server : enable cache_prompt by default (#10501)

ggml-ci

.clang-format

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

.devops/full-cuda.Dockerfile

@@ -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_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc) && \
     cp build/bin/* .
 

.devops/full-musa.Dockerfile

@@ -19,7 +19,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc) && \
     cp build/bin/* .
 

.devops/llama-cli-cann.Dockerfile

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

.devops/llama-cli-cuda.Dockerfile

@@ -22,16 +22,17 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    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)
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1
 
-COPY --from=build /app/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
+COPY --from=build /app/lib/ /
+COPY --from=build /app/build/bin/llama-cli /
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-intel.Dockerfile

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

.devops/llama-cli-musa.Dockerfile

@@ -15,16 +15,17 @@ WORKDIR /app
 
 COPY . .
 
-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)
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1
 
-COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
-COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/lib/ /
 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_VULKAN=1 && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 && \
     cmake --build build --config Release --target llama-cli
 
 # Clean up

.devops/llama-server-cuda.Dockerfile

@@ -22,16 +22,17 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    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)
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
-COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/lib/ /
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 # Must be set to 0.0.0.0 so it can listen to requests from host machine

.devops/llama-server-intel.Dockerfile

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

.devops/llama-server-musa.Dockerfile

@@ -15,16 +15,17 @@ WORKDIR /app
 
 COPY . .
 
-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)
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-server -j$(nproc) && \
+    mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
 
 FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
-COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/lib/ /
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 # Must be set to 0.0.0.0 so it can listen to requests from host machine

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

.devops/nix/package.nix

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

.editorconfig

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

.github/ISSUE_TEMPLATE/01-bug-low.yml

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

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

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

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

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

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

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

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

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

.github/ISSUE_TEMPLATE/020-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/03-bug-high.yml

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

.github/ISSUE_TEMPLATE/030-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/04-bug-critical.yml

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

.github/ISSUE_TEMPLATE/040-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/workflows/build.yml

@@ -55,7 +55,13 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake .. \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -113,7 +119,12 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -394,15 +405,36 @@ jobs:
       - name: Build with native CMake HIP support
         id: cmake_build
         run: |
-          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
       - name: Build with legacy HIP support
         id: cmake_build_legacy_hip
         run: |
-          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIP=ON
           cmake --build build2 --config Release -j $(nproc)
 
+  ubuntu-22-cmake-musa:
+    runs-on: ubuntu-22.04
+    container: mthreads/musa:rc3.1.0-devel-ubuntu22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          apt-get update
+          apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+      - name: Build with native CMake MUSA support
+        id: cmake_build
+        run: |
+          cmake -B build -S . -DGGML_MUSA=ON
+          cmake --build build --config Release -j $(nproc)
+
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -569,6 +601,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -599,6 +632,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -734,7 +768,7 @@ jobs:
         id: clone_kompute
         if: ${{ matrix.build == 'kompute-x64' }}
         run: |
-          git submodule update --init ggml/src/kompute
+          git submodule update --init ggml/src/ggml-kompute/kompute
 
       - name: Download OpenBLAS
         id: get_openblas
@@ -917,8 +951,8 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
@@ -928,7 +962,8 @@ jobs:
           fetch-depth: 0
 
       - name: Install
-        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+        run:  |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
       - name: Build
         id: cmake_build
@@ -947,26 +982,34 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
-      - name: Pack artifacts
+      - name: Build the release package
         id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
+
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+      - name: Upload the release package
+        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
@@ -1001,7 +1044,7 @@ 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_HIPBLAS=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_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
 
   windows-latest-cmake-hip-release:
@@ -1037,7 +1080,7 @@ 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_HIPBLAS=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="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
           cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"

.gitignore

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

CMakeLists.txt

@@ -46,6 +46,13 @@ if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 endif()
 
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/execution-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/execution-charset:utf-8>")
+endif()
+
 #
 # option list
 #
@@ -140,7 +147,6 @@ 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

CMakePresets.json

@@ -24,11 +24,12 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
-    { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
-    { "name": "sycl_f16",  "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
+    { "name": "debug",    "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
+    { "name": "reldbg",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",   "hidden": true, "cacheVariables": { "GGML_STATIC":      "ON" } },
+    { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
+    { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
@@ -57,25 +58,28 @@
         }
     },
 
-    { "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-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-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" ] }
   ]
 }

Makefile

@@ -48,7 +48,6 @@ 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 \
@@ -252,7 +251,7 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon
+MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -DGGML_USE_CPU
 MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
@@ -291,6 +290,7 @@ 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,
@@ -359,6 +359,10 @@ 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 \
@@ -523,70 +527,59 @@ 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
-		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
-		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
-		MK_LDFLAGS  += -framework Accelerate
-		OBJ_GGML    += ggml/src/ggml-blas.o
+		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
 	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    += ggml/src/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_EXT += ggml/src/ggml-blas/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    += ggml/src/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_EXT += ggml/src/ggml-blas/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    += ggml/src/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_EXT += ggml/src/ggml-blas/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    += ggml/src/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_EXT += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_NVPL
 
 ifndef GGML_NO_LLAMAFILE
-	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJ_GGML    += ggml/src/llamafile/sgemm.o
+	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    += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
+	OBJ_GGML_EXT += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
 endif
 
 ifdef GGML_RPC
-	MK_CPPFLAGS += -DGGML_USE_RPC
-	OBJ_GGML    += ggml/src/ggml-rpc.o
+	MK_CPPFLAGS  += -DGGML_USE_RPC
+	OBJ_GGML_EXT += ggml/src/ggml-rpc.o
 endif # GGML_RPC
 
 OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@@ -601,41 +594,27 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_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$(CUDA_PATH)/lib -L/usr/lib64
-		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
+	ifneq ('', '$(wildcard /opt/cuda)')
+		CUDA_PATH ?= /opt/cuda
 	else
-		ifneq ('', '$(wildcard /opt/cuda)')
-			CUDA_PATH ?= /opt/cuda
-		else
-			CUDA_PATH ?= /usr/local/cuda
-		endif
+		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$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-		MK_NVCCFLAGS += -use_fast_math
-	endif # GGML_MUSA
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(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
 
-	OBJ_GGML += ggml/src/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
+	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)
 
 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
@@ -648,11 +627,7 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	ifdef GGML_MUSA
-		NVCC = $(CCACHE) mcc
-	else
-		NVCC = $(CCACHE) nvcc
-	endif # GGML_MUSA
+	NVCC = $(CCACHE) nvcc
 endif # GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
@@ -661,10 +636,6 @@ 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
@@ -673,20 +644,6 @@ 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
@@ -695,12 +652,6 @@ 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
@@ -724,15 +675,9 @@ 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: \
@@ -742,8 +687,8 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-cuda/common.cuh
 	$(NVCC_COMPILE)
 
-ggml/src/ggml-cuda.o: \
-	ggml/src/ggml-cuda.cu \
+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 \
@@ -754,9 +699,9 @@ ggml/src/ggml-cuda.o: \
 endif # GGML_CUDA
 
 ifdef GGML_VULKAN
-	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
+	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
 
 ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
@@ -786,10 +731,10 @@ 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/vulkan-shaders
+_ggml_vk_input_dir = ggml/src/ggml-vulkan/vulkan-shaders
 _ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
 
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
+ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan/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)
@@ -801,8 +746,8 @@ $(_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/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+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
 
 endif # GGML_VULKAN
 
@@ -815,11 +760,7 @@ ifdef GGML_HIPBLAS
 		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 
-	GGML_CUDA_DMMV_X       ?= 32
-	GGML_CUDA_MMV_Y        ?= 1
-	GGML_CUDA_KQUANTS_ITER ?= 2
-
-	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
+	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
 ifdef GGML_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
@@ -832,13 +773,6 @@ 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
@@ -852,12 +786,12 @@ ifdef GGML_CUDA_NO_PEER_COPY
 	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY
 
-	OBJ_GGML += ggml/src/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
+	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.o: \
-	ggml/src/ggml-cuda.cu \
+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 \
@@ -874,70 +808,168 @@ ggml/src/ggml-cuda/%.o: \
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 endif # GGML_HIPBLAS
 
+ifdef GGML_MUSA
+	ifeq ($(wildcard /opt/musa),)
+		MUSA_PATH ?= /usr/local/musa
+	else
+		MUSA_PATH ?= /opt/musa
+	endif
+	MTGPU_TARGETS ?= mp_21 mp_22
+
+	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
+	MK_LDFLAGS += -L$(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 += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
+
+ifdef GGML_CUDA_FORCE_MMQ
+	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # GGML_CUDA_FORCE_MMQ
+
+ifdef GGML_CUDA_FORCE_CUBLAS
+	MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
+endif # GGML_CUDA_FORCE_CUBLAS
+
+ifdef GGML_CUDA_F16
+	MUSAFLAGS += -DGGML_CUDA_F16
+endif # GGML_CUDA_F16
+
+ifdef GGML_CUDA_DMMV_F16
+	MUSAFLAGS += -DGGML_CUDA_F16
+endif # GGML_CUDA_DMMV_F16
+
+ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(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) -x musa -mtgpu -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) -x musa -mtgpu -c -o $@ $<
+endif # GGML_MUSA
+
 ifdef GGML_METAL
-	MK_CPPFLAGS += -DGGML_USE_METAL
-	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML	+= ggml/src/ggml-metal.o
+	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
 ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
 ifdef GGML_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML   += ggml/src/ggml-metal-embed.o
+	MK_CPPFLAGS  += -DGGML_METAL_EMBED_LIBRARY
+	OBJ_GGML_EXT += ggml/src/ggml-metal-embed.o
 endif
 endif # GGML_METAL
 
 ifdef GGML_METAL
-ggml/src/ggml-metal.o: \
-	ggml/src/ggml-metal.m \
+ggml/src/ggml-metal/ggml-metal.o: \
+	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	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.metal \
+	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@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
+	@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
 	$(eval TEMP_ASSEMBLY=$(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-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/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
 
-OBJ_GGML += \
-	ggml/src/ggml.o \
-	ggml/src/ggml-cpu.o \
-	ggml/src/ggml-alloc.o \
-	ggml/src/ggml-backend.o \
-	ggml/src/ggml-quants.o \
-	ggml/src/ggml-aarch64.o
+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_LLAMA = \
-	src/llama.o \
-	src/llama-vocab.o \
-	src/llama-grammar.o \
-	src/llama-sampling.o \
-	src/unicode.o \
-	src/unicode-data.o
+	$(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
 
 OBJ_COMMON = \
-	common/common.o \
-	common/arg.o \
-	common/log.o \
-	common/console.o \
-	common/ngram-cache.o \
-	common/sampling.o \
-	common/build-info.o \
-	common/json-schema-to-grammar.o
+	$(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
 
 OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
 
@@ -993,7 +1025,6 @@ $(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
@@ -1003,7 +1034,6 @@ 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 )
 
@@ -1040,224 +1070,78 @@ endif
 # Build libraries
 #
 
-# ggml
+# Libraries
+LIB_GGML   = libggml.so
+LIB_GGML_S = libggml.a
 
-ggml/src/ggml.o: \
-	ggml/src/ggml.c \
-	ggml/include/ggml.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+LIB_LLAMA   = libllama.so
+LIB_LLAMA_S = libllama.a
 
-ggml/src/ggml-cpu.o: \
-	ggml/src/ggml-cpu.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+LIB_COMMON   = libcommon.so
+LIB_COMMON_S = libcommon.a
 
-ggml/src/ggml-alloc.o: \
-	ggml/src/ggml-alloc.c \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+# Targets
+BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
 
-ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.cpp \
-	ggml/src/ggml-backend-impl.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Dependency files
+DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
 
-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 $@
+# Default target
+all: $(BUILD_TARGETS)
 
-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 $@
-
-ggml/src/ggml-blas.o: \
-	ggml/src/ggml-blas.cpp \
-	ggml/include/ggml-blas.h
-	$(CXX) $(CXXFLAGS) -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
-
-ifndef GGML_NO_AMX
-ggml/src/ggml-amx.o: \
-	ggml/src/ggml-amx.cpp \
-	ggml/include/ggml-amx.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-ggml/src/ggml-amx/mmq.o: \
-	ggml/src/ggml-amx/mmq.cpp \
-	ggml/src/ggml-amx/mmq.h \
-	ggml/include/ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif
-
-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)
-	ar rcs $(LIB_GGML_S) $^
-
-# 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 $@
-
-src/llama.o: \
-	src/llama.cpp \
-	src/llama-impl.h \
-	src/llama-vocab.h \
-	src/llama-grammar.h \
-	src/llama-sampling.h \
-	src/unicode.h \
-	include/llama.h \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml-metal.h \
+# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
+#       g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
+$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
+	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/include/ggml-backend.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-alloc.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-cpu.h \
+	ggml/src/ggml-impl.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 $@
+# Rules for building object files
+$(DIR_GGML)/%.o: $(DIR_GGML)/%.c
+	$(CC) $(CFLAGS) -MMD -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 $@
+$(DIR_GGML)/%.o: $(DIR_GGML)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -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 $@
+$(DIR_LLAMA)/%.o: $(DIR_LLAMA)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
 
-$(LIB_LLAMA): \
-	$(OBJ_LLAMA) \
-	$(LIB_GGML)
+$(DIR_COMMON)/%.o: $(DIR_COMMON)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
+
+# Rules for building libraries
+$(LIB_GGML): $(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_LLAMA_S): \
-	$(OBJ_LLAMA)
+$(LIB_GGML_S): $(OBJ_GGML)
+	ar rcs $(LIB_GGML_S) $^
+
+$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_LLAMA_S): $(OBJ_LLAMA)
 	ar rcs $(LIB_LLAMA_S) $^
 
-# 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/ngram-cache.o: \
-	common/ngram-cache.cpp \
-	common/ngram-cache.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_COMMON): \
-	$(OBJ_COMMON) \
-	$(LIB_LLAMA) \
-	$(LIB_GGML)
+$(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 *.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 -vrf ggml/src/ggml-amx/*.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
+	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
 
 #
 # Examples
@@ -1455,22 +1339,13 @@ 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)
@@ -1572,11 +1447,6 @@ 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

@@ -10,11 +10,16 @@ var sources = [
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
-    "ggml/src/ggml-cpu.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] = []
@@ -22,6 +27,7 @@ var linkerSettings: [LinkerSetting] = []
 var cSettings: [CSetting] =  [
     .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
     .unsafeFlags(["-fno-objc-arc"]),
+    .headerSearchPath("ggml/src"),
     // NOTE: NEW_LAPACK will required iOS version 16.4+
     // We should consider add this in the future when we drop support for iOS 14
     // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
@@ -30,13 +36,15 @@ var cSettings: [CSetting] =  [
 ]
 
 #if canImport(Darwin)
-sources.append("ggml/src/ggml-metal.m")
-resources.append(.process("ggml/src/ggml-metal.metal"))
+sources.append("ggml/src/ggml-common.h")
+sources.append("ggml/src/ggml-metal/ggml-metal.m")
+resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
 linkerSettings.append(.linkedFramework("Accelerate"))
 cSettings.append(
     contentsOf: [
         .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL")
+        .define("GGML_USE_METAL"),
+        .define("GGML_USE_CPU")
     ]
 )
 #endif
@@ -61,13 +69,15 @@ let package = Package(
             name: "llama",
             path: ".",
             exclude: [
+               "build",
                "cmake",
                "examples",
                "scripts",
                "models",
                "tests",
                "CMakeLists.txt",
-               "Makefile"
+               "Makefile",
+               "ggml/src/ggml-metal-embed.metal"
             ],
             sources: sources,
             resources: resources,

README.md

@@ -131,6 +131,7 @@ Typically finetunes of the base models below are supported as well.
 - 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)
@@ -458,14 +459,14 @@ 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 documentations
+## Other documentation
 
 - [main (cli)](./examples/main/README.md)
 - [server](./examples/server/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [GBNF grammars](./grammars/README.md)
 
-**Development documentations**
+**Development documentation**
 
 - [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"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then

cmake/llama-config.cmake.in

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

common/CMakeLists.txt

@@ -66,6 +66,8 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     sampling.cpp
     sampling.h
+    speculative.cpp
+    speculative.h
     )
 
 if (BUILD_SHARED_LIBS)

common/arg.cpp

@@ -233,10 +233,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         }
     }
 
-    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams,       nullptr);
     postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
-    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
-    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
+    postprocess_cpu_params(params.speculative.cpuparams,       &params.cpuparams);
+    postprocess_cpu_params(params.speculative.cpuparams_batch, &params.cpuparams_batch);
 
     if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
@@ -251,7 +252,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & antiprompt : params.antiprompt) {
             string_process_escapes(antiprompt);
         }
-        for (auto & seq_breaker : params.sparams.dry_sequence_breakers) {
+        for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
             string_process_escapes(seq_breaker);
         }
     }
@@ -297,6 +298,27 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
     print_options(specific_options);
 }
 
+static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & value) {
+    std::vector<ggml_backend_dev_t> devices;
+    auto dev_names = string_split<std::string>(value, ',');
+    if (dev_names.empty()) {
+        throw std::invalid_argument("no devices specified");
+    }
+    if (dev_names.size() == 1 && dev_names[0] == "none") {
+        devices.push_back(nullptr);
+    } else {
+        for (const auto & device : dev_names) {
+            auto * dev = ggml_backend_dev_by_name(device.c_str());
+            if (!dev || ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+                throw std::invalid_argument(string_format("invalid device: %s", device.c_str()));
+            }
+            devices.push_back(dev);
+        }
+        devices.push_back(nullptr);
+    }
+    return devices;
+}
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
@@ -323,13 +345,16 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // load dynamic backends
+    ggml_backend_load_all();
+
     common_params_context ctx_arg(params);
     ctx_arg.print_usage = print_usage;
     ctx_arg.ex          = ex;
 
     std::string sampler_type_chars;
     std::string sampler_type_names;
-    for (const auto & sampler : params.sparams.samplers) {
+    for (const auto & sampler : params.sampling.samplers) {
         sampler_type_chars += common_sampler_type_to_chr(sampler);
         sampler_type_names += common_sampler_type_to_str(sampler) + ";";
     }
@@ -407,26 +432,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
-    add_opt(common_arg(
-        {"-td", "--threads-draft"}, "N",
-        "number of threads to use during generation (default: same as --threads)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.n_threads = value;
-            if (params.draft_cpuparams.n_threads <= 0) {
-                params.draft_cpuparams.n_threads = std::thread::hardware_concurrency();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-tbd", "--threads-batch-draft"}, "N",
-        "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.n_threads = value;
-            if (params.draft_cpuparams_batch.n_threads <= 0) {
-                params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-C", "--cpu-mask"}, "M",
         "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")",
@@ -515,108 +520,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.cpuparams_batch.poll = value;
         }
     ));
-    add_opt(common_arg(
-        {"-Cd", "--cpu-mask-draft"}, "M",
-        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
-        [](common_params & params, const std::string & mask) {
-            params.draft_cpuparams.mask_valid = true;
-            if (!parse_cpu_mask(mask, params.draft_cpuparams.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Crd", "--cpu-range-draft"}, "lo-hi",
-        "Ranges of CPUs for affinity. Complements --cpu-mask-draft",
-        [](common_params & params, const std::string & range) {
-            params.draft_cpuparams.mask_valid = true;
-            if (!parse_cpu_range(range, params.draft_cpuparams.cpumask)) {
-                throw std::invalid_argument("invalid range");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--cpu-strict-draft"}, "<0|1>",
-        "Use strict CPU placement for draft model (default: same as --cpu-strict)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.strict_cpu = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--prio-draft"}, "N",
-        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams.priority),
-        [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.draft_cpuparams.priority = (enum ggml_sched_priority) prio;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--poll-draft"}, "<0|1>",
-        "Use polling to wait for draft model work (default: same as --poll])",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.poll = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Cbd", "--cpu-mask-batch-draft"}, "M",
-        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
-        [](common_params & params, const std::string & mask) {
-            params.draft_cpuparams_batch.mask_valid = true;
-            if (!parse_cpu_mask(mask, params.draft_cpuparams_batch.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi",
-        "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)",
-        [](common_params & params, const std::string & range) {
-            params.draft_cpuparams_batch.mask_valid = true;
-            if (!parse_cpu_range(range, params.draft_cpuparams_batch.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--cpu-strict-batch-draft"}, "<0|1>",
-        "Use strict CPU placement for draft model (default: --cpu-strict-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.strict_cpu = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--prio-batch-draft"}, "N",
-        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams_batch.priority),
-        [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) prio;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--poll-batch-draft"}, "<0|1>",
-        "Use polling to wait for draft model work (default: --poll-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.poll = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--draft"}, "N",
-        string_format("number of tokens to draft for speculative decoding (default: %d)", params.n_draft),
-        [](common_params & params, int value) {
-            params.n_draft = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
-    add_opt(common_arg(
-        {"-ps", "--p-split"}, "N",
-        string_format("speculative decoding split probability (default: %.1f)", (double)params.p_split),
-        [](common_params & params, const std::string & value) {
-            params.p_split = std::stof(value);
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-lcs", "--lookup-cache-static"}, "FNAME",
         "path to static lookup cache to use for lookup decoding (not updated by generation)",
@@ -701,7 +604,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
         [](common_params & params) {
             params.no_perf = true;
-            params.sparams.no_perf = true;
+            params.sampling.no_perf = true;
         }
     ).set_env("LLAMA_ARG_NO_PERF"));
     add_opt(common_arg(
@@ -883,155 +786,155 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
         [](common_params & params, const std::string & value) {
             const auto sampler_names = string_split<std::string>(value, ';');
-            params.sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"-s", "--seed"}, "SEED",
-        string_format("RNG seed (default: %d, use random seed for %d)", params.sparams.seed, LLAMA_DEFAULT_SEED),
+        string_format("RNG seed (default: %d, use random seed for %d)", params.sampling.seed, LLAMA_DEFAULT_SEED),
         [](common_params & params, const std::string & value) {
-            params.sparams.seed = std::stoul(value);
+            params.sampling.seed = std::stoul(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--sampling-seq"}, "SEQUENCE",
         string_format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()),
         [](common_params & params, const std::string & value) {
-            params.sparams.samplers = common_sampler_types_from_chars(value);
+            params.sampling.samplers = common_sampler_types_from_chars(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--ignore-eos"},
         "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)",
         [](common_params & params) {
-            params.sparams.ignore_eos = true;
+            params.sampling.ignore_eos = true;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--penalize-nl"},
-        string_format("penalize newline tokens (default: %s)", params.sparams.penalize_nl ? "true" : "false"),
+        string_format("penalize newline tokens (default: %s)", params.sampling.penalize_nl ? "true" : "false"),
         [](common_params & params) {
-            params.sparams.penalize_nl = true;
+            params.sampling.penalize_nl = true;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--temp"}, "N",
-        string_format("temperature (default: %.1f)", (double)params.sparams.temp),
+        string_format("temperature (default: %.1f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
-            params.sparams.temp = std::stof(value);
-            params.sparams.temp = std::max(params.sparams.temp, 0.0f);
+            params.sampling.temp = std::stof(value);
+            params.sampling.temp = std::max(params.sampling.temp, 0.0f);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--top-k"}, "N",
-        string_format("top-k sampling (default: %d, 0 = disabled)", params.sparams.top_k),
+        string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
         [](common_params & params, int value) {
-            params.sparams.top_k = value;
+            params.sampling.top_k = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--top-p"}, "N",
-        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sparams.top_p),
+        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.top_p = std::stof(value);
+            params.sampling.top_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--min-p"}, "N",
-        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sparams.min_p),
+        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.min_p = std::stof(value);
+            params.sampling.min_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
-        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_probability),
+        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
-            params.sparams.xtc_probability = std::stof(value);
+            params.sampling.xtc_probability = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-threshold"}, "N",
-        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sparams.xtc_threshold),
+        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
-            params.sparams.xtc_threshold = std::stof(value);
+            params.sampling.xtc_threshold = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--typical"}, "N",
-        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sparams.typ_p),
+        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.typ_p = std::stof(value);
+            params.sampling.typ_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-last-n"}, "N",
-        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sparams.penalty_last_n),
+        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
         [](common_params & params, int value) {
-            params.sparams.penalty_last_n = value;
-            params.sparams.n_prev = std::max(params.sparams.n_prev, params.sparams.penalty_last_n);
+            params.sampling.penalty_last_n = value;
+            params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-penalty"}, "N",
-        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sparams.penalty_repeat),
+        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_repeat = std::stof(value);
+            params.sampling.penalty_repeat = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--presence-penalty"}, "N",
-        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_present),
+        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_present = std::stof(value);
+            params.sampling.penalty_present = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
-        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_freq),
+        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_freq = std::stof(value);
+            params.sampling.penalty_freq = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-multiplier"}, "N",
-        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sparams.dry_multiplier),
+        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
         [](common_params & params, const std::string & value) {
-            params.sparams.dry_multiplier = std::stof(value);
+            params.sampling.dry_multiplier = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-base"}, "N",
-        string_format("set DRY sampling base value (default: %.2f)", (double)params.sparams.dry_base),
+        string_format("set DRY sampling base value (default: %.2f)", (double)params.sampling.dry_base),
         [](common_params & params, const std::string & value) {
             float potential_base = std::stof(value);
             if (potential_base >= 1.0f)
             {
-                params.sparams.dry_base = potential_base;
+                params.sampling.dry_base = potential_base;
             }
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-allowed-length"}, "N",
-        string_format("set allowed length for DRY sampling (default: %d)", params.sparams.dry_allowed_length),
+        string_format("set allowed length for DRY sampling (default: %d)", params.sampling.dry_allowed_length),
         [](common_params & params, int value) {
-            params.sparams.dry_allowed_length = value;
+            params.sampling.dry_allowed_length = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-penalty-last-n"}, "N",
-        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sparams.dry_penalty_last_n),
+        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
         [](common_params & params, int value) {
-            params.sparams.dry_penalty_last_n = value;
+            params.sampling.dry_penalty_last_n = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-sequence-breaker"}, "STRING",
         string_format("add sequence breaker for DRY sampling, clearing out default breakers (%s) in the process; use \"none\" to not use any sequence breakers\n",
-            params.sparams.dry_sequence_breakers.empty() ? "none" :
-            std::accumulate(std::next(params.sparams.dry_sequence_breakers.begin()),
-                params.sparams.dry_sequence_breakers.end(),
-                std::string("'") + (params.sparams.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sparams.dry_sequence_breakers[0]) + "'",
+            params.sampling.dry_sequence_breakers.empty() ? "none" :
+            std::accumulate(std::next(params.sampling.dry_sequence_breakers.begin()),
+                params.sampling.dry_sequence_breakers.end(),
+                std::string("'") + (params.sampling.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sampling.dry_sequence_breakers[0]) + "'",
                 [](const std::string& a, const std::string& b) {
                     std::string formatted_b = (b == "\n") ? "\\n" : b;
                     return a + ", '" + formatted_b + "'";
@@ -1040,51 +943,51 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             static bool defaults_cleared = false;
 
             if (!defaults_cleared) {
-                params.sparams.dry_sequence_breakers.clear();
+                params.sampling.dry_sequence_breakers.clear();
                 defaults_cleared = true;
             }
 
             if (value == "none") {
-                params.sparams.dry_sequence_breakers.clear();
+                params.sampling.dry_sequence_breakers.clear();
             } else {
-                params.sparams.dry_sequence_breakers.emplace_back(value);
+                params.sampling.dry_sequence_breakers.emplace_back(value);
             }
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
-        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range),
+        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
         [](common_params & params, const std::string & value) {
-            params.sparams.dynatemp_range = std::stof(value);
+            params.sampling.dynatemp_range = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-exp"}, "N",
-        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sparams.dynatemp_exponent),
+        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sampling.dynatemp_exponent),
         [](common_params & params, const std::string & value) {
-            params.sparams.dynatemp_exponent = std::stof(value);
+            params.sampling.dynatemp_exponent = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat"}, "N",
         string_format("use Mirostat sampling.\nTop K, Nucleus and Locally Typical samplers are ignored if used.\n"
-        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat),
+        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat),
         [](common_params & params, int value) {
-            params.sparams.mirostat = value;
+            params.sampling.mirostat = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-lr"}, "N",
-        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sparams.mirostat_eta),
+        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
-            params.sparams.mirostat_eta = std::stof(value);
+            params.sampling.mirostat_eta = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-ent"}, "N",
-        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sparams.mirostat_tau),
+        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
-            params.sparams.mirostat_tau = std::stof(value);
+            params.sampling.mirostat_tau = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1100,7 +1003,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             try {
                 if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
                     const float bias = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
-                    params.sparams.logit_bias.push_back({key, bias});
+                    params.sampling.logit_bias.push_back({key, bias});
                 } else {
                     throw std::invalid_argument("invalid input format");
                 }
@@ -1111,9 +1014,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--grammar"}, "GRAMMAR",
-        string_format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sparams.grammar.c_str()),
+        string_format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sampling.grammar.c_str()),
         [](common_params & params, const std::string & value) {
-            params.sparams.grammar = value;
+            params.sampling.grammar = value;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1127,7 +1030,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             std::copy(
                 std::istreambuf_iterator<char>(file),
                 std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sparams.grammar)
+                std::back_inserter(params.sampling.grammar)
             );
         }
     ).set_sparam());
@@ -1135,7 +1038,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-j", "--json-schema"}, "SCHEMA",
         "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
         [](common_params & params, const std::string & value) {
-            params.sparams.grammar = json_schema_to_grammar(json::parse(value));
+            params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1433,6 +1336,30 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_NUMA"));
+    add_opt(common_arg(
+        {"-dev", "--device"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.devices = parse_device_list(value);
+        }
+    ).set_env("LLAMA_ARG_DEVICE"));
+    add_opt(common_arg(
+        {"--list-devices"},
+        "print list of available devices and exit",
+        [](common_params &) {
+            printf("Available devices:\n");
+            for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                auto * dev = ggml_backend_dev_get(i);
+                if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+                    size_t free, total;
+                    ggml_backend_dev_memory(dev, &free, &total);
+                    printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
+                }
+            }
+            exit(0);
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1444,17 +1371,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_N_GPU_LAYERS"));
-    add_opt(common_arg(
-        {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
-        "number of layers to store in VRAM for the draft model",
-        [](common_params & params, int value) {
-            params.n_gpu_layers_draft = value;
-            if (!llama_supports_gpu_offload()) {
-                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
-                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-sm", "--split-mode"}, "{none,layer,row}",
         "how to split the model across multiple GPUs, one of:\n"
@@ -1468,10 +1384,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             } else if (arg_next == "layer") {
                 params.split_mode = LLAMA_SPLIT_MODE_LAYER;
             } else if (arg_next == "row") {
-#ifdef GGML_USE_SYCL
-                fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
-                exit(1);
-#endif // GGML_USE_SYCL
                 params.split_mode = LLAMA_SPLIT_MODE_ROW;
             } else {
                 throw std::invalid_argument("invalid value");
@@ -1593,13 +1505,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model = value;
         }
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
-    add_opt(common_arg(
-        {"-md", "--model-draft"}, "FNAME",
-        "draft model for speculative decoding (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.model_draft = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
@@ -1939,17 +1844,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.simple_io = true;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
-    add_opt(common_arg(
-        {"-ld", "--logdir"}, "LOGDIR",
-        "path under which to save YAML logs (no logging if unset)",
-        [](common_params & params, const std::string & value) {
-            params.logdir = value;
-
-            if (params.logdir.back() != DIRECTORY_SEPARATOR) {
-                params.logdir += DIRECTORY_SEPARATOR;
-            }
-        }
-    ));
     add_opt(common_arg(
         {"--positive-file"}, "FNAME",
         string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),
@@ -2048,5 +1942,176 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_LOG_TIMESTAMPS"));
 
+    // speculative parameters
+    add_opt(common_arg(
+        {"-td", "--threads-draft"}, "N",
+        "number of threads to use during generation (default: same as --threads)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.n_threads = value;
+            if (params.speculative.cpuparams.n_threads <= 0) {
+                params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-tbd", "--threads-batch-draft"}, "N",
+        "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.n_threads = value;
+            if (params.speculative.cpuparams_batch.n_threads <= 0) {
+                params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Cd", "--cpu-mask-draft"}, "M",
+        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
+        [](common_params & params, const std::string & mask) {
+            params.speculative.cpuparams.mask_valid = true;
+            if (!parse_cpu_mask(mask, params.speculative.cpuparams.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Crd", "--cpu-range-draft"}, "lo-hi",
+        "Ranges of CPUs for affinity. Complements --cpu-mask-draft",
+        [](common_params & params, const std::string & range) {
+            params.speculative.cpuparams.mask_valid = true;
+            if (!parse_cpu_range(range, params.speculative.cpuparams.cpumask)) {
+                throw std::invalid_argument("invalid range");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--cpu-strict-draft"}, "<0|1>",
+        "Use strict CPU placement for draft model (default: same as --cpu-strict)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.strict_cpu = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--prio-draft"}, "N",
+        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.speculative.cpuparams.priority),
+        [](common_params & params, int prio) {
+            if (prio < 0 || prio > 3) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.speculative.cpuparams.priority = (enum ggml_sched_priority) prio;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--poll-draft"}, "<0|1>",
+        "Use polling to wait for draft model work (default: same as --poll])",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.poll = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Cbd", "--cpu-mask-batch-draft"}, "M",
+        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
+        [](common_params & params, const std::string & mask) {
+            params.speculative.cpuparams_batch.mask_valid = true;
+            if (!parse_cpu_mask(mask, params.speculative.cpuparams_batch.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi",
+        "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)",
+        [](common_params & params, const std::string & range) {
+            params.speculative.cpuparams_batch.mask_valid = true;
+            if (!parse_cpu_range(range, params.speculative.cpuparams_batch.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--cpu-strict-batch-draft"}, "<0|1>",
+        "Use strict CPU placement for draft model (default: --cpu-strict-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.strict_cpu = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--prio-batch-draft"}, "N",
+        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.speculative.cpuparams_batch.priority),
+        [](common_params & params, int prio) {
+            if (prio < 0 || prio > 3) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.speculative.cpuparams_batch.priority = (enum ggml_sched_priority) prio;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--poll-batch-draft"}, "<0|1>",
+        "Use polling to wait for draft model work (default: --poll-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.poll = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--draft-max", "--draft", "--draft-n"}, "N",
+        string_format("number of tokens to draft for speculative decoding (default: %d)", params.speculative.n_max),
+        [](common_params & params, int value) {
+            params.speculative.n_max = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--draft-min", "--draft-n-min"}, "N",
+        string_format("minimum number of draft tokens to use for speculative decoding (default: %d)", params.speculative.n_min),
+        [](common_params & params, int value) {
+            params.speculative.n_min = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--draft-p-split"}, "P",
+        string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
+        [](common_params & params, const std::string & value) {
+            params.speculative.p_split = std::stof(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--draft-p-min"}, "P",
+        string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
+        [](common_params & params, const std::string & value) {
+            params.speculative.p_min = std::stof(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-cd", "--ctx-size-draft"}, "N",
+        string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx),
+        [](common_params & params, int value) {
+            params.speculative.n_ctx = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-devd", "--device-draft"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.speculative.devices = parse_device_list(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
+        "number of layers to store in VRAM for the draft model",
+        [](common_params & params, int value) {
+            params.speculative.n_gpu_layers = value;
+            if (!llama_supports_gpu_offload()) {
+                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
+                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-md", "--model-draft"}, "FNAME",
+        "draft model for speculative decoding (default: unused)",
+        [](common_params & params, const std::string & value) {
+            params.speculative.model = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+
     return ctx_arg;
 }

common/common.cpp

@@ -536,12 +536,12 @@ 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 << " ]";
@@ -875,6 +875,12 @@ 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);
@@ -919,9 +925,9 @@ struct common_init_result common_init_from_params(common_params & params) {
         common_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+    if (params.sampling.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.sparams.ignore_eos = false;
+        params.sampling.ignore_eos = false;
     }
 
     if (params.warmup) {
@@ -973,9 +979,12 @@ void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_l
     }
 }
 
-struct llama_model_params common_model_params_to_llama(const common_params & params) {
+struct llama_model_params common_model_params_to_llama(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;
     }
@@ -1484,6 +1493,66 @@ 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
 //
@@ -1890,218 +1959,3 @@ 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) {
-    ggml_cpu_init(); // some ARM features are detected at runtime
-
-    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, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
-    fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
-    fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
-    fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
-    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, "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, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
-    fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
-    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

@@ -33,6 +33,8 @@ 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;
@@ -101,8 +103,8 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
-// sampler parameters
-struct common_sampler_params {
+// sampling parameters
+struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
     int32_t n_prev             = 64;    // number of previous tokens to remember
@@ -153,21 +155,30 @@ struct common_sampler_params {
     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_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)
@@ -178,27 +189,31 @@ 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          = -1.0f; // KV cache defragmentation threshold
+    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
+
+    // offload params
+    std::vector<ggml_backend_dev_t> devices;         // devices to use for offloading
+    int32_t n_gpu_layers                    =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu                        =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]               =   {0}; // how split tensors should be distributed across GPUs
+    enum llama_split_mode        split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
-    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_sampler_params sparams;
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
 
     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
@@ -209,7 +224,6 @@ 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
@@ -452,7 +466,7 @@ struct common_init_result {
 
 struct common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model_params     common_model_params_to_llama  (const common_params & params);
+struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
@@ -462,7 +476,9 @@ struct llama_model * common_load_model_from_hf(const char * repo, const char * f
 // 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);
 
@@ -473,6 +489,16 @@ 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
 //
@@ -584,15 +610,3 @@ 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/sampling.cpp

@@ -99,7 +99,7 @@ struct ring_buffer {
 };
 
 struct common_sampler {
-    common_sampler_params params;
+    common_params_sampling params;
 
     struct llama_sampler * grmr;
     struct llama_sampler * chain;
@@ -125,7 +125,7 @@ struct common_sampler {
     }
 };
 
-std::string common_sampler_params::print() const {
+std::string common_params_sampling::print() const {
     char result[1024];
 
     snprintf(result, sizeof(result),
@@ -141,7 +141,7 @@ std::string common_sampler_params::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
@@ -320,6 +320,45 @@ 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);
 }

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_sampler_params & params);
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
@@ -60,6 +60,27 @@ 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

@@ -0,0 +1,269 @@
+#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__);
+        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

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

convert_hf_to_gguf.py

@@ -2707,7 +2707,7 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_token_scores(scores)
         self.gguf_writer.add_token_types(toktypes)
         self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
         self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
         if precompiled_charsmap:
             self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
@@ -3040,6 +3040,11 @@ 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

docs/backend/SYCL.md

@@ -34,13 +34,16 @@ The SYCL backend would be broken by some PRs due to no online CI.
 
 The following release is verified with good quality:
 
-|Commit ID|Tag|Release|Verified  Platform|
-|-|-|-|-|
-|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
+|Commit ID|Tag|Release|Verified  Platform| Update date|
+|-|-|-|-|-|
+|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
 
 ## News
 
+- 2024.11
+  - Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer.
 
 - 2024.8
   - Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
@@ -310,12 +313,14 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_
 export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
+# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
+GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -333,8 +338,9 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE
 
 ## AMD
 # Use FP32, FP16 is not supported
-# 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=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
+GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -644,6 +650,7 @@ 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

@@ -186,13 +186,9 @@ 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.                                                                                                  |
 
@@ -230,7 +226,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
   On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
@@ -247,7 +243,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
   ```bash
   HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
   HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+      cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build -- -j 16
   ```
 
@@ -259,7 +255,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
   ```bash
   set PATH=%HIP_PATH%\bin;%PATH%
-  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
   cmake --build build
   ```
   Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
@@ -268,13 +264,6 @@ 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
 
@@ -282,9 +271,9 @@ The following compilation options are also available to tweak performance (yes,
 
 #### w64devkit
 
-Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
+Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
-Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
+Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
 
 Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
 ```sh
@@ -302,6 +291,29 @@ 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
@@ -375,7 +387,7 @@ cmake --build build --config release
 
 You can test with:
 
-`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
+`./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
 
 If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
 ```bash

examples/CMakeLists.txt

@@ -12,13 +12,10 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(cvector-generator)
     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)
@@ -27,28 +24,35 @@ else()
     add_subdirectory(imatrix)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
-    add_subdirectory(llava)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(main)
     add_subdirectory(parallel)
     add_subdirectory(passkey)
     add_subdirectory(perplexity)
-    add_subdirectory(quantize-stats)
     add_subdirectory(quantize)
     add_subdirectory(retrieval)
-    if (GGML_RPC)
-        add_subdirectory(rpc)
-    endif()
     if (LLAMA_BUILD_SERVER)
-    add_subdirectory(server)
-    endif()
-    if (GGML_SYCL)
-        add_subdirectory(sycl)
+        add_subdirectory(server)
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(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/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.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));
+    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));
 
     if (ctx == NULL) {
         LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);

examples/chat-persistent.sh

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

examples/convert_legacy_llama.py

@@ -840,6 +840,8 @@ 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:
@@ -849,12 +851,32 @@ 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/eval-callback/CMakeLists.txt

@@ -5,5 +5,6 @@ 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/infill/infill.cpp

@@ -43,50 +43,6 @@ 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) {
@@ -96,7 +52,6 @@ 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");
@@ -118,7 +73,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sparams;
+    auto & sparams = params.sampling;
 
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
@@ -625,7 +580,6 @@ 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/llava/llava-cli.cpp

@@ -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->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
     if (!smpl) {
         LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);

examples/llava/minicpmv-cli.cpp

@@ -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->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
     return smpl;
 }
 

examples/lookahead/lookahead.cpp

@@ -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.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
 
     // 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.n_draft;
+    const int n_draft = params.speculative.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -40,6 +40,7 @@ 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.n_draft;
+    const int n_draft = params.speculative.n_max;
 
     const bool dump_kv_cache = params.dump_kv_cache;
 
@@ -102,7 +102,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
 
     std::vector<llama_token> draft;
 

examples/main/main.cpp

@@ -62,49 +62,6 @@ 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) {
@@ -115,7 +72,6 @@ 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");
@@ -144,7 +100,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sparams;
+    auto & sparams = params.sampling;
 
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
@@ -209,6 +165,10 @@ 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 =
@@ -218,7 +178,7 @@ 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(&tpp_batch);
+        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
         if (!threadpool_batch) {
             LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
             return 1;
@@ -228,7 +188,7 @@ int main(int argc, char ** argv) {
         tpp.paused = true;
     }
 
-    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+    struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
     if (!threadpool) {
         LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
         return 1;
@@ -926,7 +886,6 @@ 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);
 
@@ -935,8 +894,8 @@ int main(int argc, char ** argv) {
 
     llama_backend_free();
 
-    ggml_threadpool_free(threadpool);
-    ggml_threadpool_free(threadpool_batch);
+    ggml_threadpool_free_fn(threadpool);
+    ggml_threadpool_free_fn(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.sparams);
+        client.smpl = common_sampler_init(model, params.sampling);
     }
 
     std::vector<llama_token> tokens_system;

examples/perplexity/perplexity.cpp

@@ -34,55 +34,6 @@ 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];
@@ -2072,8 +2023,6 @@ 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, bool use_reference,
+    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,
     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.from_float(input_scratch, quantized_scratch, chunk_size);
+        qfns_cpu.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, bool use_reference,
+    std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, 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, use_reference, input_scratch_ptr, quantized_scratch.data(),
+        test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, 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, nelements, layer, use_reference, input_scratch_ptr,
+        auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, 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, use_reference, input_scratch_ptr + offset,
+                test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,
                         quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
             }
         };
@@ -371,8 +371,9 @@ 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);
-        if (qfns->from_float && qfns->to_float) {
+        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) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }
@@ -393,7 +394,7 @@ int main(int argc, char ** argv) {
                 test_roundtrip_on_layer(
                         layer_name,
                         params.per_layer_stats,
-                        *qfns,
+                        *qfns, *qfns_cpu,
                         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.sparams.top_k);
-            for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
+            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("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/save-load-state/save-load-state.cpp

@@ -9,7 +9,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.prompt = "The quick brown fox";
-    params.sparams.seed = 1234;
+    params.sampling.seed = 1234;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
@@ -42,7 +42,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sampling.seed));
 
     // tokenize prompt
     auto tokens = common_tokenize(ctx, params.prompt, true);
@@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sampling.seed));
 
     printf("\nsecond run: %s", params.prompt.c_str());
 
@@ -169,7 +169,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sampling.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
 

examples/server/CMakeLists.txt

@@ -15,22 +15,13 @@ set(TARGET_SRCS
     httplib.h
 )
 set(PUBLIC_ASSETS
-    colorthemes.css
-    style.css
-    theme-beeninorder.css
-    theme-ketivah.css
-    theme-mangotango.css
-    theme-playground.css
-    theme-polarnight.css
-    theme-snowstorm.css
     index.html
-    index-new.html
-    index.js
     completion.js
-    system-prompts.js
-    prompt-formats.js
-    json-schema-to-grammar.mjs
     loading.html
+    deps_daisyui.min.css
+    deps_markdown-it.js
+    deps_tailwindcss.js
+    deps_vue.esm-browser.js
 )
 
 foreach(asset ${PUBLIC_ASSETS})

examples/server/README.md

@@ -39,7 +39,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
 | `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
 | `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
-| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
+| `-c, --ctx-size N` | size of the prompt context (default: 4096, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
 | `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
 | `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
 | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
@@ -64,7 +64,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
-| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
@@ -85,7 +85,6 @@ The project is under active development, and we are [looking for feedback and co
 | `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
 | `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
 | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
-| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
 | `--log-disable` | Log disable |
 | `--log-file FNAME` | Log to file |
 | `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
@@ -99,25 +98,27 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;typ_p;top_p;min_p;temperature) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: dry;top_k;typ_p;top_p;min_p;xtc;temperature) |
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
-| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
+| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
 | `--penalize-nl` | penalize newline tokens (default: false) |
 | `--temp N` | temperature (default: 0.8) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
 | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
 | `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
-| `--dry-base N` | DRY sampling base value (default: 1.75) |
-| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
-| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
-| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | set DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | set 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 sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers<br/> |
 | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
 | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
@@ -381,6 +382,10 @@ node index.js
 
     `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
 
+    `xtc_probability`: Set the chance for token removal via XTC sampler. Default: `0.0`, which is disabled.
+
+    `xtc_threshold`: Set a minimum probability threshold for tokens to be removed via XTC sampler. Default: `0.1` (> `0.5` disables XTC)
+
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -407,9 +412,9 @@ node index.js
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
-    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
+    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true`
 
-    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 
 **Response format**
 
@@ -928,6 +933,16 @@ Apart from error types supported by OAI, we also have custom types that are spec
 }
 ```
 
+### Legacy completion web UI
+
+A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
+
+For example:
+
+```sh
+./llama-server -m my_model.gguf -c 8192 --path ./examples/server/public_legacy
+```
+
 ### Extending or building alternative Web Front End
 
 You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.

examples/server/chat.mjs

@@ -1,7 +1,7 @@
 import * as readline from 'node:readline'
 import { stdin, stdout } from 'node:process'
 import { readFileSync } from 'node:fs'
-import { SchemaConverter }  from './public/json-schema-to-grammar.mjs'
+import { SchemaConverter }  from './public_legacy/json-schema-to-grammar.mjs'
 
 const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(

examples/server/deps.sh

@@ -6,5 +6,20 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 PUBLIC=$DIR/public
 
 echo "download js bundle files"
-curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
-echo >> $PUBLIC/index.js # add newline
+
+# Note for contributors: Always pin to a specific version "maj.min.patch" to avoid breaking the CI
+
+curl -L https://cdn.tailwindcss.com/3.4.14 > $PUBLIC/deps_tailwindcss.js
+echo >> $PUBLIC/deps_tailwindcss.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/styled.min.css > $PUBLIC/deps_daisyui.min.css
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/themes.min.css >> $PUBLIC/deps_daisyui.min.css
+echo >> $PUBLIC/deps_daisyui.min.css # add newline
+
+curl -L https://unpkg.com/vue@3.5.12/dist/vue.esm-browser.js > $PUBLIC/deps_vue.esm-browser.js
+echo >> $PUBLIC/deps_vue.esm-browser.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/markdown-it/13.0.2/markdown-it.js > $PUBLIC/deps_markdown-it.js
+echo >> $PUBLIC/deps_markdown-it.js # add newline
+
+ls -lah $PUBLIC

examples/server/public/completion.js

@@ -1,12 +1,16 @@
 const paramDefaults = {
   stream: true,
-  n_predict: 500,
   temperature: 0.2,
-  stop: ["</s>"]
 };
 
 let generation_settings = null;
 
+export class CompletionError extends Error {
+  constructor(message, name, data) {
+    super(message);
+    this.name = name;
+  }
+};
 
 // Completes the prompt as a generator. Recommended for most use cases.
 //
@@ -29,7 +33,7 @@ export async function* llama(prompt, params = {}, config = {}) {
 
   const completionParams = { ...paramDefaults, ...params, prompt };
 
-  const response = await fetch(`${api_url}/completion`, {
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
     method: 'POST',
     body: JSON.stringify(completionParams),
     headers: {
@@ -41,6 +45,18 @@ export async function* llama(prompt, params = {}, config = {}) {
     signal: controller.signal,
   });
 
+  const status = response.status;
+  if (status !== 200) {
+    try {
+      const body = await response.json();
+      if (body && body.error && body.error.message) {
+        throw new CompletionError(body.error.message, 'ServerError');
+      }
+    } catch (err) {
+      throw new CompletionError(err.message, 'ServerError');
+    }
+  }
+
   const reader = response.body.getReader();
   const decoder = new TextDecoder();
 
@@ -78,7 +94,12 @@ export async function* llama(prompt, params = {}, config = {}) {
       for (const line of lines) {
         const match = regex.exec(line);
         if (match) {
-          result[match[1]] = match[2]
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
           // since we know this is llama.cpp, let's just decode the json in data
           if (result.data) {
             result.data = JSON.parse(result.data);

examples/server/public_legacy/completion.js

@@ -0,0 +1,209 @@
+const paramDefaults = {
+  stream: true,
+  n_predict: 500,
+  temperature: 0.2,
+  stop: ["</s>"]
+};
+
+let generation_settings = null;
+
+
+// Completes the prompt as a generator. Recommended for most use cases.
+//
+// Example:
+//
+//    import { llama } from '/completion.js'
+//
+//    const request = llama("Tell me a joke", {n_predict: 800})
+//    for await (const chunk of request) {
+//      document.write(chunk.data.content)
+//    }
+//
+export async function* llama(prompt, params = {}, config = {}) {
+  let controller = config.controller;
+  const api_url = config.api_url?.replace(/\/+$/, '') || "";
+
+  if (!controller) {
+    controller = new AbortController();
+  }
+
+  const completionParams = { ...paramDefaults, ...params, prompt };
+
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
+    method: 'POST',
+    body: JSON.stringify(completionParams),
+    headers: {
+      'Connection': 'keep-alive',
+      'Content-Type': 'application/json',
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
+    },
+    signal: controller.signal,
+  });
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+
+  let content = "";
+  let leftover = ""; // Buffer for partially read lines
+
+  try {
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        break;
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
+
+      // Split the text into lines
+      let lines = text.split('\n');
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
+
+            // yield
+            yield result;
+
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
+          if (result.error) {
+            try {
+              result.error = JSON.parse(result.error);
+              if (result.error.message.includes('slot unavailable')) {
+                // Throw an error to be caught by upstream callers
+                throw new Error('slot unavailable');
+              } else {
+                console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
+              }
+            } catch(e) {
+              console.error(`llama.cpp error ${result.error}`)
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    if (e.name !== 'AbortError') {
+      console.error("llama error: ", e);
+    }
+    throw e;
+  }
+  finally {
+    controller.abort();
+  }
+
+  return content;
+}
+
+// Call llama, return an event target that you can subscribe to
+//
+// Example:
+//
+//    import { llamaEventTarget } from '/completion.js'
+//
+//    const conn = llamaEventTarget(prompt)
+//    conn.addEventListener("message", (chunk) => {
+//      document.write(chunk.detail.content)
+//    })
+//
+export const llamaEventTarget = (prompt, params = {}, config = {}) => {
+  const eventTarget = new EventTarget();
+  (async () => {
+    let content = "";
+    for await (const chunk of llama(prompt, params, config)) {
+      if (chunk.data) {
+        content += chunk.data.content;
+        eventTarget.dispatchEvent(new CustomEvent("message", { detail: chunk.data }));
+      }
+      if (chunk.data.generation_settings) {
+        eventTarget.dispatchEvent(new CustomEvent("generation_settings", { detail: chunk.data.generation_settings }));
+      }
+      if (chunk.data.timings) {
+        eventTarget.dispatchEvent(new CustomEvent("timings", { detail: chunk.data.timings }));
+      }
+    }
+    eventTarget.dispatchEvent(new CustomEvent("done", { detail: { content } }));
+  })();
+  return eventTarget;
+}
+
+// Call llama, return a promise that resolves to the completed text. This does not support streaming
+//
+// Example:
+//
+//     llamaPromise(prompt).then((content) => {
+//       document.write(content)
+//     })
+//
+//     or
+//
+//     const content = await llamaPromise(prompt)
+//     document.write(content)
+//
+export const llamaPromise = (prompt, params = {}, config = {}) => {
+  return new Promise(async (resolve, reject) => {
+    let content = "";
+    try {
+      for await (const chunk of llama(prompt, params, config)) {
+        content += chunk.data.content;
+      }
+      resolve(content);
+    } catch (error) {
+      reject(error);
+    }
+  });
+};
+
+/**
+ * (deprecated)
+ */
+export const llamaComplete = async (params, controller, callback) => {
+  for await (const chunk of llama(params.prompt, params, { controller })) {
+    callback(chunk);
+  }
+}
+
+// Get the model info from the server. This is useful for getting the context window and so on.
+export const llamaModelInfo = async (config = {}) => {
+  if (!generation_settings) {
+    const api_url = config.api_url?.replace(/\/+$/, '') || "";
+    const props = await fetch(`${api_url}/props`).then(r => r.json());
+    generation_settings = props.default_generation_settings;
+  }
+  return generation_settings;
+}

examples/server/public_legacy/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

examples/server/server.cpp

@@ -2,10 +2,11 @@
 
 #include "arg.h"
 #include "common.h"
-#include "log.h"
-#include "sampling.h"
 #include "json-schema-to-grammar.h"
 #include "llama.h"
+#include "log.h"
+#include "sampling.h"
+#include "speculative.h"
 
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
@@ -14,22 +15,13 @@
 #define MIMETYPE_JSON "application/json; charset=utf-8"
 
 // auto generated files (update with ./deps.sh)
-#include "colorthemes.css.hpp"
-#include "style.css.hpp"
-#include "theme-beeninorder.css.hpp"
-#include "theme-ketivah.css.hpp"
-#include "theme-mangotango.css.hpp"
-#include "theme-playground.css.hpp"
-#include "theme-polarnight.css.hpp"
-#include "theme-snowstorm.css.hpp"
 #include "index.html.hpp"
-#include "index-new.html.hpp"
-#include "index.js.hpp"
 #include "completion.js.hpp"
-#include "system-prompts.js.hpp"
-#include "prompt-formats.js.hpp"
-#include "json-schema-to-grammar.mjs.hpp"
 #include "loading.html.hpp"
+#include "deps_daisyui.min.css.hpp"
+#include "deps_markdown-it.js.hpp"
+#include "deps_tailwindcss.js.hpp"
+#include "deps_vue.esm-browser.js.hpp"
 
 #include <atomic>
 #include <condition_variable>
@@ -111,9 +103,15 @@ struct server_task_result {
     bool error;
 };
 
+struct server_static_file {
+    const unsigned char * data;
+    unsigned int size;
+    const char * mime_type;
+};
+
 struct slot_params {
     bool stream       = true;
-    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+    bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt
 
     int32_t n_keep    =  0; // number of tokens to keep from initial prompt
     int32_t n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
@@ -124,12 +122,21 @@ struct slot_params {
     int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
 
     std::vector<std::string> antiprompt;
+
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
 };
 
 struct server_slot {
     int id;
     int id_task = -1;
 
+    llama_batch batch_spec;
+
+    llama_context * ctx_dft = nullptr;
+
+    common_speculative * spec = nullptr;
+
     // the index relative to completion multi-task request
     size_t index = 0;
 
@@ -178,7 +185,6 @@ struct server_slot {
     // sampling
     json json_schema;
 
-    struct common_sampler_params sparams;
     struct common_sampler * smpl = nullptr;
 
     llama_token sampled;
@@ -215,7 +221,7 @@ struct server_slot {
         generated_token_probs.clear();
     }
 
-    bool has_budget(common_params &global_params) {
+    bool has_budget(const common_params & global_params) {
         if (params.n_predict == -1 && global_params.n_predict == -1) {
             return true; // limitless
         }
@@ -235,6 +241,10 @@ struct server_slot {
         return state != SLOT_STATE_IDLE;
     }
 
+    bool can_speculate() const {
+        return ctx_dft && params.speculative.n_max > 0 && params.cache_prompt;
+    }
+
     void add_token(const completion_token_output & token) {
         if (!is_processing()) {
             SLT_WRN(*this, "%s", "slot is not processing\n");
@@ -594,11 +604,14 @@ struct server_response {
 };
 
 struct server_context {
+    common_params params_base;
+
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
     std::vector<common_lora_adapter_container> loras;
 
-    common_params params;
+    llama_model * model_dft = nullptr;
+    llama_context_params cparams_dft;
 
     llama_batch batch = {};
 
@@ -631,27 +644,41 @@ struct server_context {
             model = nullptr;
         }
 
+        if (model_dft) {
+            llama_free_model(model_dft);
+            model_dft = nullptr;
+        }
+
         // Clear any sampling context
         for (server_slot & slot : slots) {
-            if (slot.smpl != nullptr) {
-                common_sampler_free(slot.smpl);
-            }
+            common_sampler_free(slot.smpl);
+            slot.smpl = nullptr;
+
+            llama_free(slot.ctx_dft);
+            slot.ctx_dft = nullptr;
+
+            common_speculative_free(slot.spec);
+            slot.spec = nullptr;
+
+            llama_batch_free(slot.batch_spec);
         }
 
         llama_batch_free(batch);
     }
 
-    bool load_model(const common_params & params_) {
-        params = params_;
+    bool load_model(const common_params & params) {
+        SRV_INF("loading model '%s'\n", params.model.c_str());
 
-        common_init_result llama_init = common_init_from_params(params);
+        params_base = params;
+
+        common_init_result llama_init = common_init_from_params(params_base);
 
         model = llama_init.model;
         ctx   = llama_init.context;
         loras = llama_init.lora_adapters;
 
         if (model == nullptr) {
-            SRV_ERR("failed to load model, '%s'\n", params.model.c_str());
+            SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
             return false;
         }
 
@@ -660,32 +687,88 @@ struct server_context {
         add_bos_token = llama_add_bos_token(model);
         has_eos_token = !llama_add_eos_token(model);
 
+        if (!params_base.speculative.model.empty()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
+
+            auto params_dft = params_base;
+
+            params_dft.devices      = params_base.speculative.devices;
+            params_dft.model        = params_base.speculative.model;
+            params_dft.n_ctx        = params_base.speculative.n_ctx;
+            params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
+
+            common_init_result llama_init_dft = common_init_from_params(params_dft);
+
+            model_dft = llama_init_dft.model;
+
+            if (model_dft == nullptr) {
+                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
+                return false;
+            }
+
+            if (!common_speculative_are_compatible(ctx, llama_init_dft.context)) {
+                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
+
+                llama_free      (llama_init_dft.context);
+                llama_free_model(llama_init_dft.model);
+
+                return false;
+            }
+
+            cparams_dft = common_context_params_to_llama(params_base);
+            cparams_dft.n_batch = llama_n_ctx(llama_init_dft.context);
+
+            // the context is not needed - we will create one for each slot
+            llama_free(llama_init_dft.context);
+        }
+
         return true;
     }
 
     bool validate_model_chat_template() const {
-        llama_chat_message chat[] = {{"user", "test"}};
-
-        const int res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0);
-
-        return res > 0;
+        std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
+        std::string template_key = "tokenizer.chat_template";
+        int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
+        if (res >= 0) {
+            llama_chat_message chat[] = {{"user", "test"}};
+            std::string tmpl = std::string(model_template.data(), model_template.size());
+            int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
+            return chat_res > 0;
+        }
+        return false;
     }
 
     void init() {
-        const int32_t n_ctx_slot = n_ctx / params.n_parallel;
+        const int32_t n_ctx_slot = n_ctx / params_base.n_parallel;
 
-        SRV_INF("initializing slots, n_slots = %d\n", params.n_parallel);
+        SRV_INF("initializing slots, n_slots = %d\n", params_base.n_parallel);
 
-        for (int i = 0; i < params.n_parallel; i++) {
+        for (int i = 0; i < params_base.n_parallel; i++) {
             server_slot slot;
 
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
-            slot.n_predict = params.n_predict;
+            slot.n_predict = params_base.n_predict;
+
+            if (model_dft) {
+                slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
+
+                slot.ctx_dft = llama_new_context_with_model(model_dft, cparams_dft);
+                if (slot.ctx_dft == nullptr) {
+                    SRV_ERR("%s", "failed to create draft context\n");
+                    return;
+                }
+
+                slot.spec = common_speculative_init(slot.ctx_dft);
+                if (slot.spec == nullptr) {
+                    SRV_ERR("%s", "failed to create speculator\n");
+                    return;
+                }
+            }
 
             SLT_INF(slot, "new slot n_ctx_slot = %d\n", slot.n_ctx);
 
-            slot.sparams = params.sparams;
+            slot.params.sampling = params_base.sampling;
 
             slot.callback_on_release = [this](int) {
                 queue_tasks.pop_deferred_task();
@@ -705,7 +788,7 @@ struct server_context {
             const int32_t n_batch = llama_n_batch(ctx);
 
             // only a single seq_id per token is needed
-            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params_base.n_parallel), 0, 1);
         }
 
         metrics.init();
@@ -741,7 +824,7 @@ struct server_context {
                 }
 
                 // length of the Longest Common Subsequence between the current slot's prompt and the input prompt
-                int cur_lcs_len = longest_common_subsequence(slot.cache_tokens, task.prompt_tokens);
+                int cur_lcs_len = common_lcs(slot.cache_tokens, task.prompt_tokens);
 
                 // fraction of the common subsequence length compared to the current slot's prompt length
                 float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(slot.cache_tokens.size());
@@ -784,9 +867,11 @@ struct server_context {
     }
 
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
-        slot_params default_params;
         // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
-        auto default_sparams = params.sparams;
+        slot_params defaults;
+        defaults.sampling    = params_base.sampling;
+        defaults.speculative = params_base.speculative;
+
         const auto & data = task.data;
 
         if (data.count("__oaicompat") != 0) {
@@ -797,42 +882,48 @@ struct server_context {
             slot.oaicompat_model = "";
         }
 
-        slot.params.stream              = json_value(data, "stream",             false);
-        slot.params.cache_prompt        = json_value(data, "cache_prompt",       false);
-        slot.params.n_predict           = json_value(data, "n_predict",          json_value(data, "max_tokens", default_params.n_predict));
-        slot.params.n_indent            = json_value(data, "n_indent",           default_params.n_indent);
-        slot.sparams.top_k              = json_value(data, "top_k",              default_sparams.top_k);
-        slot.sparams.top_p              = json_value(data, "top_p",              default_sparams.top_p);
-        slot.sparams.min_p              = json_value(data, "min_p",              default_sparams.min_p);
-        slot.sparams.xtc_probability    = json_value(data, "xtc_probability",    default_sparams.xtc_probability);
-        slot.sparams.xtc_threshold      = json_value(data, "xtc_threshold",      default_sparams.xtc_threshold);
-        slot.sparams.typ_p              = json_value(data, "typical_p",          default_sparams.typ_p);
-        slot.sparams.temp               = json_value(data, "temperature",        default_sparams.temp);
-        slot.sparams.dynatemp_range     = json_value(data, "dynatemp_range",     default_sparams.dynatemp_range);
-        slot.sparams.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  default_sparams.dynatemp_exponent);
-        slot.sparams.penalty_last_n     = json_value(data, "repeat_last_n",      default_sparams.penalty_last_n);
-        slot.sparams.penalty_repeat     = json_value(data, "repeat_penalty",     default_sparams.penalty_repeat);
-        slot.sparams.penalty_freq       = json_value(data, "frequency_penalty",  default_sparams.penalty_freq);
-        slot.sparams.penalty_present    = json_value(data, "presence_penalty",   default_sparams.penalty_present);
-        slot.sparams.dry_multiplier     = json_value(data, "dry_multiplier",     default_sparams.dry_multiplier);
-        slot.sparams.dry_base           = json_value(data, "dry_base",           default_sparams.dry_base);
-        slot.sparams.dry_allowed_length = json_value(data, "dry_allowed_length", default_sparams.dry_allowed_length);
-        slot.sparams.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", default_sparams.dry_penalty_last_n);
-        slot.sparams.mirostat           = json_value(data, "mirostat",           default_sparams.mirostat);
-        slot.sparams.mirostat_tau       = json_value(data, "mirostat_tau",       default_sparams.mirostat_tau);
-        slot.sparams.mirostat_eta       = json_value(data, "mirostat_eta",       default_sparams.mirostat_eta);
-        slot.sparams.penalize_nl        = json_value(data, "penalize_nl",        default_sparams.penalize_nl);
-        slot.params.n_keep              = json_value(data, "n_keep",             default_params.n_keep);
-        slot.params.n_discard           = json_value(data, "n_discard",          default_params.n_discard);
-        slot.sparams.seed               = json_value(data, "seed",               default_sparams.seed);
-        slot.sparams.n_probs            = json_value(data, "n_probs",            default_sparams.n_probs);
-        slot.sparams.min_keep           = json_value(data, "min_keep",           default_sparams.min_keep);
-      //slot.params.t_max_prompt_ms     = json_value(data, "t_max_prompt_ms",    default_params.t_max_prompt_ms); // TODO: implement
-        slot.params.t_max_predict_ms    = json_value(data, "t_max_predict_ms",   default_params.t_max_predict_ms);
+        slot.params.stream           = json_value(data, "stream",             false);
+        slot.params.cache_prompt     = json_value(data, "cache_prompt",       true);
+        slot.params.n_predict        = json_value(data, "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
+        slot.params.n_indent         = json_value(data, "n_indent",           defaults.n_indent);
+        slot.params.n_keep           = json_value(data, "n_keep",             defaults.n_keep);
+        slot.params.n_discard        = json_value(data, "n_discard",          defaults.n_discard);
+      //slot.params.t_max_prompt_ms  = json_value(data, "t_max_prompt_ms",    defaults.t_max_prompt_ms); // TODO: implement
+        slot.params.t_max_predict_ms = json_value(data, "t_max_predict_ms",   defaults.t_max_predict_ms);
 
-        if (slot.sparams.dry_base < 1.0f)
-        {
-           slot.sparams.dry_base = default_sparams.dry_base;
+        slot.params.sampling.top_k              = json_value(data, "top_k",              defaults.sampling.top_k);
+        slot.params.sampling.top_p              = json_value(data, "top_p",              defaults.sampling.top_p);
+        slot.params.sampling.min_p              = json_value(data, "min_p",              defaults.sampling.min_p);
+        slot.params.sampling.xtc_probability    = json_value(data, "xtc_probability",    defaults.sampling.xtc_probability);
+        slot.params.sampling.xtc_threshold      = json_value(data, "xtc_threshold",      defaults.sampling.xtc_threshold);
+        slot.params.sampling.typ_p              = json_value(data, "typical_p",          defaults.sampling.typ_p);
+        slot.params.sampling.temp               = json_value(data, "temperature",        defaults.sampling.temp);
+        slot.params.sampling.dynatemp_range     = json_value(data, "dynatemp_range",     defaults.sampling.dynatemp_range);
+        slot.params.sampling.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  defaults.sampling.dynatemp_exponent);
+        slot.params.sampling.penalty_last_n     = json_value(data, "repeat_last_n",      defaults.sampling.penalty_last_n);
+        slot.params.sampling.penalty_repeat     = json_value(data, "repeat_penalty",     defaults.sampling.penalty_repeat);
+        slot.params.sampling.penalty_freq       = json_value(data, "frequency_penalty",  defaults.sampling.penalty_freq);
+        slot.params.sampling.penalty_present    = json_value(data, "presence_penalty",   defaults.sampling.penalty_present);
+        slot.params.sampling.dry_multiplier     = json_value(data, "dry_multiplier",     defaults.sampling.dry_multiplier);
+        slot.params.sampling.dry_base           = json_value(data, "dry_base",           defaults.sampling.dry_base);
+        slot.params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length);
+        slot.params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n);
+        slot.params.sampling.mirostat           = json_value(data, "mirostat",           defaults.sampling.mirostat);
+        slot.params.sampling.mirostat_tau       = json_value(data, "mirostat_tau",       defaults.sampling.mirostat_tau);
+        slot.params.sampling.mirostat_eta       = json_value(data, "mirostat_eta",       defaults.sampling.mirostat_eta);
+        slot.params.sampling.penalize_nl        = json_value(data, "penalize_nl",        defaults.sampling.penalize_nl);
+        slot.params.sampling.seed               = json_value(data, "seed",               defaults.sampling.seed);
+        slot.params.sampling.n_probs            = json_value(data, "n_probs",            defaults.sampling.n_probs);
+        slot.params.sampling.min_keep           = json_value(data, "min_keep",           defaults.sampling.min_keep);
+
+        slot.params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min);
+        slot.params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max);
+        slot.params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
+
+        slot.params.speculative.n_min = std::min(slot.params.speculative.n_max, slot.params.speculative.n_min);
+
+        if (slot.params.sampling.dry_base < 1.0f) {
+           slot.params.sampling.dry_base = defaults.sampling.dry_base;
         }
 
         // sequence breakers for DRY
@@ -841,8 +932,8 @@ struct server_context {
             // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39
 
             if (data.contains("dry_sequence_breakers")) {
-                slot.sparams.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
-                if (slot.sparams.dry_sequence_breakers.empty()) {
+                slot.params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
+                if (slot.params.sampling.dry_sequence_breakers.empty()) {
                     send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST);
                     return false;
                 }
@@ -856,14 +947,14 @@ struct server_context {
         }
         if (data.contains("json_schema") && !data.contains("grammar")) {
             try {
-                auto schema          = json_value(data, "json_schema", json::object());
-                slot.sparams.grammar = json_schema_to_grammar(schema);
+                auto schema                  = json_value(data, "json_schema", json::object());
+                slot.params.sampling.grammar = json_schema_to_grammar(schema);
             } catch (const std::exception & e) {
                 send_error(task, std::string("\"json_schema\": ") + e.what(), ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
         } else {
-            slot.sparams.grammar = json_value(data, "grammar", default_sparams.grammar);
+            slot.params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar);
         }
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@@ -873,10 +964,10 @@ struct server_context {
         }
 
         {
-            slot.sparams.logit_bias.clear();
+            slot.params.sampling.logit_bias.clear();
 
             if (json_value(data, "ignore_eos", false) && has_eos_token) {
-                slot.sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
+                slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY});
             }
 
             const auto & logit_bias = data.find("logit_bias");
@@ -897,12 +988,12 @@ struct server_context {
                         if (el[0].is_number_integer()) {
                             llama_token tok = el[0].get<llama_token>();
                             if (tok >= 0 && tok < n_vocab) {
-                                slot.sparams.logit_bias.push_back({tok, bias});
+                                slot.params.sampling.logit_bias.push_back({tok, bias});
                             }
                         } else if (el[0].is_string()) {
                             auto toks = common_tokenize(model, el[0].get<std::string>(), false);
                             for (auto tok : toks) {
-                                slot.sparams.logit_bias.push_back({tok, bias});
+                                slot.params.sampling.logit_bias.push_back({tok, bias});
                             }
                         }
                     }
@@ -925,16 +1016,24 @@ struct server_context {
 
         {
             const auto & samplers = data.find("samplers");
-            if (samplers != data.end() && samplers->is_array()) {
-                std::vector<std::string> sampler_names;
-                for (const auto & name : *samplers) {
-                    if (name.is_string()) {
-                        sampler_names.emplace_back(name);
+            if (samplers != data.end()) {
+                if (samplers->is_array()) {
+                    std::vector<std::string> sampler_names;
+                    for (const auto & name : *samplers) {
+                        if (name.is_string()) {
+                            sampler_names.emplace_back(name);
+                        }
                     }
+                    slot.params.sampling.samplers = common_sampler_types_from_names(sampler_names, false);
+                } else if (samplers->is_string()){
+                    std::string sampler_string;
+                    for (const auto & name : *samplers) {
+                        sampler_string += name;
+                    }
+                    slot.params.sampling.samplers = common_sampler_types_from_chars(sampler_string);
                 }
-                slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
             } else {
-                slot.sparams.samplers = default_sparams.samplers;
+                slot.params.sampling.samplers = defaults.sampling.samplers;
             }
         }
 
@@ -943,7 +1042,7 @@ struct server_context {
                 common_sampler_free(slot.smpl);
             }
 
-            slot.smpl = common_sampler_init(model, slot.sparams);
+            slot.smpl = common_sampler_init(model, slot.params.sampling);
             if (slot.smpl == nullptr) {
                 // for now, the only error that may happen here is invalid grammar
                 send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
@@ -951,6 +1050,12 @@ struct server_context {
             }
         }
 
+        if (slot.ctx_dft) {
+            llama_batch_free(slot.batch_spec);
+
+            slot.batch_spec = llama_batch_init(slot.params.speculative.n_max + 1, 0, 1);
+        }
+
         slot.state = SLOT_STATE_STARTED;
 
         SLT_INF(slot, "%s", "processing task\n");
@@ -968,7 +1073,7 @@ struct server_context {
 
     bool process_token(completion_token_output & result, server_slot & slot) {
         // remember which tokens were sampled - used for repetition penalties during sampling
-        const std::string token_str = common_token_to_piece(ctx, result.tok, params.special);
+        const std::string token_str = common_token_to_piece(ctx, result.tok, params_base.special);
         slot.sampled = result.tok;
 
         // search stop word and delete it
@@ -1033,7 +1138,7 @@ struct server_context {
         }
 
         // check the limits
-        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params)) {
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params_base)) {
             slot.stopped_limit  = true;
             slot.has_next_token = false;
 
@@ -1126,50 +1231,54 @@ struct server_context {
 
     json get_formated_generation(const server_slot & slot) const {
         std::vector<std::string> samplers;
-        samplers.reserve(slot.sparams.samplers.size());
-        for (const auto & sampler : slot.sparams.samplers) {
+        samplers.reserve(slot.params.sampling.samplers.size());
+        for (const auto & sampler : slot.params.sampling.samplers) {
             samplers.emplace_back(common_sampler_type_to_str(sampler));
         }
 
         return json {
             {"n_ctx",                     slot.n_ctx},
             {"n_predict",                 slot.n_predict},     // Server configured n_predict
-            {"model",                     params.model_alias},
-            {"seed",                      slot.sparams.seed},
+            {"model",                     params_base.model_alias},
+            {"seed",                      slot.params.sampling.seed},
             {"seed_cur",                  slot.smpl ? common_sampler_get_seed(slot.smpl) : 0},
-            {"temperature",               slot.sparams.temp},
-            {"dynatemp_range",            slot.sparams.dynatemp_range},
-            {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},
-            {"top_k",                     slot.sparams.top_k},
-            {"top_p",                     slot.sparams.top_p},
-            {"min_p",                     slot.sparams.min_p},
-            {"xtc_probability",           slot.sparams.xtc_probability},
-            {"xtc_threshold",             slot.sparams.xtc_threshold},
-            {"typical_p",                 slot.sparams.typ_p},
-            {"repeat_last_n",             slot.sparams.penalty_last_n},
-            {"repeat_penalty",            slot.sparams.penalty_repeat},
-            {"presence_penalty",          slot.sparams.penalty_present},
-            {"frequency_penalty",         slot.sparams.penalty_freq},
-            {"dry_multiplier",            slot.sparams.dry_multiplier},
-            {"dry_base",                  slot.sparams.dry_base},
-            {"dry_allowed_length",        slot.sparams.dry_allowed_length},
-            {"dry_penalty_last_n",        slot.sparams.dry_penalty_last_n},
-            {"dry_sequence_breakers",     slot.sparams.dry_sequence_breakers},
-            {"mirostat",                  slot.sparams.mirostat},
-            {"mirostat_tau",              slot.sparams.mirostat_tau},
-            {"mirostat_eta",              slot.sparams.mirostat_eta},
-            {"penalize_nl",               slot.sparams.penalize_nl},
+            {"temperature",               slot.params.sampling.temp},
+            {"dynatemp_range",            slot.params.sampling.dynatemp_range},
+            {"dynatemp_exponent",         slot.params.sampling.dynatemp_exponent},
+            {"top_k",                     slot.params.sampling.top_k},
+            {"top_p",                     slot.params.sampling.top_p},
+            {"min_p",                     slot.params.sampling.min_p},
+            {"xtc_probability",           slot.params.sampling.xtc_probability},
+            {"xtc_threshold",             slot.params.sampling.xtc_threshold},
+            {"typical_p",                 slot.params.sampling.typ_p},
+            {"repeat_last_n",             slot.params.sampling.penalty_last_n},
+            {"repeat_penalty",            slot.params.sampling.penalty_repeat},
+            {"presence_penalty",          slot.params.sampling.penalty_present},
+            {"frequency_penalty",         slot.params.sampling.penalty_freq},
+            {"dry_multiplier",            slot.params.sampling.dry_multiplier},
+            {"dry_base",                  slot.params.sampling.dry_base},
+            {"dry_allowed_length",        slot.params.sampling.dry_allowed_length},
+            {"dry_penalty_last_n",        slot.params.sampling.dry_penalty_last_n},
+            {"dry_sequence_breakers",     slot.params.sampling.dry_sequence_breakers},
+            {"mirostat",                  slot.params.sampling.mirostat},
+            {"mirostat_tau",              slot.params.sampling.mirostat_tau},
+            {"mirostat_eta",              slot.params.sampling.mirostat_eta},
+            {"penalize_nl",               slot.params.sampling.penalize_nl},
             {"stop",                      slot.params.antiprompt},
             {"max_tokens",                slot.params.n_predict}, // User configured n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
-            {"ignore_eos",                slot.sparams.ignore_eos},
+            {"ignore_eos",                slot.params.sampling.ignore_eos},
             {"stream",                    slot.params.stream},
-          //{"logit_bias",                slot.sparams.logit_bias},
-            {"n_probs",                   slot.sparams.n_probs},
-            {"min_keep",                  slot.sparams.min_keep},
-            {"grammar",                   slot.sparams.grammar},
+          //{"logit_bias",                slot.params.sampling.logit_bias},
+            {"n_probs",                   slot.params.sampling.n_probs},
+            {"min_keep",                  slot.params.sampling.min_keep},
+            {"grammar",                   slot.params.sampling.grammar},
             {"samplers",                  samplers},
+            {"speculative",               slot.can_speculate()},
+            {"speculative.n_max",         slot.params.speculative.n_max},
+            {"speculative.n_min",         slot.params.speculative.n_min},
+            {"speculative.p_min",         slot.params.speculative.p_min},
         };
     }
 
@@ -1206,7 +1315,7 @@ struct server_context {
             {"index",      slot.index},
         };
 
-        if (slot.sparams.n_probs > 0) {
+        if (slot.params.sampling.n_probs > 0) {
             const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
             const size_t probs_pos      = std::min(slot.n_sent_token_probs,                       slot.generated_token_probs.size());
             const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size());
@@ -1239,7 +1348,7 @@ struct server_context {
             {"content",             !slot.params.stream ? slot.generated_text : ""},
             {"id_slot",             slot.id},
             {"stop",                true},
-            {"model",               params.model_alias},
+            {"model",               params_base.model_alias},
             {"tokens_predicted",    slot.n_decoded},
             {"tokens_evaluated",    slot.n_prompt_tokens},
             {"generation_settings", get_formated_generation(slot)},
@@ -1255,7 +1364,7 @@ struct server_context {
             {"index",               slot.index},
         };
 
-        if (slot.sparams.n_probs > 0) {
+        if (slot.params.sampling.n_probs > 0) {
             std::vector<completion_token_output> probs;
             if (!slot.params.stream && slot.stopped_word) {
                 const llama_tokens stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
@@ -1412,10 +1521,10 @@ struct server_context {
                             data.at("input_prefix"),
                             data.at("input_suffix"),
                             data.at("input_extra"),
-                            params.n_batch,
-                            params.n_predict,
+                            params_base.n_batch,
+                            params_base.n_predict,
                             slots[0].n_ctx, // TODO: there should be a better way
-                            params.spm_infill,
+                            params_base.spm_infill,
                             tokenized_prompts[i]
                         );
                         create_task(data, tokens);
@@ -1788,7 +1897,7 @@ struct server_context {
         // TODO: simplify and improve
         for (server_slot & slot : slots) {
             if (slot.is_processing() && slot.n_past + 1 >= slot.n_ctx) {
-                if (!params.ctx_shift) {
+                if (!params_base.ctx_shift) {
                     // this check is redundant (for good)
                     // we should never get here, because generation should already stopped in process_token()
                     slot.release();
@@ -1854,7 +1963,7 @@ struct server_context {
         int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
 
         // next, batch any pending prompts without exceeding n_batch
-        if (params.cont_batching || batch.n_tokens == 0) {
+        if (params_base.cont_batching || batch.n_tokens == 0) {
             for (auto & slot : slots) {
                 // this slot still has a prompt to be processed
                 if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
@@ -1907,7 +2016,7 @@ struct server_context {
                                 continue;
                             }
                         } else {
-                            if (!params.ctx_shift) {
+                            if (!params_base.ctx_shift) {
                                 // if context shift is disabled, we make sure prompt size is smaller than KV size
                                 // TODO: there should be a separate parameter that control prompt truncation
                                 //       context shift should be applied only during the generation phase
@@ -1950,14 +2059,14 @@ struct server_context {
 
                             if (slot.params.cache_prompt) {
                                 // reuse any previously computed tokens that are common with the new prompt
-                                slot.n_past = longest_common_prefix(slot.cache_tokens, prompt_tokens);
+                                slot.n_past = common_lcp(slot.cache_tokens, prompt_tokens);
 
                                 // reuse chunks from the cached prompt by shifting their KV cache in the new position
-                                if (params.n_cache_reuse > 0) {
+                                if (params_base.n_cache_reuse > 0) {
                                     size_t head_c = slot.n_past; // cache
                                     size_t head_p = slot.n_past; // current prompt
 
-                                    SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params.n_cache_reuse, slot.n_past);
+                                    SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params_base.n_cache_reuse, slot.n_past);
 
                                     while (head_c < slot.cache_tokens.size() &&
                                            head_p < prompt_tokens.size()) {
@@ -1970,7 +2079,7 @@ struct server_context {
                                             n_match++;
                                         }
 
-                                        if (n_match >= (size_t) params.n_cache_reuse) {
+                                        if (n_match >= (size_t) params_base.n_cache_reuse) {
                                             SLT_INF(slot, "reusing chunk with size %zu, shifting KV cache [%zu, %zu) -> [%zu, %zu)\n", n_match, head_c, head_c + n_match, head_p, head_p + n_match);
                                             //for (size_t i = head_p; i < head_p + n_match; i++) {
                                             //    SLT_DBG(slot, "cache token %3zu: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str());
@@ -2158,38 +2267,99 @@ struct server_context {
                     continue; // continue loop of slots
                 }
 
-                completion_token_output result;
-                const llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
+                llama_token id;
 
-                common_sampler_accept(slot.smpl, id, true);
+                {
+                    completion_token_output result;
 
-                slot.n_decoded += 1;
-                if (slot.n_decoded == 1) {
-                    slot.t_start_generation = ggml_time_us();
-                    slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
-                    metrics.on_prompt_eval(slot);
+                    id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
+
+                    slot.i_batch = -1;
+
+                    common_sampler_accept(slot.smpl, id, true);
+
+                    slot.n_decoded += 1;
+                    if (slot.n_decoded == 1) {
+                        slot.t_start_generation = ggml_time_us();
+                        slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
+                        metrics.on_prompt_eval(slot);
+                    }
+
+                    result.tok = id;
+
+                    const auto * cur_p = common_sampler_get_candidates(slot.smpl);
+
+                    for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
+                        result.probs.push_back({
+                            cur_p->data[i].id,
+                                i >= cur_p->size ? 0.0f : cur_p->data[i].p,
+                        });
+                    }
+
+                    if (!process_token(result, slot)) {
+                        // release slot because of stop condition
+                        slot.release();
+                        slot.print_timings();
+                        send_final_response(slot);
+                        metrics.on_prediction(slot);
+                        continue;
+                    }
                 }
 
-                result.tok = id;
-
-                const auto * cur_p = common_sampler_get_candidates(slot.smpl);
-
-                for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
-                    result.probs.push_back({
-                        cur_p->data[i].id,
-                        i >= cur_p->size ? 0.0f : cur_p->data[i].p,
-                    });
+                // check if the slot supports speculative decoding
+                if (!slot.can_speculate()) {
+                    continue;
                 }
 
-                if (!process_token(result, slot)) {
-                    // release slot because of stop condition
-                    slot.release();
-                    slot.print_timings();
-                    send_final_response(slot);
-                    metrics.on_prediction(slot);
+                struct common_speculative_params params_spec;
+                params_spec.n_draft   = slot.params.speculative.n_max;
+                params_spec.n_reuse   = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max;
+                params_spec.p_min     = slot.params.speculative.p_min;
+
+                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+
+                // ignore small drafts
+                if (slot.params.speculative.n_min > (int) draft.size()) {
+                    continue;
                 }
 
-                slot.i_batch = -1;
+                // construct the speculation batch
+                common_batch_clear(slot.batch_spec);
+                common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
+
+                for (size_t i = 0; i < draft.size(); ++i) {
+                    common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true);
+                }
+
+                llama_decode(ctx, slot.batch_spec);
+
+                // the accepted tokens from the speculation
+                const auto ids = common_sampler_sample_and_accept_n(slot.smpl, ctx, draft);
+
+                slot.n_past    += ids.size();
+                slot.n_decoded += ids.size();
+
+                slot.cache_tokens.push_back(id);
+                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+
+                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+
+                for (size_t i = 0; i < ids.size(); ++i) {
+                    completion_token_output result;
+
+                    result.tok = ids[i];
+
+                    if (!process_token(result, slot)) {
+                        // release slot because of stop condition
+                        slot.release();
+                        slot.print_timings();
+                        send_final_response(slot);
+                        metrics.on_prediction(slot);
+                        break;
+                    }
+                }
+
+                SRV_DBG("accepted %d/%d draft tokens\n", (int) ids.size() - 1, (int) draft.size());
             }
         }
 
@@ -2263,6 +2433,16 @@ int main(int argc, char ** argv) {
     LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     LOG_INF("\n");
 
+    // static files
+    std::map<std::string, server_static_file> static_files = {
+        { "/",                        { index_html,              index_html_len,              "text/html; charset=utf-8" }},
+        { "/completion.js",           { completion_js,           completion_js_len,           "text/javascript; charset=utf-8" }},
+        { "/deps_daisyui.min.css",    { deps_daisyui_min_css,    deps_daisyui_min_css_len,    "text/css; charset=utf-8" }},
+        { "/deps_markdown-it.js",     { deps_markdown_it_js,     deps_markdown_it_js_len,     "text/javascript; charset=utf-8" }},
+        { "/deps_tailwindcss.js",     { deps_tailwindcss_js,     deps_tailwindcss_js_len,     "text/javascript; charset=utf-8" }},
+        { "/deps_vue.esm-browser.js", { deps_vue_esm_browser_js, deps_vue_esm_browser_js_len, "text/javascript; charset=utf-8" }},
+    };
+
     std::unique_ptr<httplib::Server> svr;
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
     if (params.ssl_file_key != "" && params.ssl_file_cert != "") {
@@ -2285,16 +2465,6 @@ int main(int argc, char ** argv) {
     std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
 
     svr->set_default_headers({{"Server", "llama.cpp"}});
-
-    // CORS preflight
-    svr->Options(R"(.*)", [](const httplib::Request &, httplib::Response & res) {
-        // Access-Control-Allow-Origin is already set by middleware
-        res.set_header("Access-Control-Allow-Credentials", "true");
-        res.set_header("Access-Control-Allow-Methods",     "POST");
-        res.set_header("Access-Control-Allow-Headers",     "*");
-        return res.set_content("", "text/html"); // blank response, no data
-    });
-
     svr->set_logger(log_server_request);
 
     auto res_error = [](httplib::Response & res, const json & error_data) {
@@ -2353,7 +2523,7 @@ int main(int argc, char ** argv) {
     // Middlewares
     //
 
-    auto middleware_validate_api_key = [&params, &res_error](const httplib::Request & req, httplib::Response & res) {
+    auto middleware_validate_api_key = [&params, &res_error, &static_files](const httplib::Request & req, httplib::Response & res) {
         static const std::unordered_set<std::string> public_endpoints = {
             "/health",
             "/models",
@@ -2365,8 +2535,8 @@ int main(int argc, char ** argv) {
             return true;
         }
 
-        // If path is public, skip validation
-        if (public_endpoints.find(req.path) != public_endpoints.end()) {
+        // If path is public or is static file, skip validation
+        if (public_endpoints.find(req.path) != public_endpoints.end() || static_files.find(req.path) != static_files.end()) {
             return true;
         }
 
@@ -2407,6 +2577,14 @@ int main(int argc, char ** argv) {
     // register server middlewares
     svr->set_pre_routing_handler([&middleware_validate_api_key, &middleware_server_state](const httplib::Request & req, httplib::Response & res) {
         res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        // If this is OPTIONS request, skip validation because browsers don't include Authorization header
+        if (req.method == "OPTIONS") {
+            res.set_header("Access-Control-Allow-Credentials", "true");
+            res.set_header("Access-Control-Allow-Methods",     "GET, POST");
+            res.set_header("Access-Control-Allow-Headers",     "*");
+            res.set_content("", "text/html"); // blank response, no data
+            return httplib::Server::HandlerResponse::Handled; // skip further processing
+        }
         if (!middleware_server_state(req, res)) {
             return httplib::Server::HandlerResponse::Handled;
         }
@@ -2679,7 +2857,7 @@ int main(int argc, char ** argv) {
     const auto handle_props = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
-            { "total_slots",                 ctx_server.params.n_parallel },
+            { "total_slots",                 ctx_server.params_base.n_parallel },
             { "chat_template",               llama_get_chat_template(ctx_server.model) },
         };
 
@@ -2687,7 +2865,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_props_change = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params.endpoint_props) {
+        if (!ctx_server.params_base.endpoint_props) {
             res_error(res, format_error_response("This server does not support changing global properties. Start it with `--props`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -2700,7 +2878,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -2806,7 +2984,7 @@ int main(int argc, char ** argv) {
 
     // TODO: maybe merge this function with "handle_completions_generic"
     const auto handle_chat_completions = [&ctx_server, &params, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -2983,7 +3161,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params.reranking || ctx_server.params.embedding) {
+        if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -3102,13 +3280,6 @@ int main(int argc, char ** argv) {
         res.status = 200; // HTTP OK
     };
 
-    auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
-        return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
-            res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
-            return false;
-        };
-    };
-
     //
     // Router
     //
@@ -3116,33 +3287,20 @@ int main(int argc, char ** argv) {
     // register static assets routes
     if (!params.public_path.empty()) {
         // Set the base directory for serving static files
-        svr->set_base_dir(params.public_path);
-    }
-
-    if (!params.api_keys.empty()) {
-        // for now, if API key is set, web UI is unusable
-        svr->Get("/", [&](const httplib::Request &, httplib::Response & res) {
-            return res.set_content("Web UI is disabled because API key is set.", "text/html; charset=utf-8");
-        });
+        bool is_found = svr->set_mount_point("/", params.public_path);
+        if (!is_found) {
+            LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str());
+            return 1;
+        }
     } else {
         // using embedded static files
-        svr->Get("/",                           handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
-        svr->Get("/index.js",                   handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/completion.js",              handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/json-schema-to-grammar.mjs", handle_static_file(json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8"));
-
-        // add new-ui files
-        svr->Get("/colorthemes.css",       handle_static_file(colorthemes_css, colorthemes_css_len, "text/css; charset=utf-8"));
-        svr->Get("/style.css",             handle_static_file(style_css, style_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-beeninorder.css", handle_static_file(theme_beeninorder_css, theme_beeninorder_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-ketivah.css",     handle_static_file(theme_ketivah_css, theme_ketivah_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-mangotango.css",  handle_static_file(theme_mangotango_css, theme_mangotango_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-playground.css",  handle_static_file(theme_playground_css, theme_playground_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-polarnight.css",  handle_static_file(theme_polarnight_css, theme_polarnight_css_len, "text/css; charset=utf-8"));
-        svr->Get("/theme-snowstorm.css",   handle_static_file(theme_snowstorm_css, theme_snowstorm_css_len, "text/css; charset=utf-8"));
-        svr->Get("/index-new.html",        handle_static_file(index_new_html, index_new_html_len, "text/html; charset=utf-8"));
-        svr->Get("/system-prompts.js",     handle_static_file(system_prompts_js, system_prompts_js_len, "text/javascript; charset=utf-8"));
-        svr->Get("/prompt-formats.js",     handle_static_file(prompt_formats_js, prompt_formats_js_len, "text/javascript; charset=utf-8"));
+        for (const auto & it : static_files) {
+            const server_static_file & static_file = it.second;
+            svr->Get(it.first.c_str(), [&static_file](const httplib::Request &, httplib::Response & res) {
+                res.set_content(reinterpret_cast<const char*>(static_file.data), static_file.size, static_file.mime_type);
+                return false;
+            });
+        }
     }
 
     // register API routes

examples/server/tests/features/security.feature

@@ -64,5 +64,5 @@ Feature: Security
       | localhost       | Access-Control-Allow-Origin      | localhost         |
       | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
       | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
+      | web.mydomain.fr | Access-Control-Allow-Methods     | GET, POST         |
       | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |

examples/server/utils.hpp

@@ -24,7 +24,6 @@
 #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::ordered_json;
-using llama_tokens = std::vector<llama_token>;
 
 #define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
 #define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
@@ -439,62 +438,6 @@ static std::string gen_chatcmplid() {
 // other common utils
 //
 
-static size_t longest_common_prefix(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;
-}
-
-static size_t longest_common_subsequence(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;
-}
-
 static bool ends_with(const std::string & str, const std::string & suffix) {
     return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
 }

examples/simple-chat/simple-chat.cpp

@@ -62,6 +62,9 @@ int main(int argc, char ** argv) {
         }
     }, nullptr);
 
+    // load dynamic backends
+    ggml_backend_load_all();
+
     // initialize the model
     llama_model_params model_params = llama_model_default_params();
     model_params.n_gpu_layers = ngl;

examples/simple/simple.cpp

@@ -74,6 +74,10 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // load dynamic backends
+
+    ggml_backend_load_all();
+
     // initialize the model
 
     llama_model_params model_params = llama_model_default_params();

examples/speculative-simple/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-speculative-simple)
+add_executable(${TARGET} speculative-simple.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/speculative-simple/README.md

@@ -0,0 +1,12 @@
+# llama.cpp/examples/speculative-simple
+
+Demonstration of basic greedy speculative decoding
+
+```bash
+./bin/llama-speculative-simple \
+    -m  ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
+    -md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
+    -f test.txt -c 0 -ngl 99 --color \
+    --sampling-seq k --top-k 1 -fa --temp 0.0 \
+    -ngld 99 --draft-max 16 --draft-min 5 --draft-p-min 0.9
+```

examples/speculative-simple/speculative-simple.cpp

@@ -0,0 +1,274 @@
+#include "arg.h"
+#include "common.h"
+#include "sampling.h"
+#include "speculative.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
+        return 1;
+    }
+
+    if (params.n_predict < -1) {
+        LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
+        return 1;
+    }
+
+    common_init();
+
+    if (params.speculative.model.empty()) {
+        LOG_ERR("%s: --model-draft is required\n", __func__);
+        return 1;
+    }
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model_tgt = NULL;
+    llama_model * model_dft = NULL;
+
+    llama_context * ctx_tgt = NULL;
+    llama_context * ctx_dft = NULL;
+
+    // load the target model
+    common_init_result llama_init_tgt = common_init_from_params(params);
+
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt   = llama_init_tgt.context;
+
+    // load the draft model
+    params.devices      = params.speculative.devices;
+    params.model        = params.speculative.model;
+    params.n_ctx        = params.speculative.n_ctx;
+    params.n_batch      = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
+    params.n_gpu_layers = params.speculative.n_gpu_layers;
+
+    if (params.speculative.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
+    }
+
+    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    common_init_result llama_init_dft = common_init_from_params(params);
+
+    model_dft = llama_init_dft.model;
+    ctx_dft   = llama_init_dft.context;
+
+    if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
+        return 1;
+    }
+
+    // Tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = common_tokenize(ctx_tgt, params.prompt, true, true);
+
+    if (llama_n_ctx(ctx_tgt) < (int) inp.size()) {
+        LOG_ERR("%s: the prompt exceeds the context size (%d tokens, ctx %d)\n", __func__, (int) inp.size(), llama_n_ctx(ctx_tgt));
+
+        return 1;
+    }
+
+    if (llama_n_batch(ctx_tgt) < (int) inp.size()) {
+        LOG_ERR("%s: the prompt exceeds the batch size (%d tokens, batch %d)\n", __func__, (int) inp.size(), llama_n_batch(ctx_tgt));
+
+        return 1;
+    }
+
+    LOG("\n\n");
+
+    for (auto id : inp) {
+        LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
+    }
+
+    // how many tokens to draft each time
+    int n_draft     = params.speculative.n_max;
+    int n_draft_min = params.speculative.n_min;
+
+    float p_min = params.speculative.p_min;
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    // used to determine end of generation
+    bool has_eos = false;
+
+    // ================================================
+    // everything until here is standard initialization
+    // the relevant stuff for speculative decoding starts here
+
+    const auto t_enc_start = ggml_time_us();
+
+    // target model sampling context
+    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
+
+    // eval the prompt
+    llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
+
+    // note: keep the last token separate!
+    llama_token id_last = inp.back();
+
+    // all tokens currently in the target context
+    auto prompt_tgt = std::vector<llama_token>(inp.begin(), inp.end() - 1);
+
+    int n_past = inp.size() - 1;
+
+    // init the speculator
+    struct common_speculative_params params_spec;
+    params_spec.n_draft = n_draft;
+    params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
+    params_spec.p_min   = p_min;
+
+    struct common_speculative * spec = common_speculative_init(ctx_dft);
+
+    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
+
+    const auto t_enc_end = ggml_time_us();
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // optionally, generate draft tokens that can be appended to the target batch
+        //
+        // this is the most important part of the speculation. the more probable tokens that are provided here
+        // the better the performance will be. in theory, this computation can be performed asynchronously and even
+        // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
+        // from a cache or lookup tables.
+        //
+        llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last);
+
+        //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
+
+        // always have a token to evaluate from before - id_last
+        common_batch_clear(batch_tgt);
+        common_batch_add  (batch_tgt, id_last, n_past++, { 0 }, true);
+
+        // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
+        {
+            // do not waste time on small drafts
+            if (draft.size() < n_draft_min) {
+                draft.clear();
+            }
+
+            for (size_t i = 0; i < draft.size(); ++i) {
+                common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+            }
+
+            //LOG_DBG("target batch: %s\n", string_from(ctx_tgt, batch_tgt).c_str());
+
+            llama_decode(ctx_tgt, batch_tgt);
+        }
+
+        // sample from the full target batch and return the accepted tokens based on the target sampler
+        //
+        // for each token to be accepted, the sampler would have to sample that same token
+        // in such cases, instead of decoding the sampled token as we normally do, we simply continue with the
+        // available logits from the batch and sample the next token until we run out of logits or the sampler
+        // disagrees with the draft
+        //
+        const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
+
+        //LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
+
+        GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
+
+        n_past    += ids.size() - 1;
+        n_drafted += batch_tgt.n_tokens - 1;
+        n_accept  += ids.size() - 1;
+
+        // process the accepted tokens and update contexts
+        //
+        // this is the standard token post-processing that we normally do
+        // in this case, we do it for a group of accepted tokens at once
+        //
+        {
+            llama_token id;
+            std::string token_str;
+
+            for (size_t i = 0; i < ids.size(); ++i) {
+                id = ids[i];
+
+                ++n_predict;
+
+                if (llama_token_is_eog(model_tgt, id)) {
+                    has_eos = true;
+                    break;
+                }
+
+                token_str = common_token_to_piece(ctx_tgt, id);
+
+                if (params.use_color && i + 1 < ids.size()) {
+                    LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
+                } else {
+                    LOG("%s", token_str.c_str());
+                }
+            }
+
+            if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
+                break;
+            }
+
+            LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d, '%s')\n", (int) ids.size() - 1, (int) draft.size(), id, token_str.c_str());
+
+            {
+                LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
+
+                llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+            }
+
+            prompt_tgt.push_back(id_last);
+            prompt_tgt.insert(prompt_tgt.end(), ids.begin(), ids.end() - 1);
+
+            // remember the last accepted token for the next iteration
+            id_last = id;
+        }
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    const int n_input = inp.size();
+
+    LOG("\n\n");
+
+    LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_INF("\n");
+    LOG_INF("n_draft   = %d\n", n_draft);
+    LOG_INF("n_predict = %d\n", n_predict);
+    LOG_INF("n_drafted = %d\n", n_drafted);
+    LOG_INF("n_accept  = %d\n", n_accept);
+    LOG_INF("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_INF("\n");
+    LOG_INF("draft:\n\n");
+
+    llama_perf_context_print(ctx_dft);
+
+    LOG_INF("\n");
+    LOG_INF("target:\n\n");
+    common_perf_print(ctx_tgt, smpl);
+
+    common_sampler_free(smpl);
+    common_speculative_free(spec);
+
+    llama_free(ctx_tgt);
+    llama_free_model(model_tgt);
+
+    llama_free(ctx_dft);
+    llama_free_model(model_dft);
+
+    llama_backend_free();
+
+    LOG("\n\n");
+
+    return 0;
+}

examples/speculative/speculative.cpp

@@ -12,7 +12,7 @@
 #include <string>
 #include <vector>
 
-#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
 struct seq_draft {
@@ -33,7 +33,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     // needed to get candidate probs even for temp <= 0.0
-    params.sparams.n_probs = 128;
+    params.sampling.n_probs = 128;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
@@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.model_draft.empty()) {
+    if (params.speculative.model.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -55,9 +55,9 @@ int main(int argc, char ** argv) {
     const int n_seq_dft = params.n_parallel;
 
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
-    const float p_split  = params.p_split;
+    const float p_draft_split = params.speculative.p_split;
 
-    std::default_random_engine rng(params.sparams.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sparams.seed);
+    std::default_random_engine rng(params.sampling.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sampling.seed);
     std::uniform_real_distribution<> u_dist;
 
     // init llama.cpp
@@ -76,13 +76,14 @@ int main(int argc, char ** argv) {
     ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
-    params.model = params.model_draft;
-    params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.draft_cpuparams.n_threads > 0) {
-        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
+    params.devices = params.speculative.devices;
+    params.model = params.speculative.model;
+    params.n_gpu_layers = params.speculative.n_gpu_layers;
+    if (params.speculative.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
     }
 
-    params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
+    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     common_init_result llama_init_dft = common_init_from_params(params);
     model_dft = llama_init_dft.model;
     ctx_dft = llama_init_dft.context;
@@ -170,7 +171,7 @@ int main(int argc, char ** argv) {
     //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
 
     // how many tokens to draft each time
-    int n_draft = params.n_draft;
+    int n_draft = params.speculative.n_max;
 
     int n_predict = 0;
     int n_drafted = 0;
@@ -183,14 +184,14 @@ int main(int argc, char ** argv) {
     bool has_eos = false;
 
     // target model sampling context (reuse the llama_context's sampling instance)
-    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
 
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
     for (int s = 0; s < n_seq_dft; ++s) {
         // allocate llama_sampler for each draft sequence
-        drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
+        drafts[s].smpl = common_sampler_init(model_dft, params.sampling);
     }
 
     llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
@@ -230,7 +231,7 @@ int main(int argc, char ** argv) {
             // for stochastic sampling, attempt to match the token with the drafted tokens
             {
                 bool accept = false;
-                if (params.sparams.temp > 0) {
+                if (params.sampling.temp > 0) {
                     // stochastic verification
                     common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
@@ -267,11 +268,12 @@ int main(int argc, char ** argv) {
                         for (size_t i = 0; i < dist_tgt.size; i++) {
                             if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
                                 p_tgt = dist_tgt.data[i].p;
+                                break;
                             }
+                        }
+                        for (size_t i = 0; i < dist_dft.size; i++) {
                             if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
                                 p_dft = dist_dft.data[i].p;
-                            }
-                            if (p_tgt && p_dft) {
                                 break;
                             }
                         }
@@ -493,7 +495,7 @@ int main(int argc, char ** argv) {
 
                 // attempt to split the branch if the probability is high enough
                 for (int f = 1; f < 8; ++f) {
-                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
+                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
                         llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1730200266,
-        "narHash": "sha256-l253w0XMT8nWHGXuXqyiIC/bMvh1VRszGXgdpQlfhvU=",
+        "lastModified": 1732014248,
+        "narHash": "sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "807e9154dcb16384b1b765ebe9cd2bba2ac287fd",
+        "rev": "23e89b7da85c3640bbc2173fe04f4bd114342367",
         "type": "github"
       },
       "original": {