speculative : add infill mode

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

@@ -22,7 +22,7 @@ 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

.devops/llama-cli-cuda.Dockerfile

@@ -22,7 +22,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    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 \;

.devops/llama-cli-intel.Dockerfile

@@ -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

@@ -15,7 +15,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+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 \;

.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

@@ -22,7 +22,7 @@ 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 ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-server -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;

.devops/llama-server-intel.Dockerfile

@@ -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

@@ -15,7 +15,7 @@ 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 ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-server -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;

.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

.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

@@ -952,7 +952,7 @@ jobs:
 
     env:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+      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
@@ -962,7 +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
@@ -981,26 +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

.github/workflows/docker.yml

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

.gitignore

@@ -134,3 +134,7 @@ poetry.toml
 
 # Test models for lora adapters
 /lora-tests
+
+# Local scripts
+/run-vim.sh
+/run-chat.sh

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
 #
@@ -156,8 +163,11 @@ if (GGML_TARGET_DEFINES)
     list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
 endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
-
-set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
+# all public headers
+set(LLAMA_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
+set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(

Makefile

@@ -34,6 +34,7 @@ BUILD_TARGETS = \
 	llama-server \
 	llama-simple \
 	llama-simple-chat \
+	llama-run \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -251,7 +252,7 @@ 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
@@ -290,6 +291,7 @@ 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,
@@ -635,10 +637,6 @@ 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
@@ -647,20 +645,6 @@ 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
@@ -669,12 +653,6 @@ 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
@@ -754,10 +732,10 @@ 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)
@@ -769,8 +747,8 @@ $(_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
 
@@ -783,10 +761,6 @@ 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_HIP -DGGML_USE_CUDA
 
 ifdef GGML_HIP_UMA
@@ -800,13 +774,6 @@ 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
@@ -869,10 +836,6 @@ ifdef GGML_MUSA
 
 	MUSAFLAGS += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
 
-ifdef GGML_CUDA_FORCE_DMMV
-	MUSAFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
-
 ifdef GGML_CUDA_FORCE_MMQ
 	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # GGML_CUDA_FORCE_MMQ
@@ -881,18 +844,6 @@ ifdef GGML_CUDA_FORCE_CUBLAS
 	MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
 endif # GGML_CUDA_FORCE_CUBLAS
 
-ifdef GGML_CUDA_DMMV_X
-	MUSAFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-else
-	MUSAFLAGS += -DGGML_CUDA_DMMV_X=32
-endif # GGML_CUDA_DMMV_X
-
-ifdef GGML_CUDA_MMV_Y
-	MUSAFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-else
-	MUSAFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # GGML_CUDA_MMV_Y
-
 ifdef GGML_CUDA_F16
 	MUSAFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_F16
@@ -901,12 +852,6 @@ ifdef GGML_CUDA_DMMV_F16
 	MUSAFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_DMMV_F16
 
-ifdef GGML_CUDA_KQUANTS_ITER
-	MUSAFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-else
-	MUSAFLAGS += -DK_QUANTS_PER_ITERATION=2
-endif
-
 ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
 	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
@@ -963,6 +908,7 @@ endif # GGML_METAL
 ifdef GGML_METAL
 ggml/src/ggml-metal/ggml-metal.o: \
 	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/include/ggml-metal.h \
 	ggml/include/ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
@@ -970,9 +916,11 @@ ggml/src/ggml-metal/ggml-metal.o: \
 ifdef GGML_METAL_EMBED_LIBRARY
 ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/__embed_ggml-common.h__/r ggml/src/ggml-common.h' -e '/__embed_ggml-common.h__/d' < ggml/src/ggml-metal/ggml-metal.metal > ggml/src/ggml-metal/ggml-metal-embed.metal
+	@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
@@ -1020,6 +968,7 @@ OBJ_COMMON = \
 	$(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
 
@@ -1219,6 +1168,11 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-run: examples/run/run.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-simple: examples/simple/simple.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

Package.swift

@@ -43,7 +43,8 @@ linkerSettings.append(.linkedFramework("Accelerate"))
 cSettings.append(
     contentsOf: [
         .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL")
+        .define("GGML_USE_METAL"),
+        .define("GGML_USE_CPU")
     ]
 )
 #endif

README.md

@@ -459,14 +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)

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_HIP        @GGML_HIP@)
+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)",
@@ -2037,5 +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
 //

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)
@@ -180,25 +191,29 @@ struct common_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
+    // offload params
+    std::vector<ggml_backend_dev_t> devices;         // devices to use for offloading
+    int32_t n_gpu_layers                    =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu                        =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]               =   {0}; // how split tensors should be distributed across GPUs
+    enum llama_split_mode        split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+
     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
@@ -451,7 +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);
 
@@ -461,7 +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);
 
@@ -472,6 +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
 //

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 +1,271 @@
+#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;
+    struct common_sampler * smpl_infill;
+
+    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,
+        /* .smpl_infill = */ nullptr,
+        /* .batch       = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .prompt      = */ {},
+    };
+
+    {
+        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);
+    }
+
+    {
+        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_infill = common_sampler_init(llama_get_model(ctx_dft), params);
+    }
+
+    return result;
+}
+
+void common_speculative_free(struct common_speculative * spec) {
+    common_sampler_free(spec->smpl);
+    common_sampler_free(spec->smpl_infill);
+
+    llama_batch_free(spec->batch);
+
+    delete spec;
+}
+
+bool common_speculative_are_compatible(
+        const struct llama_context * ctx_tgt,
+        const struct llama_context * ctx_dft) {
+    const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
+    const struct llama_model * model_dft = llama_get_model(ctx_dft);
+
+    const bool vocab_type_tgt = llama_vocab_type(model_tgt);
+    LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
+
+    const bool vocab_type_dft = llama_vocab_type(model_dft);
+    LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
+
+    if (vocab_type_tgt != vocab_type_dft) {
+        LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
+                     "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
+        return false;
+    }
+
+    if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
+        llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
+        llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
+        LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
+        return false;
+    }
+
+    {
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
+
+        const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
+
+        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
+            LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
+                         "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    __func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            return false;
+        }
+
+        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
+            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
+                LOG_ERR("%s: draft model vocab must match target model to use speculation but "
+                             "token %d content differs - target '%s', draft '%s'\n", __func__, i,
+                        common_token_to_piece(ctx_tgt, i).c_str(),
+                        common_token_to_piece(ctx_dft, i).c_str());
+                return false;
+            }
+        }
+    }
+
+    return true;
+}
+
+llama_tokens common_speculative_gen_draft(
+        struct common_speculative * spec,
+        struct common_speculative_params params,
+        const llama_tokens & prompt_tgt,
+        llama_token id_last) {
+    auto & batch  = spec->batch;
+    auto & ctx    = spec->ctx;
+    auto & smpl   = params.infill ? spec->smpl_infill : 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 +1,30 @@
+#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
+
+    bool infill = false; // use infill sampling (useful for FIM)
+};
+
+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

@@ -34,9 +34,10 @@ 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
@@ -312,12 +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
@@ -335,8 +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
@@ -646,6 +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.                                                                                                  |
 
@@ -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
 

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 +24,36 @@ else()
     add_subdirectory(imatrix)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
-    add_subdirectory(llava)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(main)
     add_subdirectory(parallel)
     add_subdirectory(passkey)
     add_subdirectory(perplexity)
-    add_subdirectory(quantize-stats)
     add_subdirectory(quantize)
     add_subdirectory(retrieval)
-    if (GGML_RPC)
-        add_subdirectory(rpc)
-    endif()
     if (LLAMA_BUILD_SERVER)
-    add_subdirectory(server)
-    endif()
-    if (GGML_SYCL)
-        add_subdirectory(sycl)
+        add_subdirectory(server)
     endif()
     add_subdirectory(save-load-state)
+    add_subdirectory(run)
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)
+    add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
+    if (NOT GGML_BACKEND_DL)
+        # these examples use the backends directly and cannot be built with dynamic loading
+        add_subdirectory(convert-llama2c-to-ggml)
+        add_subdirectory(cvector-generator)
+        add_subdirectory(export-lora)
+        add_subdirectory(quantize-stats)
+        add_subdirectory(llava)
+        if (GGML_RPC)
+            add_subdirectory(rpc)
+        endif()
+        if (GGML_SYCL)
+            add_subdirectory(sycl)
+        endif()
+    endif()
 endif()

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/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

@@ -73,7 +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(); });

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

@@ -100,7 +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)
@@ -165,6 +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 =
@@ -174,7 +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;
@@ -184,7 +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;
@@ -890,8 +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/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/run/CMakeLists.txt

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

examples/run/README.md

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

examples/run/run.cpp

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

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/README.md

@@ -412,7 +412,7 @@ 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: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 

examples/server/public/index.html

@@ -81,7 +81,13 @@
               <path d="M14.5 3a1 1 0 0 1-1 1H13v9a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V4h-.5a1 1 0 0 1-1-1V2a1 1 0 0 1 1-1H6a1 1 0 0 1 1-1h2a1 1 0 0 1 1 1h3.5a1 1 0 0 1 1 1zM4.118 4 4 4.059V13a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1V4.059L11.882 4zM2.5 3h11V2h-11z"/>
             </svg>
           </button>
-
+          <button v-if="messages.length > 0" class="btn mr-1" @click="downloadConv(viewingConvId)" :disabled="isGenerating">
+              <!-- download conversation button -->
+              <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-download" viewBox="0 0 16 16">
+                  <path d="M.5 9.9a.5.5 0 0 1 .5.5v2.5a1 1 0 0 0 1 1h12a1 1 0 0 0 1-1v-2.5a.5.5 0 0 1 1 0v2.5a2 2 0 0 1-2 2H2a2 2 0 0 1-2-2v-2.5a.5.5 0 0 1 .5-.5"/>
+                  <path d="M7.646 11.854a.5.5 0 0 0 .708 0l3-3a.5.5 0 0 0-.708-.708L8.5 10.293V1.5a.5.5 0 0 0-1 0v8.793L5.354 8.146a.5.5 0 1 0-.708.708z"/>
+            </svg>
+          </button>
           <button class="btn" @click="showConfigDialog = true" :disabled="isGenerating">
             <!-- edit config button -->
             <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-gear" viewBox="0 0 16 16">
@@ -526,6 +532,23 @@
             this.fetchMessages();
           }
         },
+        downloadConv(convId) {
+          const conversation = StorageUtils.getOneConversation(convId);
+          if (!conversation) {
+            alert('Conversation not found.');
+            return;
+          }
+          const conversationJson = JSON.stringify(conversation, null, 2);
+          const blob = new Blob([conversationJson], { type: 'application/json' });
+          const url = URL.createObjectURL(blob);
+          const a = document.createElement('a');
+          a.href = url;
+          a.download = `conversation_${convId}.json`;
+          document.body.appendChild(a);
+          a.click();
+          document.body.removeChild(a);
+          URL.revokeObjectURL(url);
+        },
         async sendMessage() {
           if (!this.inputMsg) return;
           const currConvId = this.viewingConvId;

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
@@ -110,7 +111,7 @@ struct server_static_file {
 
 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
@@ -121,12 +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;
 
@@ -175,7 +185,6 @@ struct server_slot {
     // sampling
     json json_schema;
 
-    struct common_sampler_params sparams;
     struct common_sampler * smpl = nullptr;
 
     llama_token sampled;
@@ -212,7 +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
         }
@@ -232,6 +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");
@@ -591,11 +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 = {};
 
@@ -628,27 +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;
         }
 
@@ -657,6 +687,41 @@ 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;
     }
 
@@ -674,20 +739,36 @@ struct server_context {
     }
 
     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();
@@ -707,7 +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();
@@ -743,7 +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());
@@ -786,9 +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) {
@@ -799,42 +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
@@ -843,8 +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;
                 }
@@ -858,14 +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) {
@@ -875,10 +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");
@@ -899,12 +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});
                             }
                         }
                     }
@@ -935,16 +1024,16 @@ struct server_context {
                             sampler_names.emplace_back(name);
                         }
                     }
-                    slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
+                    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.sparams.samplers = common_sampler_types_from_chars(sampler_string);
+                    slot.params.sampling.samplers = common_sampler_types_from_chars(sampler_string);
                 }
             } else {
-                slot.sparams.samplers = default_sparams.samplers;
+                slot.params.sampling.samplers = defaults.sampling.samplers;
             }
         }
 
@@ -953,7 +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);
@@ -961,6 +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");
@@ -978,7 +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
@@ -1043,7 +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;
 
@@ -1136,50 +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},
         };
     }
 
@@ -1216,7 +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());
@@ -1249,7 +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)},
@@ -1265,7 +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);
@@ -1422,10 +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);
@@ -1798,7 +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();
@@ -1864,7 +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) {
@@ -1917,7 +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
@@ -1960,14 +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()) {
@@ -1980,7 +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());
@@ -2168,38 +2267,100 @@ 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;
+                params_spec.infill    = slot.inf_type == SERVER_TASK_INF_TYPE_INFILL;
+
+                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());
             }
         }
 
@@ -2697,7 +2858,7 @@ 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) },
         };
 
@@ -2705,7 +2866,7 @@ 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;
         }
@@ -2718,7 +2879,7 @@ 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;
         }
@@ -2824,7 +2985,7 @@ 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;
         }
@@ -3001,7 +3162,7 @@ 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;
         }

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);
 
@@ -494,7 +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": 1730785428,
-        "narHash": "sha256-Zwl8YgTVJTEum+L+0zVAWvXAGbWAuXHax3KzuejaDyo=",
+        "lastModified": 1732014248,
+        "narHash": "sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "4aa36568d413aca0ea84a1684d2d46f55dbabad7",
+        "rev": "23e89b7da85c3640bbc2173fe04f4bd114342367",
         "type": "github"
       },
       "original": {

ggml/CMakeLists.txt

@@ -33,6 +33,7 @@ else()
 endif()
 
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 
 #
 # option list
@@ -109,6 +110,7 @@ if (NOT MSVC)
 endif()
 option(GGML_LASX        "ggml: enable lasx"             ON)
 option(GGML_LSX         "ggml: enable lsx"              ON)
+option(GGML_RVV         "ggml: enable rvv"              ON)
 option(GGML_SVE         "ggml: enable SVE"              OFF)
 
 if (WIN32)
@@ -128,14 +130,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -169,6 +166,8 @@ option(GGML_SYCL                            "ggml: use SYCL"
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
+set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
+                                            "ggml: sycl device architecture")
 
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
@@ -238,12 +237,8 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml PUBLIC_HEADER)
-
-if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml      LIBRARY)
-    install(TARGETS ggml-base LIBRARY)
-endif()
+install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml-base LIBRARY)
 
 # FIXME: this should be done in the backend cmake files
 if (GGML_METAL)

ggml/include/ggml-backend.h

@@ -190,6 +190,14 @@ extern "C" {
     typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
     // Get additional buffer types provided by the device (returns a NULL-terminated array)
     typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    // Set the abort callback for the backend
+    typedef void                         (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data);
+    // Get a list of feature flags supported by the backend (returns a NULL-terminated array)
+    struct ggml_backend_feature {
+        const char * name;
+        const char * value;
+    };
+    typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg);
 
     //
     // Backend registry
@@ -214,6 +222,13 @@ extern "C" {
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
+    // Load a backend from a dynamic library and register it
+    GGML_API ggml_backend_reg_t ggml_backend_load(const char * path);
+    // Unload a backend if loaded dynamically and unregister it
+    GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
+    // Load all known backends from dynamic libraries
+    GGML_API void               ggml_backend_load_all(void);
+
     //
     // Backend scheduler
     //

ggml/include/ggml-cpu.h

@@ -7,29 +7,6 @@
 extern "C" {
 #endif
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -75,14 +52,11 @@ extern "C" {
     GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
     GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
 
-    GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_BACKEND_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_BACKEND_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_BACKEND_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new           (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free          (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume        (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
@@ -104,10 +78,10 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
     GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
     GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
     GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
     GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
     GGML_BACKEND_API int ggml_cpu_has_fma        (void);
-    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
     GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
     GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
     GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);

ggml/include/ggml.h

@@ -2215,6 +2215,37 @@ extern "C" {
 
     GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
 
+    // ggml threadpool
+    // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend
+    // the goal should be to create an API that other backends can use move everything to the ggml base
+
+    // scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/src/CMakeLists.txt

@@ -202,6 +202,10 @@ endif()
 
 # ggml
 
+if (GGML_BACKEND_DL AND NOT BUILD_SHARED_LIBS)
+    message(FATAL_ERROR "GGML_BACKEND_DL requires BUILD_SHARED_LIBS")
+endif()
+
 add_library(ggml-base
             ../include/ggml.h
             ../include/ggml-alloc.h
@@ -226,6 +230,31 @@ add_library(ggml
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    target_link_libraries(ggml PRIVATE dl)
+endif()
+
+function(ggml_add_backend_library backend)
+    if (GGML_BACKEND_DL)
+        add_library(${backend} MODULE ${ARGN})
+        # write the shared library to the output directory
+        set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
+    else()
+        add_library(${backend} ${ARGN})
+        target_link_libraries(ggml PUBLIC ${backend})
+        install(TARGETS ${backend} LIBRARY)
+    endif()
+
+    target_link_libraries(${backend} PRIVATE ggml-base)
+    target_include_directories(${backend} PRIVATE ..)
+
+    if (${BUILD_SHARED_LIBS})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_BUILD)
+        target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
+    endif()
+endfunction()
+
 function(ggml_add_backend backend)
     string(TOUPPER "GGML_${backend}" backend_id)
     if (${backend_id})
@@ -236,14 +265,10 @@ function(ggml_add_backend backend)
         # however, currently it is necessary for AMX, since it is enabled by default on llama.cpp
         if (${backend_id})
             message(STATUS "Including ${backend} backend")
-            if (${BUILD_SHARED_LIBS})
-                target_compile_definitions(${backend_target} PRIVATE GGML_BACKEND_BUILD)
-                target_compile_definitions(${backend_target} PUBLIC  GGML_BACKEND_SHARED)
-                install(TARGETS ${backend_target} LIBRARY)
+            if (NOT GGML_BACKEND_DL)
+                string(TOUPPER "GGML_USE_${backend}" backend_use)
+                target_compile_definitions(ggml PUBLIC ${backend_use})
             endif()
-            target_link_libraries(ggml PUBLIC ${backend_target})
-            string(TOUPPER "GGML_USE_${backend}" backend_use)
-            target_compile_definitions(ggml PUBLIC ${backend_use})
         endif()
     endif()
 endfunction()
@@ -256,10 +281,10 @@ ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
 ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
+ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
-ggml_add_backend(MUSA)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-amx/CMakeLists.txt

@@ -9,12 +9,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MA
 
     file(GLOB   GGML_SOURCES_AMX "*.cpp")
 
-    add_library(ggml-amx
-                ${GGML_HEADERS_AMX}
-                ${GGML_SOURCES_AMX})
-
-    target_link_libraries(ggml-amx PRIVATE ggml-base)
-    target_include_directories(ggml-amx PRIVATE . ..)
+    ggml_add_backend_library(ggml-amx
+                             ${GGML_HEADERS_AMX}
+                             ${GGML_SOURCES_AMX}
+                            )
 
     # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
     # TODO: integrate AMX backend into the CPU backend

ggml/src/ggml-amx/ggml-amx.cpp

@@ -409,8 +409,9 @@ static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
 
 ggml_backend_reg_t ggml_backend_amx_reg(void) {
     static struct ggml_backend_reg ggml_backend_amx_reg = {
-        /* .iface   = */ ggml_backend_amx_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_amx_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_amx_reg;
@@ -444,3 +445,5 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
 }
 
 #endif
+
+GGML_BACKEND_DL_IMPL(ggml_backend_amx_reg)

ggml/src/ggml-backend-impl.h

@@ -8,6 +8,8 @@
 extern "C" {
 #endif
 
+    #define GGML_BACKEND_API_VERSION 1
+
     //
     // Backend buffer type
     //
@@ -63,20 +65,20 @@ extern "C" {
         enum ggml_backend_buffer_usage usage;
     };
 
-    ggml_backend_buffer_t ggml_backend_buffer_init(
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
                    ggml_backend_buffer_type_t buft,
             struct ggml_backend_buffer_i      iface,
                    void *                     context,
                    size_t                     size);
 
     // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     // multi-buffer
     // buffer that contains a collection of buffers
-    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend (stream)
@@ -199,17 +201,37 @@ extern "C" {
     };
 
     struct ggml_backend_reg {
-        // int api_version; // TODO: for dynamic loading
+        int api_version; // initialize to GGML_BACKEND_API_VERSION
         struct ggml_backend_reg_i iface;
         void * context;
     };
 
-
     // Internal backend registry API
-    void ggml_backend_register(ggml_backend_reg_t reg);
-    void ggml_backend_device_register(ggml_backend_dev_t device);
-    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
-    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
+    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
+
+    // Add backend dynamic loading support to the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+
+    #ifdef GGML_BACKEND_DL
+        #ifdef __cplusplus
+        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
+                extern "C" {                                                     \
+                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+                }                                                                \
+                ggml_backend_reg_t ggml_backend_init(void) {                     \
+                    return reg_fn();                                             \
+                }
+        #else
+        #    define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+                GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+                ggml_backend_reg_t ggml_backend_init(void) {                 \
+                    return reg_fn();                                         \
+                }
+        #endif
+    #else
+    #    define GGML_BACKEND_DL_IMPL(reg_fn)
+    #endif
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend-reg.cpp

@@ -1,11 +1,29 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
+#include <algorithm>
 #include <cstring>
+#include <string>
 #include <vector>
 
+#ifdef _WIN32
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#elif defined(__APPLE__)
+#    include <mach-o/dyld.h>
+#    include <dlfcn.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+
 // Backend registry
+#ifdef GGML_USE_CPU
+#include "ggml-cpu.h"
+#endif
 
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
@@ -43,8 +61,13 @@
 #include "ggml-kompute.h"
 #endif
 
+struct ggml_backend_reg_entry {
+    ggml_backend_reg_t reg;
+    void * handle;
+};
+
 struct ggml_backend_registry {
-    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_reg_entry> backends;
     std::vector<ggml_backend_dev_t> devices;
 
     ggml_backend_registry() {
@@ -75,11 +98,19 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_KOMPUTE
         register_backend(ggml_backend_kompute_reg());
 #endif
-
+#ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
+#endif
     }
 
-    void register_backend(ggml_backend_reg_t reg) {
+    ~ggml_backend_registry() {
+        while (!backends.empty()) {
+            // use silent since the log system may have been destroyed at this point
+            unload_backend(backends.back().reg, true);
+        }
+    }
+
+    void register_backend(ggml_backend_reg_t reg, void * handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -88,7 +119,7 @@ struct ggml_backend_registry {
         GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
 #endif
-        backends.push_back(reg);
+        backends.push_back({ reg, handle });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
             register_device(ggml_backend_reg_dev_get(reg, i));
         }
@@ -100,6 +131,111 @@ struct ggml_backend_registry {
 #endif
         devices.push_back(device);
     }
+
+    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+#ifdef _WIN32
+        // suppress error dialogs for missing DLLs
+        DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+        SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+        HMODULE handle = LoadLibraryA(path);
+
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s: %lu\n", __func__, path, GetLastError());
+            }
+            SetErrorMode(old_mode);
+            return nullptr;
+        }
+
+        ggml_backend_init_t backend_init = (ggml_backend_init_t) GetProcAddress(handle, "ggml_backend_init");
+
+        SetErrorMode(old_mode);
+
+        if (!backend_init) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %lu\n", __func__, path, GetLastError());
+            }
+            FreeLibrary(handle);
+            return nullptr;
+        }
+#else
+        void * handle = dlopen(path, RTLD_NOW | RTLD_LOCAL);
+
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path, dlerror());
+            }
+            return nullptr;
+        }
+
+        auto * backend_init = (ggml_backend_init_t) dlsym(handle, "ggml_backend_init");
+
+        if (!backend_init) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %s\n", __func__, path, dlerror());
+            }
+            dlclose(handle);
+            return nullptr;
+        }
+#endif
+        ggml_backend_reg_t reg = backend_init();
+
+        if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
+            if (!silent) {
+                if (!reg) {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                } else {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
+                                   __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                }
+            }
+#ifdef _WIN32
+            FreeLibrary(handle);
+#else
+            dlclose(handle);
+#endif
+            return nullptr;
+        }
+
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+        register_backend(reg, handle);
+        return reg;
+    }
+
+    void unload_backend(ggml_backend_reg_t reg, bool silent) {
+        auto it = std::find_if(backends.begin(), backends.end(),
+                                [reg](ggml_backend_reg_entry entry) { return entry.reg == reg; });
+
+        if (it == backends.end()) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: backend not found\n", __func__);
+            }
+            return;
+        }
+
+        if (!silent) {
+            GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
+        }
+
+        // remove devices
+        devices.erase(
+            std::remove_if(devices.begin(), devices.end(),
+                            [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
+            devices.end());
+
+        // unload library
+        if (it->handle) {
+#ifdef _WIN32
+            FreeLibrary((HMODULE) it->handle);
+#else
+            dlclose(it->handle);
+#endif
+        }
+
+        // remove backend
+        backends.erase(it);
+    }
 };
 
 static ggml_backend_registry & get_reg() {
@@ -117,23 +253,32 @@ void ggml_backend_device_register(ggml_backend_dev_t device) {
 }
 
 // Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
 size_t ggml_backend_reg_count() {
     return get_reg().backends.size();
 }
 
 ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
     GGML_ASSERT(index < ggml_backend_reg_count());
-    return get_reg().backends[index];
+    return get_reg().backends[index].reg;
 }
 
 ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
         ggml_backend_reg_t reg = ggml_backend_reg_get(i);
-        if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
+        if (striequals(ggml_backend_reg_name(reg), name)) {
             return reg;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Device enumeration
@@ -149,11 +294,11 @@ ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
 ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+        if (striequals(ggml_backend_dev_name(dev), name)) {
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
@@ -163,14 +308,14 @@ ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Convenience functions
 ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return ggml_backend_dev_init(dev, params);
 }
@@ -178,7 +323,7 @@ ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params)
 ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return ggml_backend_dev_init(dev, params);
 }
@@ -189,7 +334,97 @@ ggml_backend_t ggml_backend_init_best(void) {
         dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     }
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
-    return ggml_backend_dev_init(dev, NULL);
+    return ggml_backend_dev_init(dev, nullptr);
+}
+
+// Dynamic loading
+ggml_backend_reg_t ggml_backend_load(const char * path) {
+    return get_reg().load_backend(path, false);
+}
+
+void ggml_backend_unload(ggml_backend_reg_t reg) {
+    get_reg().unload_backend(reg, true);
+}
+
+void ggml_backend_load_all() {
+    std::vector<std::string> search_prefix;
+
+    // add the executable directory to the search path
+    // FIXME: this is convenient for development, but it should probably be disabled in production
+
+#if defined(__APPLE__)
+    // get executable path
+    std::vector<char> path;
+    uint32_t size;
+    while (true) {
+        size = path.size();
+        if (_NSGetExecutablePath(path.data(), &size) == 0) {
+            break;
+        }
+        path.resize(size);
+    }
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    search_prefix.push_back(base_path + "/");
+#elif defined(__linux__)
+    std::string base_path = ".";
+    std::vector<char> path(1024);
+    while (true) {
+        // get executable path
+        ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+        if (len == -1) {
+            break;
+        }
+        if (len < (ssize_t) path.size()) {
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
+        }
+        path.resize(path.size() * 2);
+    }
+
+    search_prefix.push_back(base_path + "/");
+#endif
+
+    auto & reg = get_reg();
+
+    auto try_load = [&](const std::string & name) {
+        std::string os_name;
+#ifdef _WIN32
+        os_name = "ggml-" + name + ".dll";
+#else
+        os_name = "libggml-" + name + ".so";
+#endif
+        if (reg.load_backend(os_name.c_str(), true)) {
+            return;
+        }
+        for (const auto & prefix : search_prefix) {
+            if (reg.load_backend((prefix + os_name).c_str(), true)) {
+                return;
+            }
+        }
+    };
+
+    try_load("amx");
+    try_load("blas");
+    try_load("cann");
+    try_load("cuda");
+    try_load("hip");
+    try_load("kompute");
+    try_load("metal");
+    try_load("rpc");
+    try_load("sycl");
+    try_load("vulkan");
+    try_load("musa");
+    try_load("cpu");
 }

ggml/src/ggml-backend.cpp

@@ -252,6 +252,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 }
 
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -266,6 +267,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -689,7 +691,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -722,8 +724,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -742,7 +742,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
     }
 
     // graph input
@@ -1538,12 +1538,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }

ggml/src/ggml-blas/CMakeLists.txt

@@ -11,12 +11,9 @@ find_package(BLAS)
 if (BLAS_FOUND)
     message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
-    add_library(ggml-blas
-                ggml-blas.cpp
-                )
-
-    target_link_libraries(ggml-blas PRIVATE ggml-base)
-    target_include_directories(ggml-blas PRIVATE . ..)
+    ggml_add_backend_library(ggml-blas
+                             ggml-blas.cpp
+                            )
 
     if (${GGML_BLAS_VENDOR} MATCHES "Apple")
         add_compile_definitions(ACCELERATE_NEW_LAPACK)
@@ -75,7 +72,6 @@ if (BLAS_FOUND)
 
     message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
 
-    #add_compile_options(${BLAS_LINKER_FLAGS})
     target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
 
     if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))

ggml/src/ggml-blas/ggml-blas.cpp

@@ -506,9 +506,12 @@ static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
 
 ggml_backend_reg_t ggml_backend_blas_reg(void) {
     static struct ggml_backend_reg ggml_backend_blas_reg = {
-        /* .iface   = */ ggml_backend_blas_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_blas_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_blas_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)

ggml/src/ggml-cann/CMakeLists.txt

@@ -3,6 +3,33 @@ if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOM
     message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
 endif()
 
+# Auto-detech Soc type and Soc version, if detect failed, will abort build
+set(SOC_VERSION "")
+function(detect_ascend_soc_type SOC_VERSION)
+    execute_process(
+        COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
+        OUTPUT_VARIABLE npu_info
+        RESULT_VARIABLE npu_result
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    if("${npu_info}" STREQUAL "" OR ${npu_result})
+        message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
+    endif()
+    set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
+endfunction()
+
+if(NOT SOC_TYPE)
+    detect_ascend_soc_type(SOC_VERSION)
+    set(SOC_TYPE "${SOC_VERSION}")
+    message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
+else()
+    string(TOLOWER ${SOC_TYPE} SOC_VERSION)
+endif()
+
+# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND310P.
+string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
+set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
+
 if (CANN_INSTALL_DIR)
     # Only Support Linux.
     if (NOT UNIX)
@@ -34,11 +61,13 @@ if (CANN_INSTALL_DIR)
 
     file(GLOB GGML_SOURCES_CANN "*.cpp")
 
-    add_library(ggml-cann ${GGML_SOURCES_CANN})
-    target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
-    target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
+    ggml_add_backend_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE ${CANN_INCLUDE_DIRS})
     target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
 
+    target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
+
     message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
     message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
 else()

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2312,6 +2312,14 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     switch (src0->type) {
         case GGML_TYPE_F32:
+        {
+#ifdef ASCEND_310P
+             // Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 8) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f32(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2320,7 +2328,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_F16:
+        {
+#ifdef ASCEND_310P
+             // Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 16) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f16(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2329,6 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_Q4_0:
             aclrtlaunch_ascendc_get_row_q4_0(
                 24, ctx.stream(), src0->data, src1->data, dst->data,

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2064,16 +2064,17 @@ ggml_backend_reg_t ggml_backend_cann_reg() {
                 dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
                 ggml_cann_set_device(i);
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cann_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cann_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cann_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cann_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -2126,3 +2127,5 @@ void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
     ggml_cann_set_device(device);
     ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cann_reg)

ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -1,7 +1,3 @@
-if (NOT SOC_TYPE)
-    set (SOC_TYPE "Ascend910B3")
-endif()
-
 file(GLOB SRC_FILES
     get_row_f32.cpp
     get_row_f16.cpp
@@ -13,7 +9,6 @@ file(GLOB SRC_FILES
     dup.cpp
 )
 
-string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
 set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
 
@@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
+message(STATUS "CANN: compile ascend kernels witch SOC_VERSION:${SOC_VERSION}.")
+ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

ggml/src/ggml-cann/kernels/dup.cpp

@@ -5,6 +5,7 @@
 using namespace AscendC;
 
 #define BUFFER_NUM 2
+const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
 
 template <typename SRC_T, typename DST_T>
 class DupByRows {
@@ -19,6 +20,7 @@ class DupByRows {
         // Input has four dims.
         int64_t op_block_num = GetBlockNum();
         int64_t op_block_idx = GetBlockIdx();
+        assert(op_block_idx < SUPPORTED_MAX_DIM && op_block_idx >= 0, "Invalid block index:%d, max is:%d\n", op_block_idx, SUPPORTED_MAX_DIM);
 
         // param
         num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
@@ -51,24 +53,36 @@ class DupByRows {
 
     __aicore__ inline void copy_in() {
         LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
-        DataCopyPadExtParams<SRC_T> padParams;
-        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
-
+        const size_t elem_per_block = 32 / sizeof(SRC_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
+        DataCopy(src_local, src_gm, cpy_elements_len);
         src_queue.EnQue(src_local);
     }
 
     __aicore__ inline void copy_out() {
         LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-
+#ifdef ASCEND_310P
+        const size_t elem_per_block = 32 / sizeof(DST_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t len = num_elem & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(dst_gm, dst_local, len);
+        }
+        if(tail != 0) {
+            for (size_t i = tail; i < elem_per_block; i++) {
+                dst_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
+            SetAtomicNone();
+        }
+#else
         DataCopyExtParams dataCopyParams;
         dataCopyParams.blockCount = 1;
         dataCopyParams.blockLen = num_elem * sizeof(DST_T);
         DataCopyPad(dst_gm, dst_local, dataCopyParams);
-
+#endif
         dst_queue.FreeTensor(dst_local);
     }
 

ggml/src/ggml-cann/kernels/get_row_f16.cpp

@@ -14,7 +14,7 @@ class GET_ROW_F16 {
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         // TODO, use template for F16/f32
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -59,32 +59,42 @@ class GET_ROW_F16 {
     }
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        size_t origin_len = len;
         LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(half);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(half);
-            DataCopyPadExtParams<half> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -150,6 +160,7 @@ class GET_ROW_F16 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>

ggml/src/ggml-cann/kernels/get_row_f32.cpp

@@ -13,7 +13,7 @@ class GET_ROW_F32 {
                                 int64_t *indices_ne_ub, size_t *indices_nb_ub,
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -55,31 +55,40 @@ class GET_ROW_F32 {
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
         LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPadExtParams<float> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -144,6 +153,7 @@ class GET_ROW_F32 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>

ggml/src/ggml-cann/kernels/get_row_q4_0.cpp

@@ -110,9 +110,12 @@ class GET_ROW_Q4_0 {
         LocalTensor<float> output_local = output_queue.AllocTensor<float>();
 
         // TODO: cast more data to speed up.
+#ifdef ASCEND_310P
+        // TODO: 310P support quantification
+#else
         Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
         Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
-
+#endif
         // Only mul need compile by group.
         half scale = scale_gm.GetValue(scale_offset);
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -1,14 +1,13 @@
-add_library(ggml-cpu
-            ggml-cpu.c
-            ggml-cpu.cpp
-            ggml-cpu-aarch64.c
-            ggml-cpu-aarch64.h
-            ggml-cpu-quants.c
-            ggml-cpu-quants.h
-            )
+ggml_add_backend_library(ggml-cpu
+                         ggml-cpu.c
+                         ggml-cpu.cpp
+                         ggml-cpu-aarch64.c
+                         ggml-cpu-aarch64.h
+                         ggml-cpu-quants.c
+                         ggml-cpu-quants.h
+                        )
 
-target_link_libraries(ggml-cpu PRIVATE ggml-base)
-target_include_directories(ggml-cpu PRIVATE . ..)
+target_include_directories(ggml-cpu PRIVATE .)
 
 if (APPLE AND GGML_ACCELERATE)
     find_library(ACCELERATE_FRAMEWORK Accelerate)
@@ -244,6 +243,11 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
     if (GGML_LSX)
         list(APPEND ARCH_FLAGS -mlsx)
     endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+    message(STATUS "RISC-V detected")
+    if (GGML_RVV)
+        list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+    endif()
 else()
     message(STATUS "Unknown architecture")
 endif()

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2369,7 +2369,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     // figure out which node we're on
     uint current_cpu;
     int getcpu_ret = 0;
-#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall
@@ -13578,29 +13578,6 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 
 #endif // GGML_USE_OPENMP
 
-void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
-    p->n_threads  = n_threads;
-    p->prio       = 0;     // default priority (usually means normal or inherited)
-    p->poll       = 50;    // hybrid-polling enabled
-    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
-    p->paused     = false; // threads are ready to go
-    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
-}
-
-struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
-    struct ggml_threadpool_params p;
-    ggml_threadpool_params_init(&p, n_threads);
-    return p;
-}
-
-bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
-    if (p0->n_threads      != p1->n_threads  )    return false;
-    if (p0->prio           != p1->prio       )    return false;
-    if (p0->poll           != p1->poll       )    return false;
-    if (p0->strict_cpu     != p1->strict_cpu )    return false;
-    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
-}
-
 static struct ggml_threadpool * ggml_threadpool_new_impl(
     struct ggml_threadpool_params * tpp,
                struct ggml_cgraph * cgraph,
@@ -13896,7 +13873,7 @@ int ggml_cpu_has_vsx(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return ggml_arm_arch_features.has_neon;
 #else
     return 0;
@@ -13904,7 +13881,7 @@ int ggml_cpu_has_neon(void) {
 }
 
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.has_sve;
 #else
     return 0;
@@ -13912,7 +13889,7 @@ int ggml_cpu_has_sve(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
     return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
@@ -13920,7 +13897,7 @@ int ggml_cpu_has_matmul_int8(void) {
 }
 
 int ggml_cpu_get_sve_cnt(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.sve_cnt;
 #else
     return 0;

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -541,16 +541,12 @@ static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg
     return &ggml_backend_cpu_device;
 }
 
-struct ggml_backend_feature {
-    const char * name;
-    const char * value;
-};
-
-// Not used yet
 // This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
-// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
+// and additionally to allow other backends to expose their own list of features that applications can query using the same API
 static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
     static std::vector<ggml_backend_feature> features = []() {
+        ggml_cpu_init();
+
         std::vector<ggml_backend_feature> features;
         if (ggml_cpu_has_sse3()) {
             features.push_back({ "SSE3", "1" });
@@ -561,6 +557,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_avx()) {
             features.push_back({ "AVX", "1" });
         }
+        if (ggml_cpu_has_avx_vnni()) {
+            features.push_back({ "AVX_VNNI", "1" });
+        }
         if (ggml_cpu_has_avx2()) {
             features.push_back({ "AVX2", "1" });
         }
@@ -570,9 +569,6 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_fma()) {
             features.push_back({ "FMA", "1" });
         }
-        if (ggml_cpu_has_avx_vnni()) {
-            features.push_back({ "AVX_VNNI", "1" });
-        }
         if (ggml_cpu_has_avx512()) {
             features.push_back({ "AVX512", "1" });
         }
@@ -619,6 +615,10 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_llamafile()) {
             features.push_back({ "LLAMAFILE", "1" });
         }
+        // TODO: rename this
+    #ifdef GGML_USE_CPU_AARCH64
+        features.push_back({ "AARCH64_REPACK", "1" });
+    #endif
 
         features.push_back({ nullptr, nullptr });
 
@@ -637,6 +637,29 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
     if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
         return (void *)ggml_backend_cpu_get_extra_bufts;
     }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cpu_get_features;
+    }
+    if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
+        return (void *)ggml_backend_cpu_set_abort_callback;
+    }
+    if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
+        return (void *)ggml_numa_init;
+    }
+    if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
+        return (void *)ggml_is_numa;
+    }
+
+    // threadpool - TODO:  move to ggml-base
+    if (strcmp(name, "ggml_threadpool_new") == 0) {
+        return (void *)ggml_threadpool_new;
+    }
+    if (strcmp(name, "ggml_threadpool_free") == 0) {
+        return (void *)ggml_threadpool_free;
+    }
+    if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
+        return (void *)ggml_backend_cpu_set_threadpool;
+    }
 
     return NULL;
 
@@ -655,9 +678,12 @@ ggml_backend_reg_t ggml_backend_cpu_reg(void) {
     ggml_cpu_init();
 
     static struct ggml_backend_reg ggml_backend_cpu_reg = {
-        /* .iface   = */ ggml_backend_cpu_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_cpu_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_cpu_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cpu_reg)

ggml/src/ggml-cuda/CMakeLists.txt

@@ -6,15 +6,18 @@ if (CUDAToolkit_FOUND)
     message(STATUS "CUDA Toolkit found")
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        # 52 == lowest CUDA 12 standard
-        # 60 == FP16 CUDA intrinsics
-        # 61 == integer CUDA intrinsics
-        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
-            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -43,29 +46,17 @@ if (CUDAToolkit_FOUND)
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
     endif()
 
-    add_library(ggml-cuda
-                ${GGML_HEADERS_CUDA}
-                ${GGML_SOURCES_CUDA}
-                )
+    ggml_add_backend_library(ggml-cuda
+                             ${GGML_HEADERS_CUDA}
+                             ${GGML_SOURCES_CUDA}
+                            )
 
-    target_link_libraries(ggml-cuda PRIVATE ggml-base)
-    target_include_directories(ggml-cuda PRIVATE . ..)
-
-    # TODO: change the definitions to this target only
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
     if (GGML_CUDA_GRAPHS)
         add_compile_definitions(GGML_CUDA_USE_GRAPHS)
     endif()
 
-    if (GGML_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
@@ -78,10 +69,6 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_VMM)
     endif()
 
-    if (DEFINED GGML_CUDA_DMMV_Y)
-        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
-    endif()
-
     if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
         add_compile_definitions(GGML_CUDA_F16)
     endif()
@@ -159,7 +146,7 @@ if (CUDAToolkit_FOUND)
         list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
     endif()
 
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
 else()
     message(FATAL_ERROR "CUDA Toolkit not found")
 endif()

ggml/src/ggml-cuda/argmax.cu

@@ -1,57 +1,69 @@
-#include "common.cuh"
-#include "argmax.cuh"
-#include "sum.cuh"
-
+#include <algorithm>
 #include <cstdint>
 
-static __global__ void argmax_f32(
-    const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) {
+#include "argmax.cuh"
+#include "common.cuh"
+#include "sum.cuh"
 
-    int argmax_thread = 0;
-    const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE;
+static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) {
+    const int64_t row = blockIdx.x;
 
-#pragma unroll
-    for (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) {
-        const int64_t row = row0 + row1;
+    float maxval = -FLT_MAX;
+    int   argmax = -1;
+    const float * rowx = x + row * ncols;
 
-        if (row >= nrows) {
-            break;
+    for (int32_t col = threadIdx.x; col < ncols; col += blockDim.x) {
+        const float val = rowx[col];
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
         }
-
-        float maxval = -FLT_MAX;
-        int   argmax = -1;
-
-        for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) {
-            const float val        = x[row*ncols + col];
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
-        }
-
-#pragma unroll
-        for (int mask = 16; mask > 0; mask >>= 1) {
-            const float val        = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE);
-            const int   col        = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE);
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
-        }
-
-        const int store = row1 == threadIdx.x;
-        argmax_thread += store*argmax;
     }
 
-    const int row = row0 + threadIdx.x;
-
-    if (row >= nrows) {
-        return;
+#pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+        const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
+        }
     }
 
-    dst[row] = argmax_thread;
+    const int n_warps = blockDim.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    if (n_warps > 1) {
+        constexpr int    max_warps = 1024 / WARP_SIZE;
+        __shared__ float shared_maxval[max_warps];
+        __shared__ int   shared_argmax[max_warps];
+        if (lane_id == 0) {
+            shared_maxval[warp_id] = maxval;
+            shared_argmax[warp_id] = argmax;
+        }
+
+        __syncthreads();
+
+        if (warp_id == 0) {
+            if (lane_id < n_warps) {
+                maxval = shared_maxval[lane_id];
+                argmax = shared_argmax[lane_id];
+            }
+#pragma unroll
+            for (int offset = 16; offset > 0; offset >>= 1) {
+                const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+                const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+                if (val > maxval) {
+                    maxval = val;
+                    argmax = col;
+                }
+            }
+        }
+    }
+
+    if (warp_id == 0 && lane_id == 0) {
+        dst[row] = argmax;
+    }
 }
 
 void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -70,10 +82,10 @@ void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE;
-
-    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const int64_t num_blocks = nrows;
+    const int64_t num_threads = std::min<int64_t>(1024, (ne00 + WARP_SIZE - 1) / WARP_SIZE * WARP_SIZE);
+    const dim3 blocks_dim(num_threads, 1, 1);
     const dim3 blocks_num(num_blocks, 1, 1);
 
-    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00, nrows);
+    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00);
 }

ggml/src/ggml-cuda/common.cuh

@@ -180,8 +180,8 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
@@ -189,17 +189,17 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
 }
 
 static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, offset, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, offset, 32);
     }
     return a;
 }
@@ -209,16 +209,16 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, offset, 32);
         reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
         reinterpret_cast<half&>(a.y) += __high2half(a_other);
     }
     return a;
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, offset, 32));
     }
     return a;
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -231,8 +231,8 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
     }
     return x;
 }
@@ -275,8 +275,8 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   for (int offset = 16; offset > 0; offset >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
    }
    return x;
 #else

ggml/src/ggml-cuda/dmmv.cu

@@ -1,699 +0,0 @@
-#include "dmmv.cuh"
-#include "dequantize.cuh"
-#include "convert.cuh"
-
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
-static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q2_K * x = (const block_q2_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
-    const int q_offset = 32*im + l0;
-    const int s_offset = 8*im;
-    const int y_offset = 128*im + l0;
-
-    uint32_t aux[4];
-    const uint8_t * d = (const uint8_t *)aux;
-    const uint8_t * m = (const uint8_t *)(aux + 2);
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = a[1] & 0x0f0f0f0f;
-        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
-        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
-                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
-                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
-                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
-                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
-                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
-                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
-                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
-            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
-                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
-
-        }
-        tmp += dall * sum1 - dmin * sum2;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q3_K * x = (const block_q3_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
-    const int step = 16/K_QUANTS_PER_ITERATION;
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0....15 or 0...7
-
-    const uint8_t m = 1 << (4*im);
-
-    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
-    const int q_offset =  32*im + l0;
-    const int y_offset = 128*im + l0;
-
-    uint16_t utmp[4];
-    const int8_t * s = (const int8_t *)utmp;
-
-    const uint16_t s_shift = 4*im;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-        const uint8_t * h = x[i].hmask + l0;
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
-        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
-
-        const float d = x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < n; ++l) {
-            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
-                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
-                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
-                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
-            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
-                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
-                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
-                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
-        }
-        tmp += d * sum;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q4_K * x = (const block_q4_K *)vx + ib0;
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
-
-    const int il  = tid/step;                            // 0...3
-    const int ir  = tid - step*il;                       // 0...7 or 0...3
-    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-#if K_QUANTS_PER_ITERATION == 2
-    uint32_t q32[4];
-    const uint8_t * q4 = (const uint8_t *)q32;
-#else
-    uint16_t q16[4];
-    const uint8_t * q4 = (const uint8_t *)q16;
-#endif
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y1 = yy + i*QK_K + y_offset;
-        const float   * y2 = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-#if K_QUANTS_PER_ITERATION == 2
-        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
-        const uint32_t * q2 = q1 + 16;
-
-        q32[0] = q1[0] & 0x0f0f0f0f;
-        q32[1] = q1[0] & 0xf0f0f0f0;
-        q32[2] = q2[0] & 0x0f0f0f0f;
-        q32[3] = q2[0] & 0xf0f0f0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 4; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
-            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#else
-        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
-        const uint16_t * q2 = q1 + 32;
-
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[0] & 0xf0f0;
-        q16[2] = q2[0] & 0x0f0f;
-        q16[3] = q2[0] & 0xf0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 2; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
-
-    const int row = blockIdx.x;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q5_K * x = (const block_q5_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;
-
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 2;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    const uint8_t hm1  = 1 << (2*im);
-    const uint8_t hm2  = hm1 << 4;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    uint16_t q16[8];
-    const uint8_t * q4 = (const uint8_t *)q16;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
-
-        const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * qh  = x[i].qh + l0;
-        const float   * y1  = yy + i*QK_K + y_offset;
-        const float   * y2  = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 sum = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        const uint16_t * q1 = (const uint16_t *)ql1;
-        const uint16_t * q2 = q1 + 32;
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[8] & 0x0f0f;
-        q16[2] = (q1[0] >> 4) & 0x0f0f;
-        q16[3] = (q1[8] >> 4) & 0x0f0f;
-        q16[4] = q2[0] & 0x0f0f;
-        q16[5] = q2[8] & 0x0f0f;
-        q16[6] = (q2[0] >> 4) & 0x0f0f;
-        q16[7] = (q2[8] >> 4) & 0x0f0f;
-        for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
-            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
-                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
-        }
-        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q6_K * x = (const block_q6_K *)vx + ib0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-#if K_QUANTS_PER_ITERATION == 1
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
-    const int is = 0;
-#else
-    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
-    const int is = in / 4;
-#endif
-    const int ql_offset = 64*im + l0;
-    const int qh_offset = 32*im + l0;
-    const int s_offset  =  8*im + is;
-    const int y_offset = 128*im + l0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * ql = x[i].ql + ql_offset;
-        const uint8_t * qh = x[i].qh + qh_offset;
-        const int8_t  * s  = x[i].scales + s_offset;
-
-        const float d = x[i].d;
-
-#if K_QUANTS_PER_ITERATION == 1
-        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
-                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
-                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
-                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
-                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
-                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
-                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
-                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
-        tmp += sum;
-#else
-        float sum = 0;
-        for (int l = 0; l < 4; ++l) {
-            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
-                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
-                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
-                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
-        }
-        tmp += sum;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
-    const half * x = (const half *) vx;
-    // load 2 halfs into register in a single instruction
-    const half2 x_reg = *((half2 *) &(x[ib + iqs]));
-    // automatic half -> float type cast if dfloat == float
-    v.x = __low2float(x_reg);
-    v.y = __high2float(x_reg);
-}
-
-static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
-        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
-        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
-        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
-        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
-        type == GGML_TYPE_F16 ? convert_f16 :
-        nullptr;
-}
-
-template <ggml_type type>
-static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
-    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
-    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
-
-    const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int tid = threadIdx.x;
-
-    const int iter_stride = 2*GGML_CUDA_DMMV_X;
-    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-// partial sum for each thread
-#ifdef GGML_CUDA_F16
-    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
-#else
-    float tmp = 0.0f;
-#endif // GGML_CUDA_F16
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
-        const int iqs = (col%qk)/qr; // x quant index
-        const int iybs = col - col%qk; // y block start index
-
-// processing >2 values per i iter is faster for fast GPUs
-#pragma unroll
-        for (int j = 0; j < vals_per_iter; j += 2) {
-            // process 2 vals per j iter
-
-            // dequantize
-            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
-            dfloat2 v;
-            dequantize_kernel(vx, ib, iqs + j/qr, v);
-
-            // matrix multiplication
-            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
-#ifdef GGML_CUDA_F16
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += __hmul2(v, y_reg);
-            }
-            else {
-                tmp += __hmul2(v, {
-                        y[iybs + iqs + j/qr + 0],
-                        y[iybs + iqs + j/qr + y_offset]
-                    });
-            }
-#else
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += v.x * y_reg.x;
-                tmp += v.y * y_reg.y;
-            }
-            else {
-                tmp += v.x * y[iybs + iqs + j/qr + 0];
-                tmp += v.y * y[iybs + iqs + j/qr + y_offset];
-            }
-#endif // GGML_CUDA_F16
-        }
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-#ifdef GGML_CUDA_F16
-        dst[row] = tmp.x + tmp.y;
-#else
-        dst[row] = tmp;
-#endif // GGML_CUDA_F16
-    }
-}
-
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
-}
-
-static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_F16>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-    GGML_UNUSED(ctx);
-    const int64_t ne00 = src0->ne[0];
-    const int64_t row_diff = row_high - row_low;
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_F16
-    ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
-    half * src1_dfloat = nullptr; // dfloat == half
-
-    bool src1_convert_f16 =
-        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-    if (src1_convert_f16) {
-        src1_dfloat = src1_dfloat_a.alloc(ne00);
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
-    }
-#else
-    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
-#endif // GGML_CUDA_F16
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
-
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
-}
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
-    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
-        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
-        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
-        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
-        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
-        src0_type == GGML_TYPE_F16;
-}

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -16,11 +16,11 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
-#include "ggml-cuda/dmmv.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
+#include "ggml-cuda/mmv.cuh"
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
@@ -1020,114 +1020,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 
 #define MUL_MAT_SRC1_COL_STRIDE 128
 
-static __global__ void mul_mat_p021_f16_f32(
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / (nchannels_y / nchannels_x);
-
-    const int nrows_y = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
-        const float xi = __half2float(x[ix]);
-
-        const int row_y = col_x;
-
-        // y is not transposed but permuted
-        const int iy = channel*nrows_y + row_y;
-
-        tmp += xi * y[iy];
-    }
-
-    // dst is not transposed and not permuted
-    const int idst = channel*nrows_dst + row_dst;
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / channel_x_divisor;
-
-    const int nrows_y   = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst   = row_x;
-
-    const int idst = channel*nrows_dst + row_dst;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        const int row_y = col_x;
-
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
-
-        const float xi = __half2float(x[ix]);
-
-        tmp += xi * y[iy];
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static void ggml_mul_mat_p021_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
-}
-
-static void ggml_mul_mat_vec_nc_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
-}
-
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
@@ -1654,58 +1546,6 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
-static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
-    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
-}
-
-static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    const int64_t row_stride_x = nb01 / sizeof(half);
-    const int64_t channel_stride_x = nb02 / sizeof(half);
-
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
-}
-
 static __global__ void k_compute_batched_ptrs(
         const half * src0_as_f16, const half * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
@@ -1879,21 +1719,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+    bool use_mul_mat_vec   = src0->type == GGML_TYPE_F16
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
-    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    // if mmvq is available it's a better choice than dmmv:
-#ifndef GGML_CUDA_FORCE_DMMV
-    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
-#endif // GGML_CUDA_FORCE_DMMV
-
-    bool any_gpus_with_slow_fp16 = false;
+    bool any_gpus_with_slow_fp16   = false;
+    bool any_gpus_without_fp16_mma = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1904,14 +1740,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc            = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+            const int cc              = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
         }
     } else {
-        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
     }
 
     // debug helpers
@@ -1922,18 +1760,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
-    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    if (!split && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
+        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
                && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch without FlashAttention
+        // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
@@ -3290,6 +3126,61 @@ static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t re
     return ctx->devices[index];
 }
 
+static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        std::vector<ggml_backend_feature> features;
+    #define _STRINGIFY(...) #__VA_ARGS__
+    #define STRINGIFY(...) _STRINGIFY(__VA_ARGS__)
+
+    #ifdef __CUDA_ARCH_LIST__
+        features.push_back({ "ARCHS", STRINGIFY(__CUDA_ARCH_LIST__) });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_MMQ
+        features.push_back({ "FORCE_MMQ", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_CUBLAS
+        features.push_back({ "FORCE_CUBLAS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_VMM
+        features.push_back({ "NO_VMM", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_PEER_COPY
+        features.push_back({ "NO_PEER_COPY", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_F16
+        features.push_back({ "F16", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_USE_GRAPHS
+        features.push_back({ "USE_GRAPHS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+        features.push_back({ "PEER_MAX_BATCH_SIZE", STRINGIFY(GGML_CUDA_PEER_MAX_BATCH_SIZE) });
+    #endif
+
+    #ifdef GGML_CUDA_FA_ALL_QUANTS
+        features.push_back({ "FA_ALL_QUANTS", "1" });
+    #endif
+
+    #undef _STRINGIFY
+    #undef STRINGIFY
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
@@ -3301,6 +3192,9 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
     if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
         return (void *)ggml_backend_cuda_unregister_host_buffer;
     }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cuda_get_features;
+    }
     return nullptr;
 }
 
@@ -3333,16 +3227,17 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->description = prop.name;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cuda_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cuda_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cuda_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cuda_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -3373,3 +3268,5 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
 
     return cuda_backend;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg)

ggml/src/ggml-cuda/mmv.cu

@@ -0,0 +1,223 @@
+#include "common.cuh"
+#include "mmv.cuh"
+
+template <typename type_acc, int block_size>
+static __global__ void mul_mat_vec(
+        const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
+    const int64_t row     = blockIdx.x;
+    const int64_t channel = blockIdx.z;
+    const int     tid     = threadIdx.x;
+
+    x   += (channel/channel_ratio)*stride_channel_x + row*stride_row;
+    y   +=  channel               *stride_channel_y;
+    dst +=  channel               *stride_channel_dst;
+
+    const half2  * x2 = (const half2  *) x;
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+    float * buf_iw = (float *) data_mmv;
+
+    if (block_size > WARP_SIZE) {
+        if (tid < WARP_SIZE) {
+            buf_iw[tid] = 0.0f;
+        }
+        __syncthreads();
+    }
+
+    float sumf;
+
+    if (std::is_same<type_acc, float>::value) {
+        sumf = 0.0f;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = __half22float2(x2[col2]);
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x * tmpy.x;
+            sumf += tmpx.y * tmpy.y;
+        }
+    } else {
+#ifdef FP16_AVAILABLE
+        half2 sumh2 = make_half2(0.0f, 0.0f);
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmp = y2[col2];
+            sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+        }
+
+        sumf = __low2float(sumh2) + __high2float(sumh2);
+#else
+        NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+    }
+
+    sumf = warp_reduce_sum(sumf);
+
+    if (block_size > WARP_SIZE) {
+        buf_iw[tid/WARP_SIZE] = sumf;
+        __syncthreads();
+        if (tid > WARP_SIZE) {
+            return;
+        }
+        sumf = buf_iw[tid];
+        sumf = warp_reduce_sum(sumf);
+    }
+
+    if (tid != 0) {
+        return;
+    }
+
+    dst[row] = sumf;
+}
+
+template <typename type_acc>
+static void launch_mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        cudaStream_t stream) {
+    GGML_ASSERT(ncols      % 2 == 0);
+    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(nchannels_y % nchannels_x == 0);
+    const int64_t channel_ratio = nchannels_y / nchannels_x;
+
+    int64_t block_size_best = WARP_SIZE;
+    int64_t niter_best      = (ncols + 2*WARP_SIZE - 1) / (2*WARP_SIZE);
+    for (int64_t block_size = 2*WARP_SIZE; block_size <= 256; block_size += WARP_SIZE) {
+        const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
+        if (niter < niter_best) {
+            niter_best      = niter;
+            block_size_best = block_size;
+        }
+    }
+
+    const int smem = WARP_SIZE*sizeof(float);
+    const dim3 block_nums(nrows, 1, nchannels_y);
+    const dim3 block_dims(block_size_best, 1, 1);
+    switch (block_size_best) {
+        case   32: {
+            mul_mat_vec<type_acc,  32><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   64: {
+            mul_mat_vec<type_acc,  64><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   96: {
+            mul_mat_vec<type_acc,  96><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  128: {
+            mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  160: {
+            mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  192: {
+            mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  224: {
+            mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  256: {
+            mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+static void mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        enum ggml_prec prec, cudaStream_t stream) {
+    switch (prec) {
+        case GGML_PREC_DEFAULT: {
+            launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+        case GGML_PREC_F32: {
+            launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+    }
+}
+
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+
+    GGML_ASSERT(src1->ne[1] == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const half  * src0_d = (const half  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne12 = src1->ne[2];
+    GGML_ASSERT(dst->ne[2] == ne12);
+
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT( dst->ne[3] == 1);
+
+    const int64_t stride_row         = src0->nb[1] / ggml_type_size(src0->type);
+    const int64_t channel_stride_x   = src0->nb[2] / ggml_type_size(src0->type);
+    const int64_t channel_stride_y   = src1->nb[2] / ggml_type_size(src1->type);
+    const int64_t channel_stride_dst =  dst->nb[2] / ggml_type_size( dst->type);
+
+    mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
+}
+
+void ggml_cuda_op_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1_ncols == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+
+    // ggml_cuda_op provides single, contiguous matrices
+    const int64_t stride_row         = ne00;
+    const int64_t nchannels_x        = 1;
+    const int64_t nchannels_y        = 1;
+    const int64_t channel_stride_x   = 0;
+    const int64_t channel_stride_y   = 0;
+    const int64_t channel_stride_dst = 0;
+
+    mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
+        nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}

ggml/src/ggml-cuda/mmv.cuh

@@ -1,20 +1,12 @@
 #include "common.cuh"
 
-// dmmv = dequantize_mul_mat_vec
+// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
+#define MMV_MAX_ROWS 512
 
-// TODO: remove this?
-#ifndef GGML_CUDA_DMMV_X
-#define GGML_CUDA_DMMV_X 32
-#endif
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
-#ifndef GGML_CUDA_MMV_Y
-#define GGML_CUDA_MMV_Y 1
-#endif
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
+void ggml_cuda_op_mul_mat_vec(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);

ggml/src/ggml-cuda/quantize.cu

@@ -69,8 +69,8 @@ static __global__ void quantize_mmq_q8_1(
 
     // Exchange max. abs. value between vals_per_scale/4 threads.
 #pragma unroll
-    for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) {
-        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+    for (int offset = vals_per_scale/8; offset > 0; offset >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, offset, WARP_SIZE));
     }
 
     float sum;
@@ -79,8 +79,8 @@ static __global__ void quantize_mmq_q8_1(
 
         // Exchange calculate sum across vals_per_sum/4 threads.
 #pragma unroll
-        for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) {
-            sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE);
+        for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
+            sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
         }
     }