggml-alloc.o \
-	$(DIR_GGML)/src/ggml-backend.o \
-	$(DIR_GGML)/src/ggml-backend-reg.o \
-	$(DIR_GGML)/src/ggml-opt.o \
-	$(DIR_GGML)/src/ggml-quants.o \
-	$(DIR_GGML)/src/ggml-threading.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
-	$(OBJ_GGML_EXT)
+OBJ_GGML += \
+	ggml/src/ggml.o \
+	ggml/src/ggml-alloc.o \
+	ggml/src/ggml-backend.o \
+	ggml/src/ggml-quants.o \
+	ggml/src/ggml-aarch64.o
 
 OBJ_LLAMA = \
-	$(DIR_LLAMA)/llama.o \
-	$(DIR_LLAMA)/llama-vocab.o \
-	$(DIR_LLAMA)/llama-grammar.o \
-	$(DIR_LLAMA)/llama-sampling.o \
-	$(DIR_LLAMA)/unicode.o \
-	$(DIR_LLAMA)/unicode-data.o
+	src/llama.o \
+	src/llama-vocab.o \
+	src/llama-grammar.o \
+	src/llama-sampling.o \
+	src/unicode.o \
+	src/unicode-data.o
 
 OBJ_COMMON = \
-	$(DIR_COMMON)/common.o \
-	$(DIR_COMMON)/arg.o \
-	$(DIR_COMMON)/log.o \
-	$(DIR_COMMON)/console.o \
-	$(DIR_COMMON)/ngram-cache.o \
-	$(DIR_COMMON)/sampling.o \
-	$(DIR_COMMON)/speculative.o \
-	$(DIR_COMMON)/build-info.o \
-	$(DIR_COMMON)/json-schema-to-grammar.o
+	common/common.o \
+	common/arg.o \
+	common/log.o \
+	common/console.o \
+	common/ngram-cache.o \
+	common/sampling.o \
+	common/train.o \
+	common/build-info.o \
+	common/json-schema-to-grammar.o
 
 OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
 
@@ -1027,6 +993,7 @@ $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 
 ifndef CUDA_DOCKER_ARCH
@@ -1036,6 +1003,7 @@ endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # GGML_MUSA
 endif # GGML_CUDA
 $(info )
 
@@ -1072,78 +1040,209 @@ endif
 # Build libraries
 #
 
-# Libraries
-LIB_GGML   = libggml.so
-LIB_GGML_S = libggml.a
+# ggml
 
-LIB_LLAMA   = libllama.so
-LIB_LLAMA_S = libllama.a
+ggml/src/ggml.o: \
+	ggml/src/ggml.c \
+	ggml/include/ggml.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-LIB_COMMON   = libcommon.so
-LIB_COMMON_S = libcommon.a
-
-# Targets
-BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
-
-# Dependency files
-DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
-
-# Default target
-all: $(BUILD_TARGETS)
-
-# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
-#       g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
-$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
-	ggml/src/ggml-cpu/ggml-cpu.cpp \
-	ggml/include/ggml-backend.h \
+ggml/src/ggml-alloc.o: \
+	ggml/src/ggml-alloc.c \
 	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h \
+	ggml/include/ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml/src/ggml-backend.o: \
+	ggml/src/ggml-backend.cpp \
 	ggml/src/ggml-backend-impl.h \
-	ggml/include/ggml-cpu.h \
-	ggml/src/ggml-impl.h
-	$(CXX) $(CXXFLAGS)   -c $< -o $@
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-# Rules for building object files
-$(DIR_GGML)/%.o: $(DIR_GGML)/%.c
-	$(CC) $(CFLAGS) -MMD -c $< -o $@
+ggml/src/ggml-quants.o: \
+	ggml/src/ggml-quants.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
 
-$(DIR_GGML)/%.o: $(DIR_GGML)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+ggml/src/ggml-aarch64.o: \
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
 
-$(DIR_LLAMA)/%.o: $(DIR_LLAMA)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+ggml/src/ggml-blas.o: \
+	ggml/src/ggml-blas.cpp \
+	ggml/include/ggml-blas.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-$(DIR_COMMON)/%.o: $(DIR_COMMON)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+ifndef GGML_NO_LLAMAFILE
+ggml/src/llamafile/sgemm.o: \
+	ggml/src/llamafile/sgemm.cpp \
+	ggml/src/llamafile/sgemm.h \
+	ggml/include/ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_NO_LLAMAFILE
 
-# Rules for building libraries
-$(LIB_GGML): $(OBJ_GGML)
+ifdef GGML_RPC
+ggml/src/ggml-rpc.o: \
+	ggml/src/ggml-rpc.cpp \
+	ggml/include/ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_RPC
+
+$(LIB_GGML): \
+	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_GGML_S): $(OBJ_GGML)
+$(LIB_GGML_S): \
+	$(OBJ_GGML)
 	ar rcs $(LIB_GGML_S) $^
 
-$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
+# llama
+
+src/unicode.o: \
+	src/unicode.cpp \
+	src/unicode.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+src/unicode-data.o: \
+	src/unicode-data.cpp \
+	src/unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+src/llama.o: \
+	src/llama.cpp \
+	src/llama-impl.h \
+	src/llama-vocab.h \
+	src/llama-grammar.h \
+	src/llama-sampling.h \
+	src/unicode.h \
+	include/llama.h \
+	ggml/include/ggml-cuda.h \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h \
+	ggml/include/ggml-backend.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+src/llama-vocab.o: \
+	src/llama-vocab.cpp \
+	src/llama-vocab.h \
+	src/llama-impl.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+src/llama-grammar.o: \
+	src/llama-grammar.cpp \
+	src/llama-grammar.h \
+	src/llama-impl.h \
+	src/llama-vocab.h \
+	src/llama-sampling.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+src/llama-sampling.o: \
+	src/llama-sampling.cpp \
+	src/llama-sampling.h \
+	src/llama-impl.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+$(LIB_LLAMA): \
+	$(OBJ_LLAMA) \
+	$(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_LLAMA_S): $(OBJ_LLAMA)
+$(LIB_LLAMA_S): \
+	$(OBJ_LLAMA)
 	ar rcs $(LIB_LLAMA_S) $^
 
-$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
+# common
+
+common/common.o: \
+	common/common.cpp \
+	common/common.h \
+	common/console.h \
+	common/sampling.h \
+	common/json.hpp \
+	common/json-schema-to-grammar.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/arg.o: \
+	common/arg.cpp \
+	common/arg.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/log.o: \
+	common/log.cpp \
+	common/log.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/sampling.o: \
+	common/sampling.cpp \
+	common/sampling.h \
+	include/llama.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/console.o: \
+	common/console.cpp \
+	common/console.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/json-schema-to-grammar.o: \
+	common/json-schema-to-grammar.cpp \
+	common/json-schema-to-grammar.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/train.o: \
+	common/train.cpp \
+	common/train.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/ngram-cache.o: \
+	common/ngram-cache.cpp \
+	common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+$(LIB_COMMON): \
+	$(OBJ_COMMON) \
+	$(LIB_LLAMA) \
+	$(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_COMMON_S): $(OBJ_COMMON)
+$(LIB_COMMON_S): \
+	$(OBJ_COMMON)
 	ar rcs $(LIB_COMMON_S) $^
 
-# Include dependency files
--include $(DEP_FILES)
-
-# Clean rule
 clean:
-	rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf *.a *.dll *.so *.dot
-	find ggml src common tests examples pocs -type f -name "*.o" -delete
-	find ggml src common tests examples pocs -type f -name "*.d" -delete
+	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -rvf src/*.o
+	rm -rvf tests/*.o
+	rm -rvf examples/*.o
+	rm -rvf common/*.o
+	rm -rvf *.a
+	rm -rvf *.dll
+	rm -rvf *.so
+	rm -rvf *.dot
+	rm -rvf ggml/*.a
+	rm -rvf ggml/*.dll
+	rm -rvf ggml/*.so
+	rm -vrf ggml/src/*.o
+	rm -rvf ggml/src/llamafile/*.o
+	rm -rvf common/build-info.cpp
+	rm -vrf ggml/src/ggml-metal-embed.metal
+	rm -vrf ggml/src/ggml-cuda/*.o
+	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
+	rm -rvf $(BUILD_TARGETS)
+	rm -rvf $(TEST_TARGETS)
+	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
+	rm -rvf $(LEGACY_TARGETS_CLEAN)
+	find examples pocs -type f -name "*.o" -delete
 
 #
 # Examples
@@ -1169,21 +1268,11 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-run: examples/run/run.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-simple: examples/simple/simple.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-simple-chat: examples/simple-chat/simple-chat.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-tokenize: examples/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1281,6 +1370,11 @@ llama-bench: examples/llama-bench/llama-bench.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-baby-llama: examples/baby-llama/baby-llama.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-export-lora: examples/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1346,13 +1440,22 @@ llama-server: \
 	examples/server/server.cpp \
 	examples/server/utils.hpp \
 	examples/server/httplib.h \
+	examples/server/colorthemes.css.hpp \
+	examples/server/style.css.hpp \
+	examples/server/theme-beeninorder.css.hpp \
+	examples/server/theme-ketivah.css.hpp \
+	examples/server/theme-mangotango.css.hpp \
+	examples/server/theme-playground.css.hpp \
+	examples/server/theme-polarnight.css.hpp \
+	examples/server/theme-snowstorm.css.hpp \
 	examples/server/index.html.hpp \
+	examples/server/index-new.html.hpp \
+	examples/server/index.js.hpp \
 	examples/server/completion.js.hpp \
+	examples/server/system-prompts.js.hpp \
+	examples/server/prompt-formats.js.hpp \
+	examples/server/json-schema-to-grammar.mjs.hpp \
 	examples/server/loading.html.hpp \
-	examples/server/deps_daisyui.min.css.hpp \
-	examples/server/deps_markdown-it.js.hpp \
-	examples/server/deps_tailwindcss.js.hpp \
-	examples/server/deps_vue.esm-browser.js.hpp \
 	common/json.hpp \
 	common/stb_image.h \
 	$(OBJ_ALL)
@@ -1454,6 +1557,11 @@ tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-grad0: tests/test-grad0.cpp \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-opt: tests/test-opt.cpp \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

Package.swift

@@ -10,16 +10,10 @@ var sources = [
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
-    "ggml/src/ggml-aarch64.c",
     "ggml/src/ggml-alloc.c",
     "ggml/src/ggml-backend.cpp",
-    "ggml/src/ggml-backend-reg.cpp",
-    "ggml/src/ggml-cpu/ggml-cpu.c",
-    "ggml/src/ggml-cpu/ggml-cpu.cpp",
-    "ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
-    "ggml/src/ggml-cpu/ggml-cpu-quants.c",
-    "ggml/src/ggml-threading.cpp",
     "ggml/src/ggml-quants.c",
+    "ggml/src/ggml-aarch64.c",
 ]
 
 var resources: [Resource] = []
@@ -27,7 +21,6 @@ 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)
@@ -36,15 +29,13 @@ var cSettings: [CSetting] =  [
 ]
 
 #if canImport(Darwin)
-sources.append("ggml/src/ggml-common.h")
-sources.append("ggml/src/ggml-metal/ggml-metal.m")
-resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
+sources.append("ggml/src/ggml-metal.m")
+resources.append(.process("ggml/src/ggml-metal.metal"))
 linkerSettings.append(.linkedFramework("Accelerate"))
 cSettings.append(
     contentsOf: [
         .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL"),
-        .define("GGML_USE_CPU")
+        .define("GGML_USE_METAL")
     ]
 )
 #endif
@@ -69,15 +60,13 @@ let package = Package(
             name: "llama",
             path: ".",
             exclude: [
-               "build",
                "cmake",
                "examples",
                "scripts",
                "models",
                "tests",
                "CMakeLists.txt",
-               "Makefile",
-               "ggml/src/ggml-metal-embed.metal"
+               "Makefile"
             ],
             sources: sources,
             resources: resources,

README.md

@@ -17,8 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **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 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)
 
 ----
@@ -30,9 +29,9 @@ 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, AVX512 and AMX support for x86 architectures
+- AVX, AVX2 and AVX512 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 and Moore Threads MTT GPUs via MUSA)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
@@ -79,7 +78,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
-- [x] [OLMo 2](https://allenai.org/olmo)
 - [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
 - [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
@@ -95,7 +93,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
-- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -125,18 +122,14 @@ 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:**
 
@@ -177,7 +170,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [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`)*
 
@@ -193,7 +185,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [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.
@@ -422,7 +413,7 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [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 |
+| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
 | [CUDA](./docs/build.md#cuda) | Nvidia GPU |
 | [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](./docs/build.md#vulkan) | GPU |
@@ -460,14 +451,14 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - 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
+## Other documentations
 
 - [main (cli)](./examples/main/README.md)
 - [server](./examples/server/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [GBNF grammars](./grammars/README.md)
 
-**Development documentation**
+**Development documentations**
 
 - [How to build](./docs/build.md)
 - [Running on Docker](./docs/docker.md)

ci/run.sh

@@ -39,7 +39,7 @@ SRC=`pwd`
 CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
@@ -53,7 +53,7 @@ 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
@@ -326,36 +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 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-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-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-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-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-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-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
+    (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
 
     function check_ppl {
         qnt="$1"
@@ -460,34 +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}  -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-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-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-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-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-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-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
+    (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
 
     function check_ppl {
         qnt="$1"
@@ -591,36 +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 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-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-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-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-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-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-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
+    (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
 
     function check_ppl {
         qnt="$1"
@@ -706,8 +706,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -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
+    (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
 
     set +e
 }
@@ -752,7 +752,7 @@ function gg_run_rerank_tiny {
     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
+    (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." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

cmake/arm64-apple-clang.cmake

@@ -1,16 +0,0 @@
-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/common.cmake

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

cmake/llama-config.cmake.in

@@ -3,60 +3,18 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
 
-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_BLAS       @GGML_BLAS@)
+set(GGML_CUDA       @GGML_CUDA@)
+set(GGML_METAL      @GGML_METAL@)
+set(GGML_HIPBLAS    @GGML_HIPBLAS@)
 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 @GGML_VULKAN@)
 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@)
-
+set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
+set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
+set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
+set(GGML_SYCL @GGML_SYCL@)
+set(GGML_OPENMP @GGML_OPENMP@)
 
 @PACKAGE_INIT@
 
@@ -64,111 +22,65 @@ 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)
 
-set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
-set(_llama_link_deps "")
-set(_llama_link_opts "")
-foreach(_ggml_lib ggml ggml-base)
-    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
-    find_library(${_ggml_lib_var} ${_ggml_lib}
-        REQUIRED
-        HINTS ${LLAMA_LIB_DIR}
-        NO_CMAKE_FIND_ROOT_PATH
-    )
-    list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
-    message(STATUS "Found ${${_ggml_lib_var}}")
-endforeach()
-
-foreach(backend amx blas cann cpu cuda hip kompute metal musa rpc sycl vulkan)
-    string(TOUPPER "GGML_${backend}" backend_id)
-    set(_ggml_lib "ggml-${backend}")
-    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
-
-    find_library(${_ggml_lib_var} ${_ggml_lib}
-        HINTS ${LLAMA_LIB_DIR}
-        NO_CMAKE_FIND_ROOT_PATH
-    )
-    if(${_ggml_lib_var})
-        list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
-        set(${backend_id} ON)
-        message(STATUS "Found backend ${${_ggml_lib_var}}")
-    else()
-        set(${backend_id} OFF)
-    endif()
-endforeach()
-
-if (NOT LLAMA_SHARED_LIB)
-    if (APPLE AND GGML_ACCELERATE)
-        find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
-        list(APPEND _llama_link_deps ${ACCELERATE_FRAMEWORK})
-    endif()
-
-    if (GGML_OPENMP)
-        find_package(OpenMP REQUIRED)
-        list(APPEND _llama_link_deps OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
-    endif()
-
-    if (GGML_CPU_HBM)
-        find_library(memkind memkind REQUIRED)
-        list(APPEND _llama_link_deps memkind)
-    endif()
-
-    if (GGML_BLAS)
-        find_package(BLAS REQUIRED)
-        list(APPEND _llama_link_deps ${BLAS_LIBRARIES})
-        list(APPEND _llama_link_opts ${BLAS_LINKER_FLAGS})
-    endif()
-
-    if (GGML_CUDA)
-        find_package(CUDAToolkit REQUIRED)
-    endif()
-
-    if (GGML_METAL)
-        find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-        find_library(METAL_FRAMEWORK    Metal REQUIRED)
-        find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
-        list(APPEND _llama_link_deps ${FOUNDATION_LIBRARY}
-                                     ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
-    endif()
-
-    if (GGML_VULKAN)
-        find_package(Vulkan REQUIRED)
-        list(APPEND _llama_link_deps Vulkan::Vulkan)
-    endif()
-
-    if (GGML_HIP)
-        find_package(hip     REQUIRED)
-        find_package(hipblas REQUIRED)
-        find_package(rocblas REQUIRED)
-        list(APPEND _llama_link_deps hip::host roc::rocblas roc::hipblas)
-    endif()
-
-    if (GGML_SYCL)
-        find_package(DNNL)
-        if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
-            list(APPEND _llama_link_deps DNNL::dnnl)
-        endif()
-        if (WIN32)
-            find_package(IntelSYCL REQUIRED)
-            find_package(MKL       REQUIRED)
-            list(APPEND _llama_link_deps IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
-        endif()
-    endif()
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
 endif()
 
+if (GGML_BLAS)
+    find_package(BLAS REQUIRED)
+endif()
+
+if (GGML_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK Metal REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+endif()
+
+if (GGML_VULKAN)
+    find_package(Vulkan REQUIRED)
+endif()
+
+if (GGML_HIPBLAS)
+    find_package(hip REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+endif()
+
+if (GGML_SYCL)
+    find_package(IntelSYCL REQUIRED)
+    find_package(MKL REQUIRED)
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP REQUIRED)
+endif()
+
+
+find_library(ggml_LIBRARY ggml
+    REQUIRED
+    HINTS ${LLAMA_LIB_DIR})
+
 find_library(llama_LIBRARY llama
     REQUIRED
-    HINTS ${LLAMA_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH
-)
+    HINTS ${LLAMA_LIB_DIR})
+
+set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
+set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
 
 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

@@ -2,8 +2,6 @@
 
 find_package(Threads REQUIRED)
 
-llama_add_compile_flags()
-
 # Build info header
 #
 
@@ -68,8 +66,8 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     sampling.cpp
     sampling.h
-    speculative.cpp
-    speculative.h
+    train.cpp
+    train.h
     )
 
 if (BUILD_SHARED_LIBS)

common/common.cpp

@@ -12,7 +12,6 @@
 
 #include <algorithm>
 #include <cinttypes>
-#include <climits>
 #include <cmath>
 #include <codecvt>
 #include <cstdarg>
@@ -24,10 +23,10 @@
 #include <regex>
 #include <sstream>
 #include <string>
-#include <thread>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+#include <thread>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -401,19 +400,17 @@ std::string common_params_get_system_info(const common_params & params) {
 // String utils
 //
 
-std::string string_format(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
+std::vector<std::string> string_split(std::string input, char separator) {
+    std::vector<std::string> parts;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(0, separator_pos);
+        parts.emplace_back(part);
+        input = input.substr(separator_pos + 1);
+        separator_pos = input.find(separator);
+    }
+    parts.emplace_back(input);
+    return parts;
 }
 
 std::string string_strip(const std::string & str) {
@@ -536,12 +533,12 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
                     [](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 << "\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 << " ]";
@@ -829,9 +826,9 @@ struct common_init_result common_init_from_params(common_params & params) {
     llama_model * model = nullptr;
 
     if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
+        model = common_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else if (!params.model_url.empty()) {
-        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
+        model = common_load_model_from_url(params.model_url.c_str(), params.model.c_str(), params.hf_token.c_str(), mparams);
     } else {
         model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
@@ -875,12 +872,6 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_ERR("%s: KV cache shifting is not supported for this model (--no-context-shift to disable)'\n", __func__);
-        llama_free_model(model);
-        return iparams;
-    }
-
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
         if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
@@ -925,9 +916,9 @@ struct common_init_result common_init_from_params(common_params & params) {
         common_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sampling.ignore_eos = false;
+        params.sparams.ignore_eos = false;
     }
 
     if (params.warmup) {
@@ -948,7 +939,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (llama_model_has_encoder(model)) {
-            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
             if (decoder_start_token_id == -1) {
                 decoder_start_token_id = bos;
@@ -957,7 +948,7 @@ struct common_init_result common_init_from_params(common_params & params) {
             tmp.push_back(decoder_start_token_id);
         }
         if (llama_model_has_decoder(model)) {
-            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
         }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
@@ -979,12 +970,9 @@ void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_l
     }
 }
 
-struct llama_model_params common_model_params_to_llama(common_params & params) {
+struct llama_model_params common_model_params_to_llama(const common_params & params) {
     auto mparams = llama_model_default_params();
 
-    if (!params.devices.empty()) {
-        mparams.devices = params.devices.data();
-    }
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
@@ -1012,9 +1000,6 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
-    if (s == "bf16") {
-        return GGML_TYPE_BF16;
-    }
     if (s == "q8_0") {
         return GGML_TYPE_Q8_0;
     }
@@ -1034,7 +1019,7 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
         return GGML_TYPE_Q5_1;
     }
 
-    throw std::runtime_error("Unsupported cache type: " + s);
+    throw std::runtime_error("Invalid cache type: " + s);
 }
 
 struct llama_context_params common_context_params_to_llama(const common_params & params) {
@@ -1046,7 +1031,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
-                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -1342,17 +1327,17 @@ static bool common_download_file(const std::string & url, const std::string & pa
 }
 
 struct llama_model * common_load_model_from_url(
-        const std::string & model_url,
-        const std::string & local_path,
-        const std::string & hf_token,
+        const char * model_url,
+        const char * path_model,
+        const char * hf_token,
         const struct llama_model_params & params) {
     // Basic validation of the model_url
-    if (model_url.empty()) {
+    if (!model_url || strlen(model_url) == 0) {
         LOG_ERR("%s: invalid model_url\n", __func__);
         return NULL;
     }
 
-    if (!common_download_file(model_url, local_path, hf_token)) {
+    if (!common_download_file(model_url, path_model, hf_token)) {
         return NULL;
     }
 
@@ -1363,9 +1348,9 @@ struct llama_model * common_load_model_from_url(
             /*.no_alloc = */ true,
             /*.ctx      = */ NULL,
         };
-        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
+        auto * ctx_gguf = gguf_init_from_file(path_model, gguf_params);
         if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
+            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, path_model);
             return NULL;
         }
 
@@ -1384,13 +1369,13 @@ struct llama_model * common_load_model_from_url(
         // Verify the first split file format
         // and extract split URL and PATH prefixes
         {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), path_model, 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, path_model, n_split);
                 return NULL;
             }
 
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url, 0, n_split)) {
+                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url, n_split);
                 return NULL;
             }
         }
@@ -1417,14 +1402,14 @@ struct llama_model * common_load_model_from_url(
         }
     }
 
-    return llama_load_model_from_file(local_path.c_str(), params);
+    return llama_load_model_from_file(path_model, params);
 }
 
 struct llama_model * common_load_model_from_hf(
-        const std::string & repo,
-        const std::string & remote_path,
-        const std::string & local_path,
-        const std::string & hf_token,
+        const char * repo,
+        const char * model,
+        const char * path_model,
+        const char * hf_token,
         const struct llama_model_params & params) {
     // construct hugging face model url:
     //
@@ -1438,27 +1423,27 @@ struct llama_model * common_load_model_from_hf(
     std::string model_url = "https://huggingface.co/";
     model_url += repo;
     model_url += "/resolve/main/";
-    model_url += remote_path;
+    model_url += model;
 
-    return common_load_model_from_url(model_url, local_path, hf_token, params);
+    return common_load_model_from_url(model_url.c_str(), path_model, hf_token, params);
 }
 
 #else
 
 struct llama_model * common_load_model_from_url(
-        const std::string & /*model_url*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
+        const char * /*model_url*/,
+        const char * /*path_model*/,
+        const char * /*hf_token*/,
         const struct llama_model_params & /*params*/) {
     LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
     return nullptr;
 }
 
 struct llama_model * common_load_model_from_hf(
-        const std::string & /*repo*/,
-        const std::string & /*remote_path*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
+        const char * /*repo*/,
+        const char * /*model*/,
+        const char * /*path_model*/,
+        const char * /*hf_token*/,
         const struct llama_model_params & /*params*/) {
     LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
     return nullptr;
@@ -1493,66 +1478,6 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
-//
-// Token utils
-//
-
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 //
 // Vocab utils
 //
@@ -1959,3 +1884,211 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
     return result;
 }
 
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float(FILE * stream, const char * prop_name, const std::vector<float> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%e, ", data[i]);
+    }
+    fprintf(stream, "%e]\n", data.back());
+}
+
+void yaml_dump_vector_int(FILE * stream, const char * prop_name, const std::vector<int> & data) {
+    if (data.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    fprintf(stream, "%s: [", prop_name);
+    for (size_t i = 0; i < data.size() - 1; ++i) {
+        fprintf(stream, "%d, ", data[i]);
+    }
+    fprintf(stream, "%d]\n", data.back());
+}
+
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data) {
+    std::string data_str(data == NULL ? "" : data);
+
+    if (data_str.empty()) {
+        fprintf(stream, "%s:\n", prop_name);
+        return;
+    }
+
+    size_t pos_start = 0;
+    size_t pos_found = 0;
+
+    if (std::isspace(data_str[0]) || std::isspace(data_str.back())) {
+        data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
+        data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
+        data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
+        data_str = "\"" + data_str + "\"";
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    if (data_str.find('\n') == std::string::npos) {
+        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
+        return;
+    }
+
+    fprintf(stream, "%s: |\n", prop_name);
+    while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
+        fprintf(stream, "  %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
+        pos_start = pos_found + 1;
+    }
+}
+
+void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
+                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    const auto & sparams = params.sparams;
+
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
+    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
+    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
+    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
+    fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
+    fprintf(stream, "cpu_has_cuda: %s\n",        ggml_cpu_has_cuda()        ? "true" : "false");
+    fprintf(stream, "cpu_has_vulkan: %s\n",      ggml_cpu_has_vulkan()      ? "true" : "false");
+    fprintf(stream, "cpu_has_kompute: %s\n",     ggml_cpu_has_kompute()     ? "true" : "false");
+    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
+    fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
+    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sve: %s\n",         ggml_cpu_has_sve()         ? "true" : "false");
+    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
+    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
+    fprintf(stream, "cpu_has_riscv_v: %s\n",     ggml_cpu_has_riscv_v()     ? "true" : "false");
+    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
+    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
+
+#ifdef NDEBUG
+    fprintf(stream, "debug: false\n");
+#else
+    fprintf(stream, "debug: true\n");
+#endif // NDEBUG
+
+    fprintf(stream, "model_desc: %s\n", model_desc);
+    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
+
+#ifdef __OPTIMIZE__
+    fprintf(stream, "optimize: true\n");
+#else
+    fprintf(stream, "optimize: false\n");
+#endif // __OPTIMIZE__
+
+    fprintf(stream, "time: %s\n", timestamp.c_str());
+
+    fprintf(stream, "\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "# User Inputs #\n");
+    fprintf(stream, "###############\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
+    fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
+    fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
+    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
+    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
+    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
+    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
+    yaml_dump_string_multiline(stream, "grammar", sparams.grammar.c_str());
+    fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
+    fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
+    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
+    fprintf(stream, "ignore_eos: %s # default: false\n", sparams.ignore_eos ? "true" : "false");
+
+    yaml_dump_string_multiline(stream, "in_prefix", params.input_prefix.c_str());
+    fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
+    yaml_dump_string_multiline(stream, "in_suffix", params.input_prefix.c_str());
+    fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
+    fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
+    fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
+    fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
+
+    fprintf(stream, "logit_bias:\n");
+    for (const auto & logit_bias : sparams.logit_bias) {
+        fprintf(stream, "  %d: %f", logit_bias.token, logit_bias.bias);
+    }
+
+    fprintf(stream, "lora:\n");
+    for (auto & la : params.lora_adapters) {
+        if (la.scale == 1.0f) {
+            fprintf(stream, "  - %s\n", la.path.c_str());
+        }
+    }
+    fprintf(stream, "lora_scaled:\n");
+    for (auto & la : params.lora_adapters) {
+        if (la.scale != 1.0f) {
+            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
+        }
+    }
+    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
+    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
+    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
+    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
+    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
+    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
+    fprintf(stream, "model: %s # default: %s\n", params.model.c_str(), DEFAULT_MODEL_PATH);
+    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
+    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
+    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
+    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
+    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
+    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
+    fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
+    fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
+    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
+    yaml_dump_string_multiline(stream, "prompt", params.prompt.c_str());
+    fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
+    fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
+    fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
+    yaml_dump_vector_int(stream, "prompt_tokens", prompt_tokens);
+    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
+
+    fprintf(stream, "reverse_prompt:\n");
+    for (std::string ap : params.antiprompt) {
+        size_t pos = 0;
+        while ((pos = ap.find('\n', pos)) != std::string::npos) {
+            ap.replace(pos, 1, "\\n");
+            pos += 1;
+        }
+
+        fprintf(stream, "  - %s\n", ap.c_str());
+    }
+
+    fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
+    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
+    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
+    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
+    fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
+    fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
+
+    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
+    yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
+
+    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
+    fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
+    fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
+    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
+    fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
+    fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
+}

common/common.h

@@ -33,8 +33,6 @@ struct common_lora_adapter_container : common_lora_adapter_info {
     struct llama_lora_adapter * adapter;
 };
 
-using llama_tokens = std::vector<llama_token>;
-
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -86,15 +84,12 @@ enum llama_example {
 
 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,
+    COMMON_SAMPLER_TYPE_TOP_K       = 1,
+    COMMON_SAMPLER_TYPE_TOP_P       = 2,
+    COMMON_SAMPLER_TYPE_MIN_P       = 3,
+    COMMON_SAMPLER_TYPE_TFS_Z       = 4,
+    COMMON_SAMPLER_TYPE_TYPICAL_P   = 5,
+    COMMON_SAMPLER_TYPE_TEMPERATURE = 6,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -103,48 +98,39 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
-// sampling parameters
-struct common_params_sampling {
+// sampler parameters
+struct common_sampler_params {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    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
-
+    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   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
+    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<enum common_sampler_type> samplers = {
-        COMMON_SAMPLER_TYPE_DRY,
         COMMON_SAMPLER_TYPE_TOP_K,
+        COMMON_SAMPLER_TYPE_TFS_Z,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
         COMMON_SAMPLER_TYPE_TOP_P,
         COMMON_SAMPLER_TYPE_MIN_P,
-        COMMON_SAMPLER_TYPE_XTC,
-        COMMON_SAMPLER_TYPE_TEMPERATURE,
+        COMMON_SAMPLER_TYPE_TEMPERATURE
     };
 
     std::string grammar; // optional BNF-like grammar to constrain sampling
@@ -155,30 +141,21 @@ struct common_params_sampling {
     std::string print() const;
 };
 
-struct common_params_speculative {
-    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
-    int32_t n_ctx        =     0; // draft context size
-    int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
-    int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
-    float   p_split      =  0.1f; // speculative decoding split probability
-    float   p_min        =  0.9f; // minimum speculative decoding probability (greedy)
-
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
-
-    std::string model = ""; // draft model for speculative decoding                          // NOLINT
-};
-
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =  4096; // context size
+    int32_t n_ctx                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
+    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
     int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
+    float   p_split               =  0.1f; // speculative decoding split probability
+    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -189,31 +166,27 @@ struct common_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          =  0.1f; // KV cache defragmentation threshold
-
-    // offload params
-    std::vector<ggml_backend_dev_t> devices;         // devices to use for offloading
-    int32_t n_gpu_layers                    =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu                        =     0; // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]               =   {0}; // how split tensors should be distributed across GPUs
-    enum llama_split_mode        split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    float   defrag_thold          = -1.0f; // 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;
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
+    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct common_params_sampling sampling;
-    struct common_params_speculative speculative;
+    struct common_sampler_params sparams;
 
     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
@@ -224,6 +197,7 @@ struct common_params {
     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 logdir               = ""; // directory in which to save YAML log files                     // 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
@@ -294,21 +268,21 @@ struct common_params {
 
     // embedding
     bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    int32_t embd_normalize = 2;     // normalisation for embendings (-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 embeddings
+    std::string embd_sep   = "\n";  // separator of embendings
     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 (TODO: support threadpool)
-    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
+    std::string system_prompt = "";                                                                         // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
@@ -378,27 +352,16 @@ void common_init();
 
 std::string common_params_get_system_info(const common_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 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
 //
 
-#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::vector<std::string> string_split(std::string input, char separator);
 
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
@@ -407,7 +370,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
-    static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
     std::vector<T> values;
     std::istringstream str_stream(str);
     std::string token;
@@ -420,22 +382,6 @@ 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);
 
@@ -466,28 +412,17 @@ struct common_init_result {
 
 struct common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model_params     common_model_params_to_llama  (      common_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 std::string & model_url,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-struct llama_model * common_load_model_from_hf(
-    const std::string & repo,
-    const std::string & remote_path,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
+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 common_batch_clear(struct llama_batch & batch);
 
@@ -498,16 +433,6 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
-//
-// Token utils
-//
-
-// longest common prefix
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
-
-// longet common subsequence
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
-
 //
 // Vocab utils
 //
@@ -619,3 +544,15 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 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 common_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

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/sampling.cpp

@@ -99,7 +99,7 @@ struct ring_buffer {
 };
 
 struct common_sampler {
-    common_params_sampling params;
+    common_sampler_params params;
 
     struct llama_sampler * grmr;
     struct llama_sampler * chain;
@@ -125,23 +125,21 @@ struct common_sampler {
     }
 };
 
-std::string common_params_sampling::print() const {
+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"
-            "\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"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             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,
+            top_k, tfs_z, top_p, min_p, typ_p, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
+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;
@@ -173,54 +171,54 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.penalize_nl,
                 params.ignore_eos));
 
-    if (params.mirostat == 0) {
-        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()));
-                    }
+    if (params.temp > 0.0f) {
+        if (params.mirostat == 0) {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    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_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");
+                    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_TFS_Z:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
+                        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;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
             }
+            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+            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 {
+            GGML_ASSERT(false && "unknown mirostat version");
         }
-        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 {
-        GGML_ASSERT(false && "unknown mirostat version");
+        if (params.n_probs > 0) {
+            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
+            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
+            //
+            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
+            // it is much faster, since we avoid sorting all tokens and should give a good approximation
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        }
+        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
     }
 
     return result;
@@ -320,45 +318,6 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     return cur_p.data[cur_p.selected].id;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
-    GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
-
-    std::vector<llama_token> result;
-    result.reserve(idxs.size());
-
-    size_t i = 0;
-    for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
-
-        common_sampler_accept(gsmpl, id, true);
-
-        result.push_back(id);
-
-        if (draft[i] != id) {
-            break;
-        }
-    }
-
-    if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
-
-        common_sampler_accept(gsmpl, id, true);
-
-        result.push_back(id);
-    }
-
-    return result;
-}
-
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
-    std::vector<int> idxs(draft.size() + 1);
-    for (size_t i = 0; i < idxs.size(); ++i) {
-        idxs[i] = i;
-    }
-
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
-}
-
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
     return llama_sampler_get_seed(gsmpl->chain);
 }
@@ -407,42 +366,36 @@ std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_
 
 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_TFS_Z:       return 'f';
         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_TFS_Z:       return "tfs_z";
         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<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 },
+        { "tfs_z",       COMMON_SAMPLER_TYPE_TFS_Z },
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
-        { "xtc",         COMMON_SAMPLER_TYPE_XTC },
-        { "infill",      COMMON_SAMPLER_TYPE_INFILL },
     };
 
     // since samplers names are written multiple ways
@@ -456,6 +409,8 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "tfs-z",       COMMON_SAMPLER_TYPE_TFS_Z },
+        { "tfs",         COMMON_SAMPLER_TYPE_TFS_Z },
         { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
     };
 
@@ -481,14 +436,12 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
 
 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_TFS_Z),       COMMON_SAMPLER_TYPE_TFS_Z },
         { 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 },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }
     };
 
     std::vector<common_sampler_type> samplers;

common/sampling.h

@@ -36,7 +36,7 @@ struct common_sampler;
 
 // llama_sampler API overloads
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
+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);
 
@@ -60,27 +60,6 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
 //
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
-// generalized version of common_sampler_sample
-//
-// will cross-reference the sampled tokens with a batch of draft tokens and accept those that match
-// if the sampler disagrees at some point, we stop and return the accepted tokens up to now
-//
-//      common_sampler_sample_n(gsmpl, ctx, { idx }, {});
-//
-// is equivalent to
-//
-//      common_sampler_sample(gsmpl, ctx, idx);
-//      common_sampler_accept(gsmpl, token, true);
-//
-// requires: idxs.size() == draft.size() + 1
-//
-// returns at least 1 token, up to idxs.size()
-//
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
-
-// assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
-
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 
 // helpers

common/speculative.cpp

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

common/speculative.h

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

common/train.h

@@ -0,0 +1,233 @@
+// 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.py

@@ -72,8 +72,7 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
-                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -88,7 +87,7 @@ class Model:
         self.is_safetensors = len(self.part_names) > 0
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
-        self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
+        self.hparams = Model.load_hparams(self.dir_model)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
@@ -574,9 +573,6 @@ class Model:
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
-        if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
-            # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
-            res = "bert-bge-large"
         if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
             # ref: https://huggingface.co/mosaicml/mpt-7b
             res = "mpt"
@@ -1542,17 +1538,6 @@ class LlamaModel(Model):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-        # Apply to granite small models only
-        if self.hparams.get("vocab_size", 32000) == 49152:
-            self.gguf_writer.add_add_bos_token(False)
-
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -1569,6 +1554,17 @@ class LlamaModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:
@@ -2707,7 +2703,7 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_token_scores(scores)
         self.gguf_writer.add_token_types(toktypes)
         self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
+        self.gguf_writer.add_token_type_count(1)
         self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
         if precompiled_charsmap:
             self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
@@ -2868,9 +2864,6 @@ class Rwkv6Model(Model):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
-        # hack: Add '\n\n' as the EOT token to make it chat normally
-        special_vocab._set_special_token("eot", 261)
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -3040,11 +3033,6 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("Olmo2ForCausalLM")
-class Olmo2Model(Model):
-    model_arch = gguf.MODEL_ARCH.OLMO2
-
-
 @Model.register("OlmoeForCausalLM")
 class OlmoeModel(Model):
     model_arch = gguf.MODEL_ARCH.OLMOE
@@ -3753,7 +3741,10 @@ class JaisModel(Model):
 
         # Embeddings scale
         self.embeddings_scale = 1.0
+        # note: For some JAIS flavors, output is tied to (same as) wte in original model
+        self.output_is_wte = False
         if 'mup_embeddings_scale' in self.hparams:
+            self.output_is_wte = True   # Hack (?)
             self.embeddings_scale = self.hparams['mup_embeddings_scale']
         elif 'embeddings_scale' in self.hparams:
             self.embeddings_scale = self.hparams['embeddings_scale']
@@ -3810,7 +3801,10 @@ class JaisModel(Model):
 
         if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
             tensors.append((new_name, data_torch * self.embeddings_scale))
+            if self.output_is_wte:
+                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
         elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            assert not self.output_is_wte
             tensors.append((new_name, data_torch * self.width_scale))
         else:
             tensors.append((new_name, data_torch))

convert_hf_to_gguf_update.py

@@ -72,7 +72,6 @@ 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", },

convert_lora_to_gguf.py

@@ -12,7 +12,6 @@ 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
 
@@ -231,7 +230,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
-        description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
+        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
     parser.add_argument(
         "--outfile", type=Path,
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
@@ -257,23 +256,17 @@ def parse_args() -> argparse.Namespace:
         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",
+        "--base", type=Path, required=True,
+        help="directory containing base model file",
     )
     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)",
+        help="directory containing LoRA adapter file",
     )
 
     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)
@@ -288,7 +281,7 @@ if __name__ == '__main__':
 
     ftype = ftype_map[args.outtype]
 
-    dir_base_model: Path | None = args.base
+    dir_base_model: Path = args.base
     dir_lora: Path = args.lora_path
     lora_config = dir_lora / "adapter_config.json"
     input_model = dir_lora / "adapter_model.safetensors"
@@ -308,29 +301,9 @@ if __name__ == '__main__':
         input_model = os.path.join(dir_lora, "adapter_model.bin")
         lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
 
-    # load LoRA config
-    with open(lora_config, "r") as f:
-        lparams: dict[str, Any] = json.load(f)
-
     # load base model
-    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)
-
+    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])
@@ -350,15 +323,13 @@ if __name__ == '__main__':
                 self.dir_model_card = dir_lora_model
                 self.lora_alpha = float(lora_alpha)
 
-            def set_vocab(self):
-                pass
-
             def set_type(self):
                 self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
+                super().set_gguf_parameters()
 
             def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                 # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
@@ -377,9 +348,6 @@ if __name__ == '__main__':
                         if ".base_layer.weight" in name:
                             continue
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
-                        if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
-                            logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
-                            logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
                         sys.exit(1)
 
                     if base_name in tensor_map:
@@ -413,6 +381,9 @@ if __name__ == '__main__':
                     yield (dest_name + ".lora_a", lora_a)
                     yield (dest_name + ".lora_b", lora_b)
 
+        with open(lora_config, "r") as f:
+            lparams: dict[str, Any] = json.load(f)
+
         alpha: float = lparams["lora_alpha"]
 
         model_instance = LoraModel(
@@ -425,7 +396,6 @@ if __name__ == '__main__':
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
-            hparams=hparams,
         )
 
         logger.info("Exporting model...")

docs/android.md

@@ -23,10 +23,10 @@ $ 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-cli -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
+$ ./build/bin/llama-simple -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
 ```
 
-Here, we show `llama-cli`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
+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:
 

docs/backend/CANN.md

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

docs/backend/SYCL.md

@@ -34,16 +34,13 @@ 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| 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||
+|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|
 
 
 ## 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.
@@ -313,14 +310,12 @@ 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 -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # Option 2: Use FP16
-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
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -338,9 +333,8 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE
 
 ## 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=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Find your GGML_SYCL_HIP_TARGET with rocminfo, under the key 'Name:'
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_HIP_TARGET=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -383,7 +377,7 @@ found 2 SYCL devices:
 
 |Chosen Device ID|Setting|
 |-|-|
-|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
 |1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
 |0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
@@ -650,7 +644,6 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |--------------------|---------------------------------------|---------------------------------------------|
 | 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. |

docs/build.md

@@ -186,16 +186,18 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+| GGML_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
+| GGML_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
 | GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
+| GGML_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
 ### MUSA
 
-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
@@ -207,12 +209,6 @@ This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GP
   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.
@@ -221,12 +217,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 
 - Using `make`:
   ```bash
-  make GGML_HIP=1
+  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 -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`.
@@ -243,19 +239,19 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
   HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      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_HIP=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
+  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 -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)
@@ -264,6 +260,13 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
+| Option                 | Legal values           | Default | Description                                                                                                                                                                                                                                    |
+|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+| GGML_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
+| GGML_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
 
 ### Vulkan
 
@@ -271,9 +274,9 @@ If your GPU is not officially supported you can use the environment variable [`H
 
 #### w64devkit
 
-Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
+Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
 
-Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
+Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
 
 Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
 ```sh
@@ -291,29 +294,6 @@ 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
@@ -387,7 +367,7 @@ 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`
+`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
 
 If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
 ```bash

examples/CMakeLists.txt

@@ -6,20 +6,20 @@ find_package(Threads REQUIRED)
 
 # ...
 
-# flags
-
-llama_add_compile_flags()
-
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
+    add_subdirectory(cvector-generator)
+    add_subdirectory(baby-llama)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)
+    add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
+    add_subdirectory(export-lora)
     add_subdirectory(gbnf-validator)
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
@@ -28,36 +28,27 @@ else()
     add_subdirectory(imatrix)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
+    add_subdirectory(llava)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(main)
     add_subdirectory(parallel)
     add_subdirectory(passkey)
     add_subdirectory(perplexity)
+    add_subdirectory(quantize-stats)
     add_subdirectory(quantize)
     add_subdirectory(retrieval)
+    if (GGML_RPC)
+        add_subdirectory(rpc)
+    endif()
     if (LLAMA_BUILD_SERVER)
-        add_subdirectory(server)
+    add_subdirectory(server)
+    endif()
+    if (GGML_SYCL)
+        add_subdirectory(sycl)
     endif()
     add_subdirectory(save-load-state)
-    add_subdirectory(run)
     add_subdirectory(simple)
-    add_subdirectory(simple-chat)
     add_subdirectory(speculative)
-    add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
-    if (NOT GGML_BACKEND_DL)
-        # these examples use the backends directly and cannot be built with dynamic loading
-        add_subdirectory(convert-llama2c-to-ggml)
-        add_subdirectory(cvector-generator)
-        add_subdirectory(export-lora)
-        add_subdirectory(quantize-stats)
-        add_subdirectory(llava)
-        if (GGML_RPC)
-            add_subdirectory(rpc)
-        endif()
-        if (GGML_SYCL)
-            add_subdirectory(sycl)
-        endif()
-    endif()
 endif()

examples/baby-llama/CMakeLists.txt

@@ -1,5 +1,5 @@
-set(TARGET llama-speculative-simple)
-add_executable(${TARGET} speculative-simple.cpp)
+set(TARGET llama-baby-llama)
+add_executable(${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/batched-bench.cpp

@@ -74,6 +74,7 @@ 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);

examples/batched/batched.cpp

@@ -68,10 +68,10 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sampling.top_k));
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sampling.top_p, params.sampling.min_keep));
-    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sampling.temp));
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sampling.seed));
+    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) {
         LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);

examples/chat-persistent.sh

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

examples/convert_legacy_llama.py

@@ -840,8 +840,6 @@ class OutputFile:
                         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:
@@ -851,32 +849,12 @@ class OutputFile:
                     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)
+            if metadata.datasets is not None:
+                self.gguf.add_datasets(metadata.datasets)
 
     def add_meta_arch(self, params: Params) -> None:
         # Metadata About The Neural Architecture Itself

examples/cvector-generator/cvector-generator.cpp

@@ -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()))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }

examples/eval-callback/CMakeLists.txt

@@ -5,6 +5,5 @@ target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
 set(TEST_TARGET test-eval-callback)
-add_test(NAME ${TEST_TARGET}
-        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
 set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

examples/eval-callback/eval-callback.cpp

@@ -131,7 +131,7 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;
     }

examples/gbnf-validator/gbnf-validator.cpp

@@ -11,19 +11,15 @@
 static bool llama_grammar_validate(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
     const auto cpts = unicode_cpts_from_utf8(input_str);
 
-    const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
-          llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
+    auto & stacks_cur = llama_grammar_get_stacks(grammar);
 
     size_t pos = 0;
     for (const auto & cpt : cpts) {
-        const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
-
-        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
+        llama_grammar_accept(grammar, cpt);
 
         if (stacks_cur.empty()) {
             error_pos = pos;
             error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
-            stacks_cur = stacks_prev;
             return false;
         }
         ++pos;
@@ -82,7 +78,8 @@ int main(int argc, char** argv) {
 
     llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
     if (grammar == nullptr) {
-        throw std::runtime_error("Failed to initialize llama_grammar");
+        fprintf(stdout, "Failed to initialize llama_grammar\n");
+        return 1;
     }
     // Read the input file
     std::string input_str;

examples/gguf-hash/CMakeLists.txt

@@ -4,17 +4,10 @@ install(TARGETS ${TARGET} RUNTIME)
 
 # clibs dependencies
 include_directories(deps/)
-
 add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
 target_link_libraries(${TARGET} PRIVATE xxhash)
-
 add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
 target_link_libraries(${TARGET} PRIVATE sha1)
-if (NOT MSVC)
-    # disable warnings in 3rd party code
-    target_compile_options(sha1 PRIVATE -w)
-endif()
-
 add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
 target_link_libraries(${TARGET} PRIVATE sha256)
 

examples/imatrix/imatrix.cpp

@@ -496,8 +496,6 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
-
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -510,14 +508,9 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            common_batch_clear(batch);
-            for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
-            }
-
-            if (llama_decode(ctx, batch)) {
+            // TODO: use batch.logits to save computations instead of relying on logits_all == true
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
                 return false;
             }
 
@@ -530,8 +523,6 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
             }
         }
 
-        llama_batch_free(batch);
-
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {

examples/infill/infill.cpp

@@ -43,6 +43,50 @@ static std::vector<llama_token> * g_output_tokens;
 
 static bool is_interacting = false;
 
+static void write_logfile(
+    const llama_context * ctx, const common_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = string_get_sortable_timestamp();
+
+    const bool success = fs_create_directory_with_parents(params.logdir);
+    if (!success) {
+        LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: infill\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
+
+    llama_perf_dump_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
@@ -52,6 +96,7 @@ static void sigint_handler(int signo) {
             console::cleanup();
             LOG("\n");
             common_perf_print(*g_ctx, *g_smpl);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
 
             // make sure all logs are flushed
             LOG("Interrupted by user\n");
@@ -73,7 +118,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sampling;
+    auto & sparams = params.sparams;
 
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
@@ -160,11 +205,11 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
     std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-    GGML_ASSERT(llama_token_fim_pre(model) >= 0);
-    GGML_ASSERT(llama_token_fim_suf(model) >= 0);
+    GGML_ASSERT(llama_token_prefix(model) >= 0);
+    GGML_ASSERT(llama_token_suffix(model) >= 0);
 
-    inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
-    inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
 
     embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
     embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -173,7 +218,7 @@ int main(int argc, char ** argv) {
     }
     embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
 
-    const llama_token middle_token = llama_token_fim_mid(model);
+    const llama_token middle_token = llama_token_middle(model);
     if (middle_token >= 0) {
         embd_inp.push_back(middle_token);
     }
@@ -351,7 +396,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -463,8 +508,8 @@ int main(int argc, char ** argv) {
                 std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
                 std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-                inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
-                inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
 
                 embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
                 embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -580,6 +625,7 @@ int main(int argc, char ** argv) {
 
     LOG("\n");
     common_perf_print(ctx, smpl);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);

examples/json_schema_to_grammar.py

@@ -540,7 +540,7 @@ class SchemaConverter:
         return self._add_rule(
             name,
             to_rule(transform()) if self._raw_pattern \
-                else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space")
+                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
 
 
     def _resolve_ref(self, ref):

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -283,6 +283,9 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
         nullptr,
         nullptr,
         nullptr,
+        0,
+        0,
+        0,
     };
 
     if (embd) {

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -46,6 +46,7 @@ actor LlamaContext {
         let sparams = llama_sampler_chain_default_params()
         self.sampling = llama_sampler_chain_init(sparams)
         llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
         llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 

examples/llama.vim

@@ -1,783 +0,0 @@
-" LLM-based text completion using llama.cpp
-"
-" requires:
-"
-"   - neovim or vim
-"   - curl
-"   - llama.cpp server instance
-"   - FIM-compatible model
-"
-" sample config:
-"
-"   - Tab       - accept the current suggestion
-"   - Shift+Tab - accept just the first line of the suggestion
-"   - Ctrl+F    - toggle FIM completion manually
-"
-" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
-"
-" start the llama.cpp server with a FIM-compatible model. for example:
-"
-"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
-"
-"   --batch-size [512, model max context]
-"
-"     adjust the batch size to control how much of the provided local context will be used during the inference
-"     lower values will use smaller part of the context around the cursor, which will result in faster processing
-"
-"   --ubatch-size [64, 2048]
-"
-"     chunks the batch into smaller chunks for faster processing
-"     depends on the specific hardware. use llama-bench to profile and determine the best size
-"
-"   --cache-reuse (ge:llama_config.n_predict, 1024]
-"
-"     this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
-"     using non-zero value enables context reuse on the server side which dramatically improves the performance at
-"     large contexts. a value of 256 should be good for all cases
-"
-" run this once to initialise llama.vim:
-"
-"   :call llama#init()
-"
-" more info: https://github.com/ggerganov/llama.cpp/pull/9787
-"
-
-" colors (adjust to your liking)
-highlight llama_hl_hint guifg=#ff772f ctermfg=202
-highlight llama_hl_info guifg=#77ff2f ctermfg=119
-
-" general parameters:
-"
-"   endpoint:         llama.cpp server endpoint
-"   n_prefix:         number of lines before the cursor location to include in the local prefix
-"   n_suffix:         number of lines after  the cursor location to include in the local suffix
-"   n_predict:        max number of tokens to predict
-"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
-"   t_max_predict_ms: max alloted time for the prediction
-"   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
-"   auto_fim:         trigger FIM completion automatically on cursor movement
-"   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
-"
-" ring buffer of chunks, accumulated with time upon:
-"
-"  - completion request
-"  - yank
-"  - entering a buffer
-"  - leaving a buffer
-"  - writing a file
-"
-" parameters for the ring-buffer with extra context:
-"
-"   ring_n_chunks:    max number of chunks to pass as extra context to the server (0 to disable)
-"   ring_chunk_size:  max size of the chunks (in number of lines)
-"                     note: adjust these numbers so that you don't overrun your context
-"                           at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
-"   ring_scope:       the range around the cursor position (in number of lines) for gathering chunks after FIM
-"   ring_update_ms:   how often to process queued chunks in normal mode
-"
-let s:default_config = {
-    \ 'endpoint':         'http://127.0.0.1:8012/infill',
-    \ 'n_prefix':         256,
-    \ 'n_suffix':         64,
-    \ 'n_predict':        128,
-    \ 't_max_prompt_ms':  500,
-    \ 't_max_predict_ms': 3000,
-    \ 'show_info':        2,
-    \ 'auto_fim':         v:true,
-    \ 'max_line_suffix':  8,
-    \ 'ring_n_chunks':    64,
-    \ 'ring_chunk_size':  64,
-    \ 'ring_scope':       1024,
-    \ 'ring_update_ms':   1000,
-    \ }
-
-let g:llama_config = get(g:, 'llama_config', s:default_config)
-
-function! s:get_indent(str)
-    let l:count = 0
-    for i in range(len(a:str))
-        if a:str[i] == "\t"
-            let l:count += &tabstop - 1
-        else
-            break
-        endif
-    endfor
-    return l:count
-endfunction
-
-function! s:rand(i0, i1) abort
-    return a:i0 + rand() % (a:i1 - a:i0 + 1)
-endfunction
-
-function! llama#init()
-    if !executable('curl')
-        echohl WarningMsg
-        echo 'llama.vim requires the "curl" command to be available'
-        echohl None
-        return
-    endif
-
-    let s:pos_x = 0 " cursor position upon start of completion
-    let s:pos_y = 0
-
-    let s:line_cur = ''
-
-    let s:line_cur_prefix = ''
-    let s:line_cur_suffix = ''
-
-    let s:ring_chunks = [] " current set of chunks used as extra context
-    let s:ring_queued = [] " chunks that are queued to be sent for processing
-    let s:ring_n_evict = 0
-
-    let s:hint_shown = v:false
-    let s:pos_y_pick = -9999 " last y where we picked a chunk
-    let s:pos_dx = 0
-    let s:content = []
-    let s:can_accept = v:false
-
-    let s:timer_fim = -1
-    let s:t_fim_start = reltime() " used to measure total FIM time
-    let s:t_last_move = reltime() " last time the cursor moved
-
-    let s:current_job = v:null
-
-    let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
-    let s:ghost_text_vim = has('textprop')
-
-    if s:ghost_text_vim
-        let s:hlgroup_hint = 'llama_hl_hint'
-        let s:hlgroup_info = 'llama_hl_info'
-
-        if empty(prop_type_get(s:hlgroup_hint))
-            call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
-        endif
-        if empty(prop_type_get(s:hlgroup_info))
-            call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
-        endif
-    endif
-
-    augroup llama
-        autocmd!
-        autocmd InsertEnter     * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
-        autocmd InsertLeavePre  * call llama#fim_cancel()
-
-        autocmd CursorMoved     * call s:on_move()
-        autocmd CursorMovedI    * call s:on_move()
-        autocmd CompleteChanged * call llama#fim_cancel()
-
-        if g:llama_config.auto_fim
-            autocmd CursorMovedI * call llama#fim(v:true)
-        endif
-
-        " gather chunks upon yanking
-        autocmd TextYankPost    * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
-
-        " gather chunks upon entering/leaving a buffer
-        autocmd BufEnter        * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
-        autocmd BufLeave        * call                      s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
-
-        " gather chunk upon saving the file
-        autocmd BufWritePost    * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
-    augroup END
-
-    silent! call llama#fim_cancel()
-
-    " init background update of the ring buffer
-    if g:llama_config.ring_n_chunks > 0
-        call s:ring_update()
-    endif
-endfunction
-
-" compute how similar two chunks of text are
-" 0 - no similarity, 1 - high similarity
-" TODO: figure out something better
-function! s:chunk_sim(c0, c1)
-    let l:lines0 = len(a:c0)
-    let l:lines1 = len(a:c1)
-
-    let l:common = 0
-
-    for l:line0 in a:c0
-        for l:line1 in a:c1
-            if l:line0 == l:line1
-                let l:common += 1
-                break
-            endif
-        endfor
-    endfor
-
-    return 2.0 * l:common / (l:lines0 + l:lines1)
-endfunction
-
-" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
-"
-" no_mod   - do not pick chunks from buffers with pending changes
-" do_evict - evict chunks that are very similar to the new one
-"
-function! s:pick_chunk(text, no_mod, do_evict)
-    " do not pick chunks from buffers with pending changes or buffers that are not files
-    if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
-        return
-    endif
-
-    " if the extra context option is disabled - do nothing
-    if g:llama_config.ring_n_chunks <= 0
-        return
-    endif
-
-    " don't pick very small chunks
-    if len(a:text) < 3
-        return
-    endif
-
-    if len(a:text) + 1 < g:llama_config.ring_chunk_size
-        let l:chunk = a:text
-    else
-        let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
-        let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
-
-        let l:chunk = a:text[l:l0:l:l1]
-    endif
-
-    let l:chunk_str = join(l:chunk, "\n") . "\n"
-
-    " check if this chunk is already added
-    let l:exist = v:false
-
-    for i in range(len(s:ring_chunks))
-        if s:ring_chunks[i].data == l:chunk
-            let l:exist = v:true
-            break
-        endif
-    endfor
-
-    for i in range(len(s:ring_queued))
-        if s:ring_queued[i].data == l:chunk
-            let l:exist = v:true
-            break
-        endif
-    endfor
-
-    if l:exist
-        return
-    endif
-
-    " evict queued chunks that are very similar to the new one
-    for i in range(len(s:ring_queued) - 1, 0, -1)
-        if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
-            if a:do_evict
-                call remove(s:ring_queued, i)
-                let s:ring_n_evict += 1
-            else
-                return
-            endif
-        endif
-    endfor
-
-    " also from s:ring_chunks
-    for i in range(len(s:ring_chunks) - 1, 0, -1)
-        if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
-            if a:do_evict
-                call remove(s:ring_chunks, i)
-                let s:ring_n_evict += 1
-            else
-                return
-            endif
-        endif
-    endfor
-
-    " TODO: become parameter ?
-    if len(s:ring_queued) == 16
-        call remove(s:ring_queued, 0)
-    endif
-
-    call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
-
-    "let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
-endfunction
-
-" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
-" called every g:llama_config.ring_update_ms
-function! s:ring_update()
-    call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
-
-    " update only if in normal mode or if the cursor hasn't moved for a while
-    if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
-        return
-    endif
-
-    if len(s:ring_queued) == 0
-        return
-    endif
-
-    " move the first queued chunk to the ring buffer
-    if len(s:ring_chunks) == g:llama_config.ring_n_chunks
-        call remove(s:ring_chunks, 0)
-    endif
-
-    call add(s:ring_chunks, remove(s:ring_queued, 0))
-
-    "let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
-
-    " send asynchronous job with the new extra context so that it is ready for the next FIM
-    let l:extra_context = []
-    for l:chunk in s:ring_chunks
-        call add(l:extra_context, {
-            \ 'text':     l:chunk.str,
-            \ 'time':     l:chunk.time,
-            \ 'filename': l:chunk.filename
-            \ })
-    endfor
-
-    " no samplers needed here
-    let l:request = json_encode({
-        \ 'input_prefix':     "",
-        \ 'input_suffix':     "",
-        \ 'input_extra':      l:extra_context,
-        \ 'prompt':           "",
-        \ 'n_predict':        1,
-        \ 'temperature':      0.0,
-        \ 'stream':           v:false,
-        \ 'samplers':         ["temperature"],
-        \ 'cache_prompt':     v:true,
-        \ 't_max_prompt_ms':  1,
-        \ 't_max_predict_ms': 1
-        \ })
-
-    let l:curl_command = [
-        \ "curl",
-        \ "--silent",
-        \ "--no-buffer",
-        \ "--request", "POST",
-        \ "--url", g:llama_config.endpoint,
-        \ "--header", "Content-Type: application/json",
-        \ "--data", l:request
-        \ ]
-
-    " no callbacks because we don't need to process the response
-    if s:ghost_text_nvim
-        call jobstart(l:curl_command, {})
-    elseif s:ghost_text_vim
-        call job_start(l:curl_command, {})
-    endif
-endfunction
-
-" necessary for 'inoremap <expr>'
-function! llama#fim_inline(is_auto) abort
-    call llama#fim(a:is_auto)
-    return ''
-endfunction
-
-" the main FIM call
-" takes local context around the cursor and sends it together with the extra context to the server for completion
-function! llama#fim(is_auto) abort
-    " we already have a suggestion for the current cursor position
-    if s:hint_shown && !a:is_auto
-        call llama#fim_cancel()
-        return
-    endif
-
-    call llama#fim_cancel()
-
-    " avoid sending repeated requests too fast
-    if reltimefloat(reltime(s:t_fim_start)) < 0.6
-        if s:timer_fim != -1
-            call timer_stop(s:timer_fim)
-            let s:timer_fim = -1
-        endif
-
-        let s:t_fim_start = reltime()
-        let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
-        return
-    endif
-
-    let s:t_fim_start = reltime()
-
-    let s:content = []
-    let s:can_accept = v:false
-
-    let s:pos_x = col('.') - 1
-    let s:pos_y = line('.')
-    let l:max_y = line('$')
-
-    let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
-    let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
-
-    let s:line_cur = getline('.')
-
-    let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
-    let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
-
-    if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
-        return
-    endif
-
-    let l:prefix = ""
-        \ . join(l:lines_prefix, "\n")
-        \ . "\n"
-
-    let l:prompt = ""
-        \ . s:line_cur_prefix
-
-    let l:suffix = ""
-        \ . s:line_cur_suffix
-        \ . "\n"
-        \ . join(l:lines_suffix, "\n")
-        \ . "\n"
-
-    " prepare the extra context data
-    let l:extra_context = []
-    for l:chunk in s:ring_chunks
-        call add(l:extra_context, {
-            \ 'text':     l:chunk.str,
-            \ 'time':     l:chunk.time,
-            \ 'filename': l:chunk.filename
-            \ })
-    endfor
-
-    " the indentation of the current line
-    let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
-
-    let l:request = json_encode({
-        \ 'input_prefix':     l:prefix,
-        \ 'input_suffix':     l:suffix,
-        \ 'input_extra':      l:extra_context,
-        \ 'prompt':           l:prompt,
-        \ 'n_predict':        g:llama_config.n_predict,
-        \ 'n_indent':         l:indent,
-        \ 'top_k':            40,
-        \ 'top_p':            0.99,
-        \ 'stream':           v:false,
-        \ 'samplers':         ["top_k", "top_p", "infill"],
-        \ 'cache_prompt':     v:true,
-        \ 't_max_prompt_ms':  g:llama_config.t_max_prompt_ms,
-        \ 't_max_predict_ms': g:llama_config.t_max_predict_ms
-        \ })
-
-    let l:curl_command = [
-        \ "curl",
-        \ "--silent",
-        \ "--no-buffer",
-        \ "--request", "POST",
-        \ "--url", g:llama_config.endpoint,
-        \ "--header", "Content-Type: application/json",
-        \ "--data", l:request
-        \ ]
-
-    if s:current_job != v:null
-        if s:ghost_text_nvim
-            call jobstop(s:current_job)
-        elseif s:ghost_text_vim
-            call job_stop(s:current_job)
-        endif
-    endif
-
-    " send the request asynchronously
-    if s:ghost_text_nvim
-        let s:current_job = jobstart(l:curl_command, {
-            \ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
-            \ 'on_exit':   function('s:fim_on_exit'),
-            \ 'stdout_buffered': v:true
-            \ })
-    elseif s:ghost_text_vim
-        let s:current_job = job_start(l:curl_command, {
-            \ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
-            \ 'exit_cb':   function('s:fim_on_exit')
-            \ })
-    endif
-
-    " TODO: per-file location
-    let l:delta_y = abs(s:pos_y - s:pos_y_pick)
-
-    " gather some extra context nearby and process it in the background
-    " only gather chunks if the cursor has moved a lot
-    " TODO: something more clever? reranking?
-    if a:is_auto && l:delta_y > 32
-        " expand the prefix even further
-        call s:pick_chunk(getline(max([1,       s:pos_y - g:llama_config.ring_scope]), max([1,       s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
-
-        " pick a suffix chunk
-        call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]),   min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
-
-        let s:pos_y_pick = s:pos_y
-    endif
-endfunction
-
-" if first_line == v:true accept only the first line of the response
-function! llama#fim_accept(first_line)
-    " insert the suggestion at the cursor location
-    if s:can_accept && len(s:content) > 0
-        call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
-        if len(s:content) > 1
-            if !a:first_line
-                call append(s:pos_y, s:content[1:-1])
-            endif
-        endif
-
-        " move the cursor to the end of the accepted text
-        if !a:first_line && len(s:content) > 1
-            call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
-        else
-            call cursor(s:pos_y, s:pos_x + len(s:content[0]))
-        endif
-    endif
-
-    call llama#fim_cancel()
-endfunction
-
-function! llama#fim_cancel()
-    let s:hint_shown = v:false
-
-    " clear the virtual text
-    let l:bufnr = bufnr('%')
-
-    if s:ghost_text_nvim
-        let l:id_vt_fim = nvim_create_namespace('vt_fim')
-        call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim,  0, -1)
-    elseif s:ghost_text_vim
-        call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
-        call prop_remove({'type': s:hlgroup_info, 'all': v:true})
-    endif
-
-    " remove the mappings
-    silent! iunmap <buffer> <Tab>
-    silent! iunmap <buffer> <S-Tab>
-    silent! iunmap <buffer> <Esc>
-endfunction
-
-function! s:on_move()
-    let s:t_last_move = reltime()
-
-    call llama#fim_cancel()
-endfunction
-
-" callback that processes the FIM result from the server and displays the suggestion
-function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
-    if s:ghost_text_nvim
-        let l:raw = join(a:data, "\n")
-    elseif s:ghost_text_vim
-        let l:raw = a:data
-    endif
-
-    if len(l:raw) == 0
-        return
-    endif
-
-    if a:pos_x != col('.') - 1 || a:pos_y != line('.')
-        return
-    endif
-
-    " show the suggestion only in insert mode
-    if mode() !=# 'i'
-        return
-    endif
-
-    let s:pos_x = a:pos_x
-    let s:pos_y = a:pos_y
-
-    let s:can_accept = v:true
-    let l:has_info   = v:false
-
-    if s:can_accept && v:shell_error
-        if !a:is_auto
-            call add(s:content, "<| curl error: is the server on? |>")
-        endif
-        let s:can_accept = v:false
-    endif
-
-    let l:n_prompt    = 0
-    let l:t_prompt_ms = 1.0
-    let l:s_prompt    = 0
-
-    let l:n_predict    = 0
-    let l:t_predict_ms = 1.0
-    let l:s_predict    = 0
-
-    " get the generated suggestion
-    if s:can_accept
-        let l:response = json_decode(l:raw)
-
-        for l:part in split(get(l:response, 'content', ''), "\n", 1)
-            call add(s:content, l:part)
-        endfor
-
-        " remove trailing new lines
-        while len(s:content) > 0 && s:content[-1] == ""
-            call remove(s:content, -1)
-        endwhile
-
-        let l:generation_settings = get(l:response, 'generation_settings', {})
-        let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
-
-        let l:n_cached  = get(l:response, 'tokens_cached', 0)
-        let l:truncated = get(l:response, 'truncated', v:false)
-
-        " if response.timings is available
-        if len(get(l:response, 'timings', {})) > 0
-            let l:has_info = v:true
-            let l:timings  = get(l:response, 'timings', {})
-
-            let l:n_prompt    = get(l:timings, 'prompt_n', 0)
-            let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
-            let l:s_prompt    = get(l:timings, 'prompt_per_second', 0)
-
-            let l:n_predict    = get(l:timings, 'predicted_n', 0)
-            let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
-            let l:s_predict    = get(l:timings, 'predicted_per_second', 0)
-        endif
-    endif
-
-    if len(s:content) == 0
-        call add(s:content, "")
-        let s:can_accept = v:false
-    endif
-
-    if len(s:content) == 0
-        return
-    endif
-
-    " NOTE: the following is logic for discarding predictions that repeat existing text
-    "       the code is quite ugly and there is very likely a simpler and more canonical way to implement this
-    "
-    "       still, I wonder if there is some better way that avoids having to do these special hacks?
-    "       on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
-    "       start generating whatever we have given it via the extra context. but on the other hand, it's not very
-    "       helpful to re-generate the same code that is already there
-
-    " truncate the suggestion if the first line is empty
-    if len(s:content) == 1 && s:content[0] == ""
-        let s:content = [""]
-    endif
-
-    " ... and the next lines are repeated
-    if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
-        let s:content = [""]
-    endif
-
-    " truncate the suggestion if it repeats the suffix
-    if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
-        let s:content = [""]
-    endif
-
-    " find the first non-empty line (strip whitespace)
-    let l:cmp_y = s:pos_y + 1
-    while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
-        let l:cmp_y += 1
-    endwhile
-
-    if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
-        " truncate the suggestion if it repeats the next line
-        if len(s:content) == 1
-            let s:content = [""]
-        endif
-
-        " ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
-        if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
-            let s:content = [""]
-        endif
-
-        " ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
-        if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
-            let s:content = [""]
-        endif
-    endif
-
-    " keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
-    "let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
-    "for i in range(1, len(s:content) - 1)
-    "    if strlen(matchstr(s:content[i], '^\s*')) < l:indent
-    "        let s:content = s:content[:i - 1]
-    "        break
-    "    endif
-    "endfor
-
-    let s:pos_dx = len(s:content[-1])
-
-    let s:content[-1] .= s:line_cur_suffix
-
-    call llama#fim_cancel()
-
-    " display virtual text with the suggestion
-    let l:bufnr = bufnr('%')
-
-    if s:ghost_text_nvim
-        let l:id_vt_fim = nvim_create_namespace('vt_fim')
-    endif
-
-    " construct the info message
-    if g:llama_config.show_info > 0 && l:has_info
-        let l:prefix = '   '
-
-        if l:truncated
-            let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
-                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
-                \ l:n_cached, l:n_ctx
-                \ )
-        else
-            let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
-                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
-                \ l:n_cached,  l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
-                \ l:n_prompt,  l:t_prompt_ms,  l:s_prompt,
-                \ l:n_predict, l:t_predict_ms, l:s_predict,
-                \ 1000.0 * reltimefloat(reltime(s:t_fim_start))
-                \ )
-        endif
-
-        if g:llama_config.show_info == 1
-            " display the info in the statusline
-            let &statusline = l:info
-            let l:info = ''
-        endif
-    endif
-
-    " display the suggestion and append the info to the end of the first line
-    if s:ghost_text_nvim
-        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
-            \ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
-            \ 'virt_text_win_col': virtcol('.') - 1
-            \ })
-
-        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
-            \ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
-            \ 'virt_text_win_col': virtcol('.')
-            \ })
-    elseif s:ghost_text_vim
-        let l:new_suffix = s:content[0]
-        if !empty(l:new_suffix)
-            call prop_add(s:pos_y, s:pos_x + 1, {
-                        \ 'type': s:hlgroup_hint,
-                        \ 'text': l:new_suffix
-                        \ })
-        endif
-        for line in s:content[1:]
-            call prop_add(s:pos_y, 0, {
-                        \ 'type': s:hlgroup_hint,
-                        \ 'text': line,
-                        \ 'text_padding_left': s:get_indent(line),
-                        \ 'text_align': 'below'
-                        \ })
-        endfor
-        if !empty(l:info)
-            call prop_add(s:pos_y, 0, {
-                        \ 'type': s:hlgroup_info,
-                        \ 'text': l:info,
-                        \ 'text_padding_left': col('$'),
-                        \ 'text_wrap': 'truncate'
-                        \ })
-        endif
-    endif
-
-    " setup accept shortcuts
-    inoremap <buffer> <Tab>   <C-O>:call llama#fim_accept(v:false)<CR>
-    inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
-
-    let s:hint_shown = v:true
-endfunction
-
-function! s:fim_on_exit(job_id, exit_code, event = v:null)
-    if a:exit_code != 0
-        echom "Job failed with exit code: " . a:exit_code
-    endif
-
-    let s:current_job = v:null
-endfunction

examples/llava/clip.cpp

@@ -4,7 +4,6 @@
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
 #include "ggml.h"
-#include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
@@ -40,17 +39,10 @@
 #include <cinttypes>
 #include <limits>
 
-#if defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...)
-#   define LOG_WRN(...)
-#   define LOG_ERR(...)
-#   define LOG_DBG(...)
-#else // defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#endif // defined(LLAVA_LOG_OFF)
+#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_DBG(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
 
 //#define CLIP_DEBUG_FUNCTIONS
 

examples/llava/llava-cli.cpp

@@ -20,7 +20,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
@@ -191,7 +191,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     LOG("\n");
 
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
     if (!smpl) {
         LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);

examples/llava/llava.cpp

@@ -11,17 +11,13 @@
 #include <limits>
 #include <vector>
 
-#if defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...)
-#   define LOG_WRN(...)
-#   define LOG_ERR(...)
-#   define LOG_DBG(...)
-#else // defined(LLAVA_LOG_OFF)
-#   define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#   define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
-#   define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
-#endif // defined(LLAVA_LOG_OFF)
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
 
 // RGB uint8 image
 struct clip_image_u8 {
@@ -405,39 +401,6 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     return true;
 }
 
-struct llava_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-};
-
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
     int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
 
@@ -446,9 +409,8 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        float * embd = image_embed->embed+i*n_embd;
-        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
-        if (llama_decode(ctx_llama, llava_batch.batch)) {
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        if (llama_decode(ctx_llama, batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
@@ -470,7 +432,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        LOG_ERR("%s: couldn't embed the image\n", __func__);
+        LOG_ERR("%s: coulnd't embed the image\n", __func__);
         return NULL;
     }
 
@@ -502,16 +464,10 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     errno = 0;
     size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
     if (ferror(file)) {
-        LOG_ERR("read error: %s", strerror(errno));
-        free(buffer);
-        fclose(file);
-        return false;
+        die_fmt("read error: %s", strerror(errno));
     }
     if (ret != (size_t) fileSize) {
-        LOG_ERR("unexpectedly reached end of file");
-        free(buffer);
-        fclose(file);
-        return false;
+        die("unexpectedly reached end of file");
     }
     fclose(file); // Close the file
 

examples/llava/minicpmv-cli.cpp

@@ -97,7 +97,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
@@ -237,7 +237,7 @@ static struct common_sampler * llama_init(struct llava_context * ctx_llava, comm
 
     LOG_INF("\n");
 
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
     return smpl;
 }
 

examples/lookahead/lookahead.cpp

@@ -89,8 +89,8 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
 
     for (int s = 1; s < W + G + 1; ++s) {
         llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
@@ -115,7 +115,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
 
     // target model sampling context
-    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
+    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
 
     // verification n-grams
     std::vector<ngram_data> ngrams_cur(G);

examples/lookup/lookup-stats.cpp

@@ -21,7 +21,7 @@ int main(int argc, char ** argv){
 
     common_init();
 
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.n_draft;
 
     // init llama.cpp
     llama_backend_init();
@@ -40,7 +40,6 @@ int main(int argc, char ** argv){
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
     common_ngram_cache ngram_cache_static;
-
     int64_t t_draft_flat_us = 0;
     int64_t t_draft_us = 0;
 

examples/lookup/lookup.cpp

@@ -22,7 +22,7 @@ int main(int argc, char ** argv){
     common_init();
 
     // max. number of additional tokens to draft if match is found
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.n_draft;
 
     const bool dump_kv_cache = params.dump_kv_cache;
 
@@ -89,8 +89,8 @@ int main(int argc, char ** argv){
 
     const auto t_enc_start = ggml_time_us();
 
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
 
     const auto t_enc_end = ggml_time_us();
 
@@ -102,7 +102,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
+    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
 
     std::vector<llama_token> draft;
 

examples/main/README.md

@@ -187,30 +187,6 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying
 
 Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
-### DRY Repetition Penalty
-
-DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
-
-- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
-- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
-- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
-- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
-- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
-
-The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
-
-The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
-
-The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
-
-The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
-
-The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
-
-DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
-
-Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
-
 ### Top-K Sampling
 
 -   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
@@ -235,6 +211,14 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e
 
 Example usage: `--min-p 0.05`
 
+### Tail-Free Sampling (TFS)
+
+-   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
+
+Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
+
+Example usage: `--tfs 0.95`
+
 ### Locally Typical Sampling
 
 -   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
@@ -257,19 +241,6 @@ The `--mirostat-ent` option sets the Mirostat target entropy (tau), which repres
 
 Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
 
-### XTC Sampling
-
--   `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
--   `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
-
-Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
-
-By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
-
-Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
-
-Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
-
 ### Logit Bias
 
 -   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
@@ -313,6 +284,10 @@ These options help improve the performance and memory usage of the LLaMA models.
 
  These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
+### Memory Float 32
+
+-   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
+
 ### Batch Size
 
 -   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
@@ -333,15 +308,6 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
-## LoRA (Low-Rank Adaptation) adapters
-
--   `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
--   `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
-
-You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
-
-LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
-
 ## Additional Options
 
 These options provide extra functionality and customization when running the LLaMA models:
@@ -350,4 +316,6 @@ These options provide extra functionality and customization when running the LLa
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/main/main.cpp

@@ -62,6 +62,49 @@ static bool file_is_empty(const std::string & path) {
     return f.tellg() == 0;
 }
 
+static void write_logfile(
+    const llama_context * ctx, const common_params & params, const llama_model * model,
+    const std::vector<llama_token> & input_tokens, const std::string & output,
+    const std::vector<llama_token> & output_tokens
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    const std::string timestamp = string_get_sortable_timestamp();
+
+    const bool success = fs_create_directory_with_parents(params.logdir);
+    if (!success) {
+        LOG_ERR("%s: failed to create logdir %s, cannot write logfile\n", __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: main\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Generation Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
+
+    llama_perf_dump_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
@@ -72,6 +115,7 @@ static void sigint_handler(int signo) {
             console::cleanup();
             LOG("\n");
             common_perf_print(*g_ctx, *g_smpl);
+            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
 
             // make sure all logs are flushed
             LOG("Interrupted by user\n");
@@ -100,7 +144,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sampling;
+    auto & sparams = params.sparams;
 
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
@@ -165,10 +209,6 @@ int main(int argc, char ** argv) {
 
     LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
 
-    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
-    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
-    auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
-
     struct ggml_threadpool_params tpp_batch =
             ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
     struct ggml_threadpool_params tpp =
@@ -178,7 +218,7 @@ int main(int argc, char ** argv) {
 
     struct ggml_threadpool * threadpool_batch = NULL;
     if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
-        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
+        threadpool_batch = ggml_threadpool_new(&tpp_batch);
         if (!threadpool_batch) {
             LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
             return 1;
@@ -188,7 +228,7 @@ int main(int argc, char ** argv) {
         tpp.paused = true;
     }
 
-    struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
+    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
     if (!threadpool) {
         LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
         return 1;
@@ -488,7 +528,7 @@ int main(int argc, char ** argv) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
 
-        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
+        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
@@ -529,30 +569,30 @@ int main(int argc, char ** argv) {
                     if (!params.ctx_shift){
                         LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
                         break;
+                    } else {
+                        if (params.n_predict == -2) {
+                            LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                            break;
+                        }
+
+                        const int n_left    = n_past - params.n_keep;
+                        const int n_discard = n_left/2;
+
+                        LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                                n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                        llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                        llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+
+                        n_past -= n_discard;
+
+                        LOG_DBG("after swap: n_past = %d\n", n_past);
+
+                        LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+
+                        LOG_DBG("clear session path\n");
+                        path_session.clear();
                     }
-
-                    if (params.n_predict == -2) {
-                        LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                        break;
-                    }
-
-                    const int n_left    = n_past - params.n_keep;
-                    const int n_discard = n_left/2;
-
-                    LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                            n_past, n_left, n_ctx, params.n_keep, n_discard);
-
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
-
-                    n_past -= n_discard;
-
-                    LOG_DBG("after swap: n_past = %d\n", n_past);
-
-                    LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
-
-                    LOG_DBG("clear session path\n");
-                    path_session.clear();
                 }
             } else {
                 // context extension via Self-Extend
@@ -608,7 +648,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -886,6 +926,7 @@ int main(int argc, char ** argv) {
 
     LOG("\n\n");
     common_perf_print(ctx, smpl);
+    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     common_sampler_free(smpl);
 
@@ -894,8 +935,8 @@ int main(int argc, char ** argv) {
 
     llama_backend_free();
 
-    ggml_threadpool_free_fn(threadpool);
-    ggml_threadpool_free_fn(threadpool_batch);
+    ggml_threadpool_free(threadpool);
+    ggml_threadpool_free(threadpool_batch);
 
     return 0;
 }

examples/parallel/parallel.cpp

@@ -160,7 +160,7 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.smpl = common_sampler_init(model, params.sampling);
+        client.smpl = common_sampler_init(model, params.sparams);
     }
 
     std::vector<llama_token> tokens_system;
@@ -308,6 +308,7 @@ 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);

examples/perplexity/perplexity.cpp

@@ -34,6 +34,55 @@ struct results_log_softmax {
     float  prob;
 };
 
+static void write_logfile(
+    const llama_context * ctx, const common_params & params, const llama_model * model,
+    const struct results_perplexity & results
+) {
+    if (params.logdir.empty()) {
+        return;
+    }
+
+    if (params.hellaswag) {
+        LOG_WRN("%s: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
+        return;
+    }
+
+    const std::string timestamp = string_get_sortable_timestamp();
+
+    const bool success = fs_create_directory_with_parents(params.logdir);
+    if (!success) {
+        LOG_WRN("%s: failed to create logdir %s, cannot write logfile\n",
+                __func__, params.logdir.c_str());
+        return;
+    }
+
+    const std::string logfile_path = params.logdir + timestamp + ".yml";
+    FILE * logfile = fopen(logfile_path.c_str(), "w");
+
+    if (logfile == NULL) {
+        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        return;
+    }
+
+    fprintf(logfile, "binary: main\n");
+    char model_desc[128];
+    llama_model_desc(model, model_desc, sizeof(model_desc));
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
+
+    fprintf(logfile, "\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "# Perplexity Results #\n");
+    fprintf(logfile, "######################\n");
+    fprintf(logfile, "\n");
+
+    yaml_dump_vector_float(logfile, "logits", results.logits);
+    fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
+    yaml_dump_vector_float(logfile, "probs", results.probs);
+
+    llama_perf_dump_yaml(logfile, ctx);
+    fclose(logfile);
+}
+
 static std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
@@ -359,21 +408,14 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
-
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
-            common_batch_clear(batch);
-            for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
-            }
-
             //LOG_DBG("    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            if (llama_decode(ctx, batch)) {
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 //LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
                 return {tokens, -1, logit_history, prob_history};
             }
 
@@ -393,8 +435,6 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
             }
         }
 
-        llama_batch_free(batch);
-
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -664,6 +704,7 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             batch.n_seq_id + i,
             batch.seq_id   + i,
             batch.logits   + i,
+            0, 0, 0, // unused
         };
 
         const int ret = llama_decode(ctx, batch_view);
@@ -1750,8 +1791,6 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
-        llama_batch batch = llama_batch_init(n_batch, 0, 1);
-
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -1764,14 +1803,9 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            common_batch_clear(batch);
-            for (int i = 0; i < batch_size; i++) {
-                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
-            }
-
-            if (llama_decode(ctx, batch)) {
+            // TODO: use llama_batch.logits instead of relying on logits_all == true
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 LOG_ERR("%s : failed to eval\n", __func__);
-                llama_batch_free(batch);
                 return;
             }
 
@@ -1784,8 +1818,6 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
             }
         }
 
-        llama_batch_free(batch);
-
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -2023,6 +2055,8 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
+    write_logfile(ctx, params, model, results);
+
     llama_free(ctx);
     llama_free_model(model);
 

examples/quantize-stats/quantize-stats.cpp

@@ -142,7 +142,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
 }
 
 static void test_roundtrip_on_chunk(
-    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, bool use_reference,
     float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
 ) {
     if (layer->type == GGML_TYPE_F16) {
@@ -156,7 +156,7 @@ static void test_roundtrip_on_chunk(
     if (use_reference) {
         qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
     } else {
-        qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);
+        qfns.from_float(input_scratch, quantized_scratch, chunk_size);
     }
     qfns.to_float(quantized_scratch, output_scratch, chunk_size);
 
@@ -166,7 +166,7 @@ static void test_roundtrip_on_chunk(
 
 // Run quantization function for a single layer and update error stats
 static void test_roundtrip_on_layer(
-    std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
+    std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, bool use_reference,
     const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
     std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
 ) {
@@ -187,13 +187,13 @@ static void test_roundtrip_on_layer(
     int num_chunks = (nelements + chunk_size - 1)/chunk_size;
 
     if (num_chunks < 2 || max_thread < 2) {
-        test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, use_reference, input_scratch_ptr, quantized_scratch.data(),
+        test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
                 output_scratch.data(), print_layer_stats ? layer_error : total_error);
     } else {
         auto & stats = print_layer_stats ? layer_error : total_error;
         std::mutex mutex;
         uint64_t counter = 0;
-        auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,
+        auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
              &quantized_scratch, &output_scratch, chunk_size] () {
             error_stats local_stats {};
             while (true) {
@@ -205,7 +205,7 @@ static void test_roundtrip_on_layer(
                 }
                 lock.unlock();
                 uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
-                test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,
+                test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
                         quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
             }
         };
@@ -371,9 +371,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        const auto *  qfns     = ggml_get_type_traits(type);
-        const auto *  qfns_cpu = ggml_get_type_traits_cpu(type);
-        if (qfns_cpu->from_float && qfns->to_float) {
+        const auto *  qfns = ggml_get_type_traits(type);
+        if (qfns->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }
@@ -394,7 +393,7 @@ int main(int argc, char ** argv) {
                 test_roundtrip_on_layer(
                         layer_name,
                         params.per_layer_stats,
-                        *qfns, *qfns_cpu,
+                        *qfns,
                         params.reference,
                         kv_tensor.second,
                         input_scratch,

examples/retrieval/retrieval.cpp

@@ -282,8 +282,8 @@ int main(int argc, char ** argv) {
                 return a.second > b.second;
             });
 
-            LOG("Top %d similar chunks:\n", params.sampling.top_k);
-            for (int i = 0; i < std::min(params.sampling.top_k, (int) chunks.size()); i++) {
+            LOG("Top %d similar chunks:\n", params.sparams.top_k);
+            for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
                 LOG("filename: %s\n", chunks[similarities[i].first].filename.c_str());
                 LOG("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
                 LOG("similarity: %f\n", similarities[i].second);

examples/rpc/rpc-server.cpp

@@ -1,5 +1,3 @@
-#include "ggml-cpu.h"
-
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif

examples/run/CMakeLists.txt

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

examples/run/README.md

@@ -1,7 +0,0 @@
-# llama.cpp/example/run
-
-The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
-
-```bash
-./llama-run Meta-Llama-3.1-8B-Instruct.gguf
-...

examples/run/run.cpp

@@ -1,409 +0,0 @@
-#if defined(_WIN32)
-#include <windows.h>
-#else
-#include <unistd.h>
-#endif
-
-#include <climits>
-#include <cstdio>
-#include <cstring>
-#include <iostream>
-#include <sstream>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-#include "llama-cpp.h"
-
-typedef std::unique_ptr<char[]> char_array_ptr;
-
-struct Argument {
-    std::string flag;
-    std::string help_text;
-};
-
-struct Options {
-    std::string model_path, prompt_non_interactive;
-    int ngl = 99;
-    int n_ctx = 2048;
-};
-
-class ArgumentParser {
-   public:
-    ArgumentParser(const char * program_name) : program_name(program_name) {}
-
-    void add_argument(const std::string & flag, std::string & var, const std::string & help_text = "") {
-        string_args[flag] = &var;
-        arguments.push_back({flag, help_text});
-    }
-
-    void add_argument(const std::string & flag, int & var, const std::string & help_text = "") {
-        int_args[flag] = &var;
-        arguments.push_back({flag, help_text});
-    }
-
-    int parse(int argc, const char ** argv) {
-        for (int i = 1; i < argc; ++i) {
-            std::string arg = argv[i];
-            if (string_args.count(arg)) {
-                if (i + 1 < argc) {
-                    *string_args[arg] = argv[++i];
-                } else {
-                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
-                    print_usage();
-                    return 1;
-                }
-            } else if (int_args.count(arg)) {
-                if (i + 1 < argc) {
-                    if (parse_int_arg(argv[++i], *int_args[arg]) != 0) {
-                        fprintf(stderr, "error: invalid value for %s: %s\n", arg.c_str(), argv[i]);
-                        print_usage();
-                        return 1;
-                    }
-                } else {
-                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
-                    print_usage();
-                    return 1;
-                }
-            } else {
-                fprintf(stderr, "error: unrecognized argument %s\n", arg.c_str());
-                print_usage();
-                return 1;
-            }
-        }
-
-        if (string_args["-m"]->empty()) {
-            fprintf(stderr, "error: -m is required\n");
-            print_usage();
-            return 1;
-        }
-
-        return 0;
-    }
-
-   private:
-    const char * program_name;
-    std::unordered_map<std::string, std::string *> string_args;
-    std::unordered_map<std::string, int *> int_args;
-    std::vector<Argument> arguments;
-
-    int parse_int_arg(const char * arg, int & value) {
-        char * end;
-        const long val = std::strtol(arg, &end, 10);
-        if (*end == '\0' && val >= INT_MIN && val <= INT_MAX) {
-            value = static_cast<int>(val);
-            return 0;
-        }
-        return 1;
-    }
-
-    void print_usage() const {
-        printf("\nUsage:\n");
-        printf("  %s [OPTIONS]\n\n", program_name);
-        printf("Options:\n");
-        for (const auto & arg : arguments) {
-            printf("  %-10s %s\n", arg.flag.c_str(), arg.help_text.c_str());
-        }
-
-        printf("\n");
-    }
-};
-
-class LlamaData {
-   public:
-    llama_model_ptr model;
-    llama_sampler_ptr sampler;
-    llama_context_ptr context;
-    std::vector<llama_chat_message> messages;
-
-    int init(const Options & opt) {
-        model = initialize_model(opt.model_path, opt.ngl);
-        if (!model) {
-            return 1;
-        }
-
-        context = initialize_context(model, opt.n_ctx);
-        if (!context) {
-            return 1;
-        }
-
-        sampler = initialize_sampler();
-        return 0;
-    }
-
-   private:
-    // Initializes the model and returns a unique pointer to it
-    llama_model_ptr initialize_model(const std::string & model_path, const int ngl) {
-        llama_model_params model_params = llama_model_default_params();
-        model_params.n_gpu_layers = ngl;
-
-        llama_model_ptr model(llama_load_model_from_file(model_path.c_str(), model_params));
-        if (!model) {
-            fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        }
-
-        return model;
-    }
-
-    // Initializes the context with the specified parameters
-    llama_context_ptr initialize_context(const llama_model_ptr & model, const int n_ctx) {
-        llama_context_params ctx_params = llama_context_default_params();
-        ctx_params.n_ctx = n_ctx;
-        ctx_params.n_batch = n_ctx;
-
-        llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params));
-        if (!context) {
-            fprintf(stderr, "%s: error: failed to create the llama_context\n", __func__);
-        }
-
-        return context;
-    }
-
-    // Initializes and configures the sampler
-    llama_sampler_ptr initialize_sampler() {
-        llama_sampler_ptr sampler(llama_sampler_chain_init(llama_sampler_chain_default_params()));
-        llama_sampler_chain_add(sampler.get(), llama_sampler_init_min_p(0.05f, 1));
-        llama_sampler_chain_add(sampler.get(), llama_sampler_init_temp(0.8f));
-        llama_sampler_chain_add(sampler.get(), llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
-
-        return sampler;
-    }
-};
-
-// Add a message to `messages` and store its content in `owned_content`
-static void add_message(const char * role, const std::string & text, LlamaData & llama_data,
-                        std::vector<char_array_ptr> & owned_content) {
-    char_array_ptr content(new char[text.size() + 1]);
-    std::strcpy(content.get(), text.c_str());
-    llama_data.messages.push_back({role, content.get()});
-    owned_content.push_back(std::move(content));
-}
-
-// Function to apply the chat template and resize `formatted` if needed
-static int apply_chat_template(const LlamaData & llama_data, std::vector<char> & formatted, const bool append) {
-    int result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
-                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
-    if (result > static_cast<int>(formatted.size())) {
-        formatted.resize(result);
-        result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
-                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
-    }
-
-    return result;
-}
-
-// Function to tokenize the prompt
-static int tokenize_prompt(const llama_model_ptr & model, const std::string & prompt,
-                           std::vector<llama_token> & prompt_tokens) {
-    const int n_prompt_tokens = -llama_tokenize(model.get(), prompt.c_str(), prompt.size(), NULL, 0, true, true);
-    prompt_tokens.resize(n_prompt_tokens);
-    if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
-                       true) < 0) {
-        GGML_ABORT("failed to tokenize the prompt\n");
-    }
-
-    return n_prompt_tokens;
-}
-
-// Check if we have enough space in the context to evaluate this batch
-static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
-    const int n_ctx = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
-    if (n_ctx_used + batch.n_tokens > n_ctx) {
-        printf("\033[0m\n");
-        fprintf(stderr, "context size exceeded\n");
-        return 1;
-    }
-
-    return 0;
-}
-
-// convert the token to a string
-static int convert_token_to_string(const llama_model_ptr & model, const llama_token token_id, std::string & piece) {
-    char buf[256];
-    int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true);
-    if (n < 0) {
-        GGML_ABORT("failed to convert token to piece\n");
-    }
-
-    piece = std::string(buf, n);
-    return 0;
-}
-
-static void print_word_and_concatenate_to_response(const std::string & piece, std::string & response) {
-    printf("%s", piece.c_str());
-    fflush(stdout);
-    response += piece;
-}
-
-// helper function to evaluate a prompt and generate a response
-static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) {
-    std::vector<llama_token> prompt_tokens;
-    const int n_prompt_tokens = tokenize_prompt(llama_data.model, prompt, prompt_tokens);
-    if (n_prompt_tokens < 0) {
-        return 1;
-    }
-
-    // prepare a batch for the prompt
-    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
-    llama_token new_token_id;
-    while (true) {
-        check_context_size(llama_data.context, batch);
-        if (llama_decode(llama_data.context.get(), batch)) {
-            GGML_ABORT("failed to decode\n");
-        }
-
-        // sample the next token, check is it an end of generation?
-        new_token_id = llama_sampler_sample(llama_data.sampler.get(), llama_data.context.get(), -1);
-        if (llama_token_is_eog(llama_data.model.get(), new_token_id)) {
-            break;
-        }
-
-        std::string piece;
-        if (convert_token_to_string(llama_data.model, new_token_id, piece)) {
-            return 1;
-        }
-
-        print_word_and_concatenate_to_response(piece, response);
-
-        // prepare the next batch with the sampled token
-        batch = llama_batch_get_one(&new_token_id, 1);
-    }
-
-    return 0;
-}
-
-static int parse_arguments(const int argc, const char ** argv, Options & opt) {
-    ArgumentParser parser(argv[0]);
-    parser.add_argument("-m", opt.model_path, "model");
-    parser.add_argument("-p", opt.prompt_non_interactive, "prompt");
-    parser.add_argument("-c", opt.n_ctx, "context_size");
-    parser.add_argument("-ngl", opt.ngl, "n_gpu_layers");
-    if (parser.parse(argc, argv)) {
-        return 1;
-    }
-
-    return 0;
-}
-
-static int read_user_input(std::string & user) {
-    std::getline(std::cin, user);
-    return user.empty();  // Indicate an error or empty input
-}
-
-// Function to generate a response based on the prompt
-static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response) {
-    // Set response color
-    printf("\033[33m");
-    if (generate(llama_data, prompt, response)) {
-        fprintf(stderr, "failed to generate response\n");
-        return 1;
-    }
-
-    // End response with color reset and newline
-    printf("\n\033[0m");
-    return 0;
-}
-
-// Helper function to apply the chat template and handle errors
-static int apply_chat_template_with_error_handling(const LlamaData & llama_data, std::vector<char> & formatted,
-                                                   const bool is_user_input, int & output_length) {
-    const int new_len = apply_chat_template(llama_data, formatted, is_user_input);
-    if (new_len < 0) {
-        fprintf(stderr, "failed to apply the chat template\n");
-        return -1;
-    }
-
-    output_length = new_len;
-    return 0;
-}
-
-// Helper function to handle user input
-static bool handle_user_input(std::string & user_input, const std::string & prompt_non_interactive) {
-    if (!prompt_non_interactive.empty()) {
-        user_input = prompt_non_interactive;
-        return true;  // No need for interactive input
-    }
-
-    printf("\033[32m> \033[0m");
-    return !read_user_input(user_input);  // Returns false if input ends the loop
-}
-
-// Function to tokenize the prompt
-static int chat_loop(LlamaData & llama_data, std::string & prompt_non_interactive) {
-    std::vector<char_array_ptr> owned_content;
-    std::vector<char> fmtted(llama_n_ctx(llama_data.context.get()));
-    int prev_len = 0;
-
-    while (true) {
-        // Get user input
-        std::string user_input;
-        if (!handle_user_input(user_input, prompt_non_interactive)) {
-            break;
-        }
-
-        add_message("user", prompt_non_interactive.empty() ? user_input : prompt_non_interactive, llama_data,
-                    owned_content);
-
-        int new_len;
-        if (apply_chat_template_with_error_handling(llama_data, fmtted, true, new_len) < 0) {
-            return 1;
-        }
-
-        std::string prompt(fmtted.begin() + prev_len, fmtted.begin() + new_len);
-        std::string response;
-        if (generate_response(llama_data, prompt, response)) {
-            return 1;
-        }
-    }
-    return 0;
-}
-
-static void log_callback(const enum ggml_log_level level, const char * text, void *) {
-    if (level == GGML_LOG_LEVEL_ERROR) {
-        fprintf(stderr, "%s", text);
-    }
-}
-
-static bool is_stdin_a_terminal() {
-#if defined(_WIN32)
-    HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE);
-    DWORD mode;
-    return GetConsoleMode(hStdin, &mode);
-#else
-    return isatty(STDIN_FILENO);
-#endif
-}
-
-static std::string read_pipe_data() {
-    std::ostringstream result;
-    result << std::cin.rdbuf();  // Read all data from std::cin
-    return result.str();
-}
-
-int main(int argc, const char ** argv) {
-    Options opt;
-    if (parse_arguments(argc, argv, opt)) {
-        return 1;
-    }
-
-    if (!is_stdin_a_terminal()) {
-        if (!opt.prompt_non_interactive.empty()) {
-            opt.prompt_non_interactive += "\n\n";
-        }
-
-        opt.prompt_non_interactive += read_pipe_data();
-    }
-
-    llama_log_set(log_callback, nullptr);
-    LlamaData llama_data;
-    if (llama_data.init(opt)) {
-        return 1;
-    }
-
-    if (chat_loop(llama_data, opt.prompt_non_interactive)) {
-        return 1;
-    }
-
-    return 0;
-}