server : fix parallel speculative decoding (#10513)

ggml-ci

.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:

CMakeLists.txt

@@ -163,8 +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,
@@ -966,6 +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
 
@@ -1165,6 +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

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 << " ]";
@@ -925,9 +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) {
@@ -979,9 +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;
     }
@@ -1490,6 +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 +1,270 @@
+#include "speculative.h"
+
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+
+#include <cstring>
+
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
+#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
+
+struct common_speculative {
+    struct llama_context * ctx;
+    struct common_sampler * smpl;
+
+    llama_batch batch;
+    llama_tokens prompt;
+};
+
+struct common_speculative * common_speculative_init(
+        struct llama_context * ctx_dft) {
+    auto * result = new common_speculative {
+        /* .ctx    = */ ctx_dft,
+        /* .smpl   = */ nullptr,
+        /* .batch  = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .prompt = */ {},
+    };
+
+    // TODO: optimize or pass from outside?
+#if 0
+    {
+        common_params_sampling params;
+        params.no_perf = false;
+
+        params.top_k = 40;
+        params.top_p = 0.9;
+
+        params.samplers = {
+            COMMON_SAMPLER_TYPE_TOP_K,
+            COMMON_SAMPLER_TYPE_TOP_P,
+            COMMON_SAMPLER_TYPE_INFILL,
+        };
+
+        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+    }
+#else
+    {
+        common_params_sampling params;
+        params.no_perf = false;
+
+        params.top_k = 10;
+
+        params.samplers = {
+            COMMON_SAMPLER_TYPE_TOP_K,
+        };
+
+        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+    }
+#endif
+
+    return result;
+}
+
+void common_speculative_free(struct common_speculative * spec) {
+    common_sampler_free(spec->smpl);
+
+    llama_batch_free(spec->batch);
+
+    delete spec;
+}
+
+bool common_speculative_are_compatible(
+        const struct llama_context * ctx_tgt,
+        const struct llama_context * ctx_dft) {
+    const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
+    const struct llama_model * model_dft = llama_get_model(ctx_dft);
+
+    const bool vocab_type_tgt = llama_vocab_type(model_tgt);
+    LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
+
+    const bool vocab_type_dft = llama_vocab_type(model_dft);
+    LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
+
+    if (vocab_type_tgt != vocab_type_dft) {
+        LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
+                     "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
+        return false;
+    }
+
+    if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
+        llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
+        llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
+        LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
+        return false;
+    }
+
+    {
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
+
+        const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
+
+        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
+            LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
+                         "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    __func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            return false;
+        }
+
+        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
+            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
+                LOG_ERR("%s: draft model vocab must match target model to use speculation but "
+                             "token %d content differs - target '%s', draft '%s'\n", __func__, i,
+                        common_token_to_piece(ctx_tgt, i).c_str(),
+                        common_token_to_piece(ctx_dft, i).c_str());
+                return false;
+            }
+        }
+    }
+
+    return true;
+}
+
+llama_tokens common_speculative_gen_draft(
+        struct common_speculative * spec,
+        struct common_speculative_params params,
+        const llama_tokens & prompt_tgt,
+        llama_token id_last) {
+    auto & batch  = spec->batch;
+    auto & ctx    = spec->ctx;
+    auto & smpl   = spec->smpl;
+    auto & prompt = spec->prompt;
+
+    int reuse_i = 0;
+    int reuse_n = 0;
+
+    const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
+
+    const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
+
+    // reuse as much as possible from the old draft context
+    // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
+    for (int i = 0; i < (int) prompt.size(); ++i) {
+        int cur = 0;
+        while (i_start + cur < (int) prompt_tgt.size() &&
+               i       + cur < (int) prompt.size() &&
+               prompt_tgt[i_start + cur] == prompt[i + cur]) {
+            cur++;
+        }
+
+        if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) {
+            reuse_i = i;
+            reuse_n = cur;
+        }
+    }
+
+    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
+
+    llama_tokens result;
+    result.reserve(params.n_draft);
+
+    if (reuse_n == 0) {
+        llama_kv_cache_clear(ctx);
+
+        prompt.clear();
+    } else {
+        // this happens when a previous draft has been discarded (for example, due to being too small), but the
+        // target model agreed with it. in this case, we simply pass back the previous results to save compute
+        if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
+            for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
+                result.push_back(prompt[i]);
+
+                if (params.n_draft <= (int) result.size()) {
+                    break;
+                }
+            }
+
+            return result;
+        }
+
+        if (reuse_i > 0) {
+            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+
+            prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
+        }
+
+        if (reuse_n < (int) prompt.size()) {
+            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+
+            prompt.erase(prompt.begin() + reuse_n, prompt.end());
+        }
+    }
+
+    // prepare a batch to evaluate any new tokens in the prompt
+    common_batch_clear(batch);
+
+    for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
+        //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
+        common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
+
+        prompt.push_back(prompt_tgt[i]);
+    }
+
+    // we should rarely end-up here during normal decoding
+    if (batch.n_tokens > 0) {
+        //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
+
+        llama_decode(ctx, batch);
+    }
+
+    const llama_pos n_past = prompt.size();
+
+    LOG_DBG("%s: n_past = %d\n", __func__, n_past);
+
+    common_batch_clear(batch);
+    common_batch_add  (batch, id_last, n_past, { 0 }, true);
+
+    prompt.push_back(id_last);
+
+    //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
+
+    llama_decode(ctx, batch);
+
+    common_sampler_reset(smpl);
+
+    // sample n_draft tokens from the draft model
+    for (int i = 0; i < params.n_draft; ++i) {
+        common_batch_clear(batch);
+
+        common_sampler_sample(smpl, ctx, 0, true);
+
+        const auto * cur_p = common_sampler_get_candidates(smpl);
+
+        for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+            LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
+        }
+
+        // add drafted token for each sequence
+        const llama_token id = cur_p->data[0].id;
+
+        // only collect very high-confidence draft tokens
+        if (cur_p->data[0].p < params.p_min) {
+            break;
+        }
+
+        common_sampler_accept(smpl, id, true);
+
+        result.push_back(id);
+
+        if (params.n_draft <= (int) result.size()) {
+            break;
+        }
+
+        common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
+
+        // evaluate the drafted tokens on the draft model
+        llama_decode(ctx, batch);
+
+        prompt.push_back(id);
+    }
+
+    return result;
+}

common/speculative.h

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

convert_hf_to_gguf.py

@@ -2707,7 +2707,7 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_token_scores(scores)
         self.gguf_writer.add_token_types(toktypes)
         self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
         self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
         if precompiled_charsmap:
             self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
@@ -3040,9 +3040,9 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("Olmo1124ForCausalLM")
-class Olmo1124Model(Model):
-    model_arch = gguf.MODEL_ARCH.OLMO_1124
+@Model.register("Olmo2ForCausalLM")
+class Olmo2Model(Model):
+    model_arch = gguf.MODEL_ARCH.OLMO2
 
 
 @Model.register("OlmoeForCausalLM")

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/llama-bench/llama-bench.cpp

@@ -1477,6 +1477,17 @@ int main(int argc, char ** argv) {
 
     cmd_params params = parse_cmd_params(argc, argv);
 
+    // initialize backends
+    ggml_backend_load_all();
+    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (!cpu_dev) {
+        fprintf(stderr, "%s: error: CPU backend is not loaded\n", __func__);
+        return 1;
+    }
+    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_threadpool_new");
+    auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_threadpool_free");
+
     // initialize llama.cpp
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
@@ -1551,7 +1562,7 @@ int main(int argc, char ** argv) {
         tpp.poll       = t.poll;
         tpp.prio       = params.prio;
 
-        struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+        struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
         if (!threadpool) {
             fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
             exit(1);
@@ -1612,7 +1623,7 @@ int main(int argc, char ** argv) {
 
         llama_free(ctx);
 
-        ggml_threadpool_free(threadpool);
+        ggml_threadpool_free_fn(threadpool);
     }
 
     llama_free_model(lmodel);

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());
@@ -2168,8 +2267,9 @@ 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_sample(slot.smpl, ctx, slot.i_batch - i);
+
+                slot.i_batch = -1;
 
                 common_sampler_accept(slot.smpl, id, true);
 
@@ -2180,14 +2280,15 @@ struct server_context {
                     metrics.on_prompt_eval(slot);
                 }
 
+                completion_token_output result;
                 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) {
+                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,
+                            i >= cur_p->size ? 0.0f : cur_p->data[i].p,
                     });
                 }
 
@@ -2197,9 +2298,67 @@ struct server_context {
                     slot.print_timings();
                     send_final_response(slot);
                     metrics.on_prediction(slot);
+                    continue;
+                }
+            }
+
+            // do speculative decoding
+            for (auto & slot : slots) {
+                if (!slot.is_processing() || !slot.can_speculate()) {
+                    continue;
                 }
 
-                slot.i_batch = -1;
+                llama_token id = slot.sampled;
+
+                struct common_speculative_params params_spec;
+                params_spec.n_draft   = slot.params.speculative.n_max;
+                params_spec.n_reuse   = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max;
+                params_spec.p_min     = slot.params.speculative.p_min;
+
+                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+
+                // ignore small drafts
+                if (slot.params.speculative.n_min > (int) draft.size()) {
+                    continue;
+                }
+
+                // 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 +2856,7 @@ 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 +2864,7 @@ 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 +2877,7 @@ 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 +2983,7 @@ 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 +3160,7 @@ 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 +1,265 @@
+#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
+    llama_tokens prompt_tgt(inp.begin(), inp.end() - 1);
+    prompt_tgt.reserve(llama_n_ctx(ctx_tgt));
+
+    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 += draft.size(); // note: we ignore the discarded small drafts
+        n_accept  += ids.size() - 1;
+        n_predict += ids.size();
+
+        // 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
+        //
+        for (size_t i = 0; i < ids.size(); ++i) {
+            prompt_tgt.push_back(id_last);
+
+            id_last = ids[i];
+
+            if (llama_token_is_eog(model_tgt, id_last)) {
+                has_eos = true;
+                break;
+            }
+
+            const std::string token_str = common_token_to_piece(ctx_tgt, id_last);
+
+            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());
+            }
+        }
+
+        LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d)\n", (int) ids.size() - 1, (int) draft.size(), id_last);
+
+        {
+            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);
+        }
+
+        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
+            break;
+        }
+    }
+
+    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": 1731676054,
-        "narHash": "sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg=",
+        "lastModified": 1732014248,
+        "narHash": "sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5e4fbfb6b3de1aa2872b76d49fafc942626e2add",
+        "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

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)
+            if (NOT GGML_BACKEND_DL)
+                string(TOUPPER "GGML_USE_${backend}" backend_use)
+                target_compile_definitions(ggml PUBLIC ${backend_use})
             endif()
-            install(TARGETS ${backend_target} LIBRARY)
-            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-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)

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

@@ -61,9 +61,9 @@ 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}")

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -21,6 +21,7 @@
  */
 
 #include "aclnn_ops.h"
+#include "ggml-impl.h"
 
 #include <aclnnop/aclnn_avgpool2d.h>
 #include <aclnnop/aclnn_cast.h>
@@ -32,6 +33,8 @@
 #include <aclnnop/aclnn_group_norm.h>
 #include <aclnnop/aclnn_index_fill_tensor.h>
 #include <aclnnop/aclnn_layer_norm.h>
+#include <aclnnop/aclnn_mm.h>
+#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_matmul.h>
 #include <aclnnop/aclnn_max_pool.h>
 #include <aclnnop/aclnn_permute.h>
@@ -241,10 +244,14 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    int64_t concat_dim = 1;
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+    int32_t acl_dim = 3 - dim;
+
     aclTensor* tensors[] = {acl_src0, acl_src1};
     aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
-    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+    aclnn_concat(ctx, tensorList, acl_dst, acl_dim);
 
     ACL_CHECK(aclDestroyTensorList(tensorList));
     ACL_CHECK(aclDestroyTensor(acl_dst));
@@ -1437,10 +1444,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
@@ -1462,9 +1465,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
-
     // memory allocated increased to 3x when is_2D == false
     const int64_t n_bytes_factor = is_2D ? 1 : 3;
 
@@ -2425,7 +2425,6 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
                           aclTensor* acl_weight, aclTensor* acl_dst) {
     int8_t cube_math_type = 1;  // ALLOW_FP32_DOWN_PRECISION, when input is
                                 // fp32, atlas a2 will transpose it to HFLOAT32.
-
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -2443,6 +2442,80 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
         aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs matrix multiplication of two 2D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
+                             aclTensor* acl_weight, aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                      cube_math_type, &workspaceSize,
+                                      &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication of two 3D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
+                             aclTensor* acl_weight, aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                               cube_math_type, &workspaceSize,
+                                               &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
 /**
  * @brief Performs matrix multiplication with floating-point precision on
  * tensors using the CANN backend.
@@ -2464,20 +2537,43 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     // broadcast, when weight ne2 or ne3 is not 1, weight need repeat.
     BCAST_MUL_MAT_SHAPE(input, weight, dst);
 
-    // transpose weight: [1,2,3,4] -> [1,2,4,3]
-    int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
-                              bcast_weight_ne[2], bcast_weight_ne[3],
-                              bcast_weight_ne[4], bcast_weight_ne[5]};
-    size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
-                             bcast_weight_nb[2], bcast_weight_nb[3],
-                             bcast_weight_nb[4], bcast_weight_nb[5]};
+    int64_t n_dims = bcast_dims;
+    if (bcast_input_ne[3] == bcast_weight_ne[3] && bcast_input_ne[3] == 1) {
+        if (bcast_input_ne[2] == 1 && bcast_weight_ne[2] == 1) {
+            n_dims = 2;
+        } else if (bcast_input_ne[2] == 1) {
+            n_dims = 3;
+        }
+    }
 
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims);
     aclTensor* acl_input_tensor =
-        ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input));
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst));
-    aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
+    int64_t transpose_ne[] = {
+        bcast_weight_ne[1], bcast_weight_ne[0],
+        bcast_weight_ne[2], bcast_weight_ne[3],
+        bcast_weight_ne[4], bcast_weight_ne[5]
+    };
+    size_t transpose_nb[] = {
+        bcast_weight_nb[1], bcast_weight_nb[0],
+        bcast_weight_nb[2], bcast_weight_nb[3],
+        bcast_weight_nb[4], bcast_weight_nb[5]
+    };
+    aclTensor* acl_weight_tensor =
+        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
+
+    switch (n_dims) {
+    case 2:
+        aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    case 3:
+        aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    default:
+        aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    }
 
     ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
     ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@@ -2503,46 +2599,40 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
-    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
-    // is regarded as batch. weight need transpose.
-    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
+    // The shape of the weight is NCHW.
+    // Matrix multiplication uses HW dims.
+    // HC is regarded as batch.
+    // weight need transpose.
     float weight_elem_size;
     if (type == GGML_TYPE_Q4_0) {
         weight_elem_size = float(sizeof(uint8_t)) / 2;
-    }
-    else if (type == GGML_TYPE_Q8_0) {
+    } else if (type == GGML_TYPE_Q8_0) {
         weight_elem_size = float(sizeof(uint8_t));
-    }
-    else {
+    } else {
         GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
     }
-    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
-
-    // size of one matrix is element_size * height * width.
-    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
+    float weight_nb[] = {src0->ne[0] * weight_elem_size, weight_elem_size};
+    size_t weight_stride = src0->ne[1] * src0->ne[0] * weight_elem_size;
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
-    GGML_ASSERT(QK4_0 == QK8_0);
-    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
-    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
-                         scale_elem_size};
-    size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0;
+    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size, scale_elem_size};
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
     char* scale_offset = (char*)src0->data + weight_size;
 
     // input
-    void* input_buffer;
     size_t input_elem_size = sizeof(uint16_t);
     int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
-    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
-    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
-
+    size_t input_nb[] = {input_elem_size,  input_ne[0] * input_elem_size};
+    size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size;
     ggml_cann_pool_alloc input_alloctor(ctx.pool());
+    void* input_buffer = src1->data;
+
+    // case in
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
-        input_buffer = input_alloctor.get();
+        input_buffer = input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
 
         int64_t* input_cast_ne = src1->ne;
         size_t input_cast_nb[GGML_MAX_DIMS];
@@ -2552,88 +2642,139 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
         }
 
         aclTensor* acl_input_tensor = ggml_cann_create_tensor(
-            input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
-            input_cast_nb, GGML_MAX_DIMS);
+            input_buffer,
+            ACL_FLOAT16,
+            input_elem_size, input_cast_ne, input_cast_nb, GGML_MAX_DIMS);
         aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
+
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         ACL_CHECK(aclDestroyTensor(acl_src1_tensor));
-    } else {
-        input_buffer = src1->data;
     }
 
     // output
     size_t output_elem_size = sizeof(uint16_t);
-    int64_t output_ne[] = {dst->ne[0], dst->ne[1]};
-    size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]};
-    ggml_cann_pool_alloc output_alloctor(
-        ctx.pool(), ggml_nelements(dst) * output_elem_size);
-    void* output_buffer = output_alloctor.get();
-    size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1];
+    size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
+    ggml_cann_pool_alloc output_allocator(ctx.pool());
+    void* output_buffer = output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
+    size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
 
     // aclnn
+    int64_t max_elem_size = 65535;
+    int64_t split_size = (src0->ne[1] / max_elem_size) + 1;
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
+    aclOpExecutor* executor = nullptr;
     uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
-
     for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) {
         for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) {
             int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]);
             int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]);
 
-            int64_t batch1 = n1 * src1->ne[2] + c1;
-            int64_t batch0 = n0 * src0->ne[2] + c0;
+            int64_t batch1 = (n1 * src1->ne[2]) + c1;
+            int64_t batch0 = (n0 * src0->ne[2]) + c0;
 
             aclTensor* acl_input_tensor = ggml_cann_create_tensor(
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
+
+            // first split
+            int64_t weight_ne_offset = 0;
+            int64_t weight_ne[2] = {max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size, src0->ne[0]};
+            int64_t scale_ne_offset = 0;
+            int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
+            int64_t output_ne_offset = 0;
+            int64_t output_ne[2] = {weight_ne[0], dst->ne[1]};
+
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 (char*)src0->data + batch0 * weight_stride,
-                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
-                weight_nb, 2);
+                ggml_cann_type_mapping(type),
+                weight_elem_size, weight_ne, weight_nb, 2,
+                ACL_FORMAT_ND, weight_ne_offset);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
-                scale_offset + batch0 * scale_stride, ACL_FLOAT16,
-                scale_elem_size, scale_ne, scale_nb, 2);
+                scale_offset + batch0 * scale_stride,
+                ACL_FLOAT16,
+                scale_elem_size, scale_ne, scale_nb, 2,
+                ACL_FORMAT_ND, scale_ne_offset);
             aclTensor* acl_output_tensor = ggml_cann_create_tensor(
-                (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
-                output_elem_size, output_ne, output_nb, 2);
+                (char*)output_buffer + batch1 * output_stride,
+                ACL_FLOAT16,
+                output_elem_size, output_ne, output_nb, 2,
+                ACL_FORMAT_ND, output_ne_offset);
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
-                acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
-                nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
-                &workspaceSize, &executor));
-
-            if (workspaceSize > 0 && workspaceAddr == nullptr) {
-                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
-                                                         workspaceSize);
-                workspaceAddr = workspace_allocator.get();
+                acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                nullptr, nullptr, nullptr, nullptr, QK8_0,
+                acl_output_tensor, &workspaceSize, &executor));
+            if (workspaceAddr == nullptr) {
+                workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
-
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
                 workspaceAddr, workspaceSize, executor, ctx.stream()));
 
-            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
             ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
             ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
             ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+
+            // other splits
+            for (int64_t split = 1; split < split_size; split++) {
+                weight_ne_offset += weight_elem_size * weight_ne[0] * weight_ne[1];
+                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1] ? src0->ne[1] - (max_elem_size * split) : max_elem_size;
+                scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
+                scale_ne[0] = weight_ne[0];
+                output_ne_offset += output_elem_size * output_ne[0] * output_ne[1];
+                output_ne[0] = weight_ne[0];
+
+                acl_weight_tensor = ggml_cann_create_tensor(
+                    (char*)src0->data + batch0 * weight_stride,
+                    ggml_cann_type_mapping(type),
+                    weight_elem_size, weight_ne, weight_nb, 2,
+                    ACL_FORMAT_ND, weight_ne_offset);
+                acl_scale_tensor = ggml_cann_create_tensor(
+                    scale_offset + batch0 * scale_stride,
+                    ACL_FLOAT16,
+                    scale_elem_size, scale_ne, scale_nb, 2,
+                    ACL_FORMAT_ND, scale_ne_offset);
+                acl_output_tensor = ggml_cann_create_tensor(
+                    (char*)output_buffer + batch1 * output_stride,
+                    ACL_FLOAT16,
+                    output_elem_size, output_ne, output_nb, 2,
+                    ACL_FORMAT_ND, output_ne_offset);
+
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
+                    acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                    nullptr, nullptr, nullptr, nullptr, QK8_0,
+                    acl_output_tensor, &workspaceSize, &executor));
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
+                    workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+                ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+            }
+
+            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         }
     }
 
     // cast out
-    int64_t* output_cast_ne = dst->ne;
-    size_t output_cast_nb[GGML_MAX_DIMS];
-    output_cast_nb[0] = sizeof(uint16_t);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t* output_cast_ne = dst->ne;
+        size_t output_cast_nb[GGML_MAX_DIMS];
+        output_cast_nb[0] = sizeof(uint16_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+        }
+
+        aclTensor* acl_output_tensor = ggml_cann_create_tensor(
+            output_buffer,
+            ACL_FLOAT16,
+            output_elem_size, output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
+
+        ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
     }
-
-    aclTensor* acl_output_tensor =
-        ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size,
-                                output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
-    aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-    aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT);
-
-    ACL_CHECK(aclDestroyTensor(acl_output_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
 }
 
 void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
@@ -2859,15 +3000,27 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ACL_CHECK(aclDestroyTensor(acl_cos_tensor));
 }
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+    const aclTensor* x, const aclTensor* cos, const aclTensor* sin,
+    int64_t mode, const aclTensor* yOut, uint64_t* workspaceSize,
+    aclOpExecutor** executor);
+aclnnStatus aclnnRotaryPositionEmbedding(void* workspace,
+                                         uint64_t workspaceSize,
+                                         aclOpExecutor* executor,
+                                         aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
-    // TODO: with freq_factors
-    GGML_ASSERT(src2 == NULL);
-
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
     // const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -2885,13 +3038,19 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    GGML_ASSERT(n_dims <= ne0);
+    // TODO: with freq_factors
+    GGML_ASSERT(src2 == NULL);
+    // TODO: attn_factor != 1
+    GGML_ASSERT(attn_factor == 1);
+    // TODO: n_dims <= ne0
+    GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
-
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
     // TODO: freq_scale != 1
     GGML_ASSERT(freq_scale == 1);
+    // TODO: type == GGML_TYPE_F16
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -2924,177 +3083,30 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
                      theta_scale, is_neox);
 
-    // roll input
-    void* input_roll_buffer;
-    aclTensor* acl_minus_one_tensor;
-    void* minus_one_scale_buffer = nullptr;
-    ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
-    ggml_cann_pool_alloc minus_one_scale_allocator(
-        ctx.pool(), sizeof(float_t) * src0->ne[0]);
-    if (!is_neox) {
-        // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
-        input_roll_buffer = roll_allocator.get();
-        int64_t input_roll_ne[4] = {2, src0->ne[1] * (src0->ne[0] / 2),
-                                    src0->ne[2], src0->ne[3]};
-        size_t input_roll_nb[GGML_MAX_DIMS];
-        input_roll_nb[0] = ggml_type_size(src0->type);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            input_roll_nb[i] = input_roll_nb[i - 1] * input_roll_ne[i - 1];
-        }
-        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
-            input_roll_buffer, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
-            GGML_MAX_DIMS);
-        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
-            src0->data, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
-            GGML_MAX_DIMS);
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
 
-        int64_t shifts[] = {1};
-        int64_t dims[] = {3};
-        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
-        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
-        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+    void* workspaceAddr = nullptr;
 
-        // init [-1, 1, -1, 1, ...]
-        minus_one_scale_buffer = minus_one_scale_allocator.get();
-
-        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
-        size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
-        }
-        acl_minus_one_tensor = aclnn_ones(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
-        int64_t dim = 3;
-        int64_t* index = new int64_t[src0->ne[0]];
-        for (int i = 0; i < src0->ne[0]; i++) {
-            index[i] = i / 2 * 2;
-        }
-        int64_t index_num = src0->ne[0];
-        float value = -1;
-        aclnn_index_fill_tensor(ctx, acl_minus_one_tensor, dim, index,
-                                index_num, value);
-    } else {
-        // roll input: [q0,q1,q2,...] ->
-        // [q_half,q_half+1,...,q_end,q0,q1,...q_half-1]
-        input_roll_buffer = roll_allocator.get();
-        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
-            input_roll_buffer, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
-        aclTensor* acl_input_tensor = ggml_cann_create_tensor(src0);
-
-        int64_t shifts[] = {src0->ne[0] / 2};
-        int64_t dims[] = {3};
-        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
-
-        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
-        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
-
-        // init [-1, -1, -1, 1, 1，1，...]
-        minus_one_scale_buffer = minus_one_scale_allocator.get();
-
-        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
-        size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
-        }
-        acl_minus_one_tensor = aclnn_ones(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
-        // -1 * first half
-        int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
-        size_t first_half_nb[GGML_MAX_DIMS];
-        first_half_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
-        }
-        aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
-            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
-            first_half_nb, GGML_MAX_DIMS);
-        bool inplace = true;
-        float scale = -1;
-        aclnn_muls(ctx, acl_first_half_tensor, scale, nullptr, inplace);
-        ACL_CHECK(aclDestroyTensor(acl_first_half_tensor));
+    int acl_mode = mode;
+    if (mode == 0) {
+        acl_mode = 1;
     }
 
-    // TODO: n_dims < ne0
-    GGML_ASSERT(n_dims == src0->ne[0]);
-
-    // input * scale
-    ggml_cann_pool_alloc roll_mul_scale_allocator(ctx.pool(),
-                                                  ggml_nbytes(src0));
-    void* input_roll_mul_scale_buffer = roll_mul_scale_allocator.get();
-    size_t input_nb[GGML_MAX_DIMS];
-    input_nb[0] = ggml_type_size(src0->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        input_nb[i] = input_nb[i - 1] * src0->ne[i - 1];
-    }
-    aclTensor* acl_input_roll_mul_scale_tensor = ggml_cann_create_tensor(
-        input_roll_mul_scale_buffer, ggml_cann_type_mapping(src0->type),
-        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
-    aclTensor* acl_input_roll_reshape_tensor = ggml_cann_create_tensor(
-        input_roll_buffer, ggml_cann_type_mapping(src0->type),
-        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
-
-    aclnn_mul(ctx, acl_input_roll_reshape_tensor, acl_minus_one_tensor,
-              acl_input_roll_mul_scale_tensor);
-
-    // output
-    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+    aclTensor* acl_x = ggml_cann_create_tensor(src0);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
-    void* output_fp32_buffer;
-    if (src0->type == GGML_TYPE_F32) {
-        aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor);
-        aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
-                          acl_sin_reshape_tensor);
-        aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst);
-        // TODO: ne0 != n_dims in mode2
-    } else if (src0->type == GGML_TYPE_F16) {
-        size_t input_fp32_nb[GGML_MAX_DIMS];
-        input_fp32_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
-        }
-        ggml_cann_pool_alloc fp32_allocator1(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        void* input_fp32_buffer1 = fp32_allocator1.get();
-        aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
-            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-        ggml_cann_pool_alloc fp32_allocator2(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        void* input_fp32_buffer2 = fp32_allocator2.get();
-        aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
-            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-
-        ggml_cann_pool_alloc fp32_allocator(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        output_fp32_buffer = fp32_allocator.get();
-        aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
-            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-        aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1);
-        aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
-                  input_fp32_tensor2);
-        aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
-                  output_fp32_tensor);
-        aclnn_cast(ctx, output_fp32_tensor, acl_dst, ACL_FLOAT16);
-
-        ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
-        ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
-        ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+    ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+        acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
     }
 
-    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+    ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
+                                           executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_x));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));
 }

ggml/src/ggml-cann/common.h

@@ -211,17 +211,20 @@ struct ggml_cann_pool_alloc {
 struct ggml_backend_cann_context {
     int32_t device;                  /**< Device ID. */
     std::string name;                /**< Name of the device. */
+    std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {
-        {nullptr}}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {}
+        : device(device), name("CANN" + std::to_string(device)) {
+        ggml_cann_set_device(device);
+        description = aclrtGetSocName();
+    }
 
     /**
      * @brief Destructor for cleaning up resources.

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

@@ -122,6 +122,10 @@ static ggml_cann_device_info ggml_cann_init() {
         ACL_CHECK(aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
             &info.devices[id].vmm_granularity));
+
+        size_t free, total;
+        ggml_backend_cann_get_device_memory(id, &free, &total);
+        info.devices[id].total_vram = free;
     }
 
     // TODO: add more device info later.
@@ -208,6 +212,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
+        const size_t alignment = 128;
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 #ifdef DEBUG_CANN_MALLOC
         int nnz = 0;
         size_t max_size = 0;
@@ -246,13 +255,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
             return ptr;
         }
         void* ptr;
-        size_t look_ahead_size = (size_t)(1.05 * size);
-        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
         ggml_cann_set_device(device);
         ACL_CHECK(
-            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = look_ahead_size;
-        pool_size += look_ahead_size;
+            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
 #ifdef DEBUG_CANN_MALLOC
         GGML_LOG_INFO(
             "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
@@ -296,7 +303,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
     /**
      * @brief The maximum size of the virtual memory pool (32 GB).
      */
-    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+    size_t max_size;
 
     /**
      * @brief The device ID associated with this buffer pool.
@@ -341,7 +348,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      */
     explicit ggml_cann_pool_vmm(int device)
         : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+        auto dev = ggml_cann_info().devices[device];
+        granularity = dev.vmm_granularity;
+        max_size = dev.total_vram;
+    }
 
     /**
      * @brief Destructor to free all buffers in the virtual memory pool.
@@ -370,17 +381,19 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
         // round up the allocation size to the alignment to ensure that all
         // allocations are aligned for all data types
         const size_t alignment = 128;
-        size = alignment * ((size + alignment - 1) / alignment);
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 
         size_t avail = pool_size - pool_used;
 
         if (size > avail) {
             // round up to the next multiple of the granularity
             size_t reserve_size = size - avail;
-            reserve_size =
-                granularity * ((reserve_size + granularity - 1) / granularity);
+            reserve_size = GGML_PAD(reserve_size, granularity);
 
-            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+            GGML_ASSERT(pool_size + reserve_size <= max_size);
 
             // allocate more physical memory
             aclrtPhysicalMemProp prop = {};
@@ -396,7 +409,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // reserve virtual address space (if not already reserved)
             if (pool_addr == 0) {
                 ACL_CHECK(aclrtReserveMemAddress(
-                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+                    &pool_addr, max_size, 0, NULL, 1));
             }
 
             // map at the end of the pool
@@ -409,10 +422,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // add to the pool
             pool_size += reserve_size;
 
-            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
-            // reserved %llu MB)\n",
-            //       device, (unsigned long long) (pool_size/1024/1024),
-            //       (unsigned long long) (reserve_size/1024/1024));
+#ifdef DEBUG_CANN_MALLOC
+             GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+                   device, (unsigned long long) (pool_size/1024/1024),
+                   (unsigned long long) (reserve_size/1024/1024));
+#endif
         }
 
         GGML_ASSERT(pool_addr != 0);
@@ -457,7 +471,6 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
 }
 
@@ -1130,10 +1143,10 @@ ggml_backend_cann_buffer_type(int32_t device) {
     static bool ggml_backend_cann_buffer_type_initialized = false;
 
     if (!ggml_backend_cann_buffer_type_initialized) {
-        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+        for (int32_t i = 0; i < ggml_cann_info().device_count; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
-                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), i),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1199,10 +1212,15 @@ static void * ggml_cann_host_malloc(size_t size) {
         return nullptr;
     }
 
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
+
     void * hostPtr = nullptr;
     aclError err = aclrtMallocHost((void **) &hostPtr, size);
     if (err != ACL_SUCCESS) {
-
         GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
                            size / 1024.0 / 1024.0, aclGetRecentErrMsg());
         return nullptr;
@@ -1669,12 +1687,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
+                case GGML_TYPE_Q8_0:
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
+                    if (op->src[0]->ne[0] <= QK8_0) {
+                        return false;
+                    }
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
-                case GGML_TYPE_Q8_0:
-                    // TODO: fix me
-                    // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
                 case GGML_TYPE_Q4_0:
                     return true;
                 default:
@@ -1706,9 +1726,61 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         }
+        case GGML_OP_CONT: {
+            // TODO: support GGML_TYPE_BF16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_ROPE: {
+            // TODO: with ops-test v == 1
+            float * freq_scale = (float*)((int32_t*)op->op_params + 6);
+            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            float * attn_factor = (float*)((int32_t*)op->op_params + 8);
+            // TODO: with freq_factors
+            if (op->src[2] != NULL) {
+                return false;
+            }
+            // TODO: n_dims <= ne0
+            if (op->src[0]->ne[0] != op->op_params[1]) {
+                return false;
+            }
+            // TODO: ext_factor != 0
+            if (*ext_factor != 0) {
+                return false;
+            }
+            // TODO: freq_scale != 1
+            if (*freq_scale != 1) {
+                return false;
+            }
+            // TODO: attn_factor != 1
+            if (*attn_factor != 1) {
+                return false;
+            }
+            //TODO: type == GGML_TYPE_F16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_UPSCALE: {
+            // aclnnUpsampleNearest2dGetWorkspaceSize not support
+            // selfDimN[2]/outDimN[2] or selfDimC[3]/outDimC[3] not equal
+            if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
+                return false;
+            }
+            return true;
+        }
+        case GGML_OP_IM2COL:
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -1722,17 +1794,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -2064,16 +2132,17 @@ 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 +2195,5 @@ 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-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)

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

@@ -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

@@ -46,13 +46,10 @@ if (CUDAToolkit_FOUND)
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
     endif()
 
-    add_library(ggml-cuda
-                ${GGML_HEADERS_CUDA}
-                ${GGML_SOURCES_CUDA}
-                )
-
-    target_link_libraries(ggml-cuda PRIVATE ggml-base)
-    target_include_directories(ggml-cuda PRIVATE . ..)
+    ggml_add_backend_library(ggml-cuda
+                             ${GGML_HEADERS_CUDA}
+                             ${GGML_SOURCES_CUDA}
+                            )
 
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 

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

@@ -3126,6 +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) {
@@ -3137,6 +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;
 }
 
@@ -3169,16 +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
             };
         }
 
@@ -3209,3 +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-hip/CMakeLists.txt

@@ -64,12 +64,10 @@ else()
     list(APPEND GGML_SOURCES_ROCM ${SRCS})
 endif()
 
-add_library(ggml-hip
-            ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_ROCM})
-
-target_link_libraries(ggml-hip PRIVATE ggml-base)
-target_include_directories(ggml-hip PRIVATE . ..)
+ggml_add_backend_library(ggml-hip
+                         ${GGML_HEADERS_ROCM}
+                         ${GGML_SOURCES_ROCM}
+                        )
 
 # TODO: do not use CUDA definitions for HIP
 target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)

ggml/src/ggml-kompute/CMakeLists.txt

@@ -6,13 +6,13 @@ if (NOT glslc_executable)
     message(FATAL_ERROR "glslc not found")
 endif()
 
-add_library(ggml-kompute
-            ggml-kompute.cpp
-            ../../include/ggml-kompute.h
-            )
+ggml_add_backend_library(ggml-kompute
+                         ggml-kompute.cpp
+                         ../../include/ggml-kompute.h
+                        )
 
 target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
-target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
 add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
 

ggml/src/ggml-kompute/ggml-kompute.cpp

@@ -2176,9 +2176,12 @@ static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
 
 ggml_backend_reg_t ggml_backend_kompute_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_kompute_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_kompute_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return ®
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg)

ggml/src/ggml-metal/CMakeLists.txt

@@ -4,19 +4,16 @@ find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
 
 message(STATUS "Metal framework found")
 
-add_library(ggml-metal
-            ggml-metal.m
-            )
+ggml_add_backend_library(ggml-metal
+                         ggml-metal.m
+                        )
 
 target_link_libraries(ggml-metal PRIVATE
-                      ggml-base
                       ${FOUNDATION_LIBRARY}
                       ${METAL_FRAMEWORK}
                       ${METALKIT_FRAMEWORK}
                       )
 
-target_include_directories(ggml-metal PRIVATE . ..)
-
 if (GGML_METAL_NDEBUG)
     add_compile_definitions(GGML_METAL_NDEBUG)
 endif()

ggml/src/ggml-metal/ggml-metal.m

@@ -1927,7 +1927,7 @@ static void ggml_metal_encode_node(
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
-                int ne11_mm_min = 1;
+                int ne11_mm_min = 4;
 
 #if 0
                 // the numbers below are measured on M2 Ultra for 7B and 13B models
@@ -1951,316 +1951,316 @@ static void ggml_metal_encode_node(
                         }
 #endif
 
-                        // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                        // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                        if ([device supportsFamily:MTLGPUFamilyApple7] &&
-                                !ggml_is_transposed(src0) &&
-                                !ggml_is_transposed(src1) &&
-                                src1t == GGML_TYPE_F32 &&
-                                ne00 % 32 == 0 && ne00 >= 64 &&
-                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
-                            //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+                // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                if ([device supportsFamily:MTLGPUFamilyApple7] &&
+                        !ggml_is_transposed(src0) &&
+                        !ggml_is_transposed(src1) &&
+                        src1t == GGML_TYPE_F32 &&
+                        ne00 % 32 == 0 && ne00 >= 64 &&
+                        (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
+                    //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
-                            // some Metal matrix data types require aligned pointers
-                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
-                                case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
-                                case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
-                                default: break;
-                            }
+                    // some Metal matrix data types require aligned pointers
+                    // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
+                        case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
+                        case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
+                        default: break;
+                    }
 
-                            id<MTLComputePipelineState> pipeline = nil;
+                    id<MTLComputePipelineState> pipeline = nil;
 
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
-                                case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
-                                case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
-                                case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
-                                default: GGML_ABORT("MUL MAT-MAT not implemented");
-                            }
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
+                        case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
+                        case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
+                        case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
+                        default: GGML_ABORT("MUL MAT-MAT not implemented");
+                    }
 
-                            ggml_metal_kargs_mul_mm args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne02 =*/ ne02,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
+                    ggml_metal_kargs_mul_mm args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne02 =*/ ne02,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
 
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args    length:sizeof(args) atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
 
-                            [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-                        } else {
-                            int nth0 = 32;
-                            int nth1 = 1;
-                            int nrows = 1;
-                            //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                } else {
+                    int nth0 = 32;
+                    int nth1 = 1;
+                    int nrows = 1;
+                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
-                            id<MTLComputePipelineState> pipeline = nil;
+                    id<MTLComputePipelineState> pipeline = nil;
 
-                            // use custom matrix x vector kernel
-                            switch (src0t) {
-                                case GGML_TYPE_F32:
-                                    {
-                                        GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                    // use custom matrix x vector kernel
+                    switch (src0t) {
+                        case GGML_TYPE_F32:
+                            {
+                                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                nrows = 4;
+                            } break;
+                        case GGML_TYPE_F16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
                                         nrows = 4;
-                                    } break;
-                                case GGML_TYPE_F16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_BF16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_Q4_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q8_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q2_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q3_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_K:
-                                    {
-                                        nth0 = 4; //1;
-                                        nth1 = 8; //32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q6_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_M:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
-                                    } break;
-                                default:
-                                    {
-                                        GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
-                                        GGML_ABORT("not implemented");
                                     }
-                            };
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_BF16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
+                                        nrows = 4;
+                                    }
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_Q4_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q8_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q2_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q3_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_K:
+                            {
+                                nth0 = 4; //1;
+                                nth1 = 8; //32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q6_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_M:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_NL:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
+                            } break;
+                        default:
+                            {
+                                GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                GGML_ABORT("not implemented");
+                            }
+                    };
 
-                            ggml_metal_kargs_mul_mv args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne01 =*/ ne01,
-                                /*.ne02 =*/ ne02,
-                                /*.nb00 =*/ nb00,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne10 =*/ ne10,
-                                /*.ne11 =*/ ne11,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
+                    ggml_metal_kargs_mul_mv args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne10 =*/ ne10,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
 
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args length:sizeof(args) atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
-                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
-                                const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q4_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q3_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q5_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q6_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            } else {
-                                const int64_t ny = (ne11 + nrows - 1)/nrows;
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                        }
+                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                        src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                        src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
+                        const int mem_size = 32*sizeof(float);
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q4_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q3_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q5_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q6_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    } else {
+                        const int64_t ny = (ne11 + nrows - 1)/nrows;
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                }
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
@@ -4372,19 +4372,45 @@ static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t r
     GGML_UNUSED(index);
 }
 
+static struct ggml_backend_feature g_ggml_backend_metal_features[] = {
+#if defined(GGML_METAL_EMBED_LIBRARY)
+    { "EMBED_LIBRARY", "1" },
+#endif
+#if defined(GGML_METAL_USE_BF16)
+    { "BF16", "1" },
+#endif
+    { nil, nil },
+};
+
+static struct ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_metal_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_metal_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
 static struct ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_name         = */ ggml_backend_metal_reg_get_name,
     /* .device_count     = */ ggml_backend_metal_reg_device_count,
     /* .device_get       = */ ggml_backend_metal_reg_device_get,
-    /* .get_proc_address = */ NULL,
+    /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
 ggml_backend_reg_t ggml_backend_metal_reg(void) {
     // TODO: make this thread-safe somehow?
     {
         g_ggml_backend_metal_reg = (struct ggml_backend_reg) {
-            /* .iface   = */ ggml_backend_metal_reg_i,
-            /* .context = */ NULL,
+            /* .api_version = */ GGML_BACKEND_API_VERSION,
+            /* .iface       = */ ggml_backend_metal_reg_i,
+            /* .context     = */ NULL,
         };
 
         g_ggml_backend_metal_device = (struct ggml_backend_device) {
@@ -4396,3 +4422,5 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) {
 
     return &g_ggml_backend_metal_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg)

ggml/src/ggml-metal/ggml-metal.metal

@@ -5447,12 +5447,12 @@ kernel void kernel_mul_mm(
     const int im = tgpig.z;
 
     // if this block is of 64x32 shape or smaller
-    short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
     simdgroup_T8x8     ma[4];
     simdgroup_float8x8 mb[2];
@@ -5467,20 +5467,23 @@ kernel void kernel_mul_mm(
     const int i12 = im%args.ne12;
     const int i13 = im/args.ne12;
 
-    uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
-    short    offset1 = il/nl;
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0
+        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*args.nb01 + offset0) + offset1;
     device const float   * y = (device const float   *)(src1
         + args.nb13*i13
         + args.nb12*i12
-        + args.nb11*(r1 * BLOCK_SIZE_N + thread_col)
+        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
         + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
         T4x4 temp_a;
         dequantize_func(x, il, temp_a);
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         #pragma unroll(16)
@@ -5490,44 +5493,46 @@ kernel void kernel_mul_mm(
             +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
+        *(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
+        threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
 
         #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
             #pragma unroll(4)
             for (short i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
+
             simdgroup_barrier(mem_flags::mem_none);
+
             #pragma unroll(2)
             for (short i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
 
-            lsma += BLOCK_SIZE_M/SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE;
-
             #pragma unroll(8)
             for (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
+
+            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
         }
     }
 
     if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) {
         device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
+            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
 
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
@@ -5536,7 +5541,7 @@ kernel void kernel_mul_mm(
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
         threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
+                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
         }

ggml/src/ggml-musa/CMakeLists.txt

@@ -47,12 +47,10 @@ if (MUSAToolkit_FOUND)
         set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
     endforeach()
 
-    add_library(ggml-musa
-                ${GGML_HEADERS_MUSA}
-                ${GGML_SOURCES_MUSA})
-
-    target_link_libraries(ggml-musa PRIVATE ggml-base)
-    target_include_directories(ggml-musa PRIVATE . ..)
+    ggml_add_backend_library(ggml-musa
+                             ${GGML_HEADERS_MUSA}
+                             ${GGML_SOURCES_MUSA}
+                            )
 
     # TODO: do not use CUDA definitions for MUSA
     target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)

ggml/src/ggml-rpc/CMakeLists.txt

@@ -1,10 +1,8 @@
 message(STATUS "Using RPC backend")
 
-add_library(ggml-rpc
-            ggml-rpc.cpp)
-
-target_link_libraries(ggml-rpc PRIVATE ggml-base)
-target_include_directories(ggml-rpc PRIVATE . ..)
+ggml_add_backend_library(ggml-rpc
+                         ggml-rpc.cpp
+                        )
 
 if (WIN32)
     target_link_libraries(ggml-rpc PRIVATE ws2_32)

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1369,8 +1369,9 @@ static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
 
 ggml_backend_reg_t ggml_backend_rpc_reg(void) {
     static struct ggml_backend_reg ggml_backend_rpc_reg = {
-        /* .iface   = */ ggml_backend_rpc_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_rpc_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_rpc_reg;
@@ -1401,3 +1402,5 @@ ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
 
     return dev;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_rpc_reg)

ggml/src/ggml-sycl/CMakeLists.txt

@@ -16,12 +16,10 @@ endif()
 message(STATUS "SYCL found")
 #todo: AOT
 
-add_library(ggml-sycl
-            ggml-sycl.cpp
-            ../../include/ggml-sycl.h)
-
-target_link_libraries(ggml-sycl PRIVATE ggml-base)
-target_include_directories(ggml-sycl PRIVATE . ..)
+ggml_add_backend_library(ggml-sycl
+                         ggml-sycl.cpp
+                         ../../include/ggml-sycl.h
+                        )
 
 if (GGML_SYCL_F16)
     if (GGML_SYCL_TARGET STREQUAL "AMD")

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4637,16 +4637,17 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                 dev_ctx->description = prop.get_name();
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_sycl_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface       = */ ggml_backend_sycl_device_interface,
+                    /* .reg         = */ &reg,
+                    /* .context     = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_sycl_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_sycl_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -4678,3 +4679,4 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     return sycl_backend;
 }
 
+GGML_BACKEND_DL_IMPL(ggml_backend_sycl_reg)

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -3,13 +3,13 @@ find_package(Vulkan COMPONENTS glslc REQUIRED)
 if (Vulkan_FOUND)
     message(STATUS "Vulkan found")
 
-    add_library(ggml-vulkan
-                ggml-vulkan.cpp
-                ../../include/ggml-vulkan.h
-                )
+    ggml_add_backend_library(ggml-vulkan
+                             ggml-vulkan.cpp
+                             ../../include/ggml-vulkan.h
+                            )
 
-    target_link_libraries(ggml-vulkan PRIVATE ggml-base Vulkan::Vulkan)
-    target_include_directories(ggml-vulkan PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+    target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+    target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
     # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
     # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -6738,8 +6738,9 @@ static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
 
 ggml_backend_reg_t ggml_backend_vk_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_vk_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_vk_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return ®
@@ -7365,3 +7366,5 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")");
 }
 #endif
+
+GGML_BACKEND_DL_IMPL(ggml_backend_vk_reg)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -474,9 +474,15 @@ void write_output_files() {
 
 int main(int argc, char** argv) {
     std::map<std::string, std::string> args;
-    for (int i = 1; i < argc; i += 2) {
-        if (i + 1 < argc) {
-            args[argv[i]] = argv[i + 1];
+    for (int i = 1; i < argc; ++i) {
+        std::string arg = argv[i];
+        if (arg.rfind("--", 0) == 0) {
+            if (i + 1 < argc && argv[i + 1][0] != '-') {
+                args[arg] = argv[i + 1];
+                ++i;
+            } else {
+                args[arg] = "";
+            }
         }
     }
 

ggml/src/ggml.c

@@ -7571,3 +7571,26 @@ void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
     g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
     g_logger_state.log_callback_user_data = user_data;
 }
+
+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;
+}

gguf-py/gguf/constants.py

@@ -243,7 +243,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
-    OLMO_1124    = auto()
+    OLMO2        = auto()
     OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
@@ -405,7 +405,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
-    MODEL_ARCH.OLMO_1124:      "olmo_1124",
+    MODEL_ARCH.OLMO2:          "olmo2",
     MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
@@ -1071,7 +1071,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.OLMO_1124: [
+    MODEL_ARCH.OLMO2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,

gguf-py/gguf/metadata.py

@@ -545,7 +545,10 @@ class Metadata:
             gguf_writer.add_size_label(self.size_label)
 
         if self.license is not None:
-            gguf_writer.add_license(self.license)
+            if isinstance(self.license, list):
+                gguf_writer.add_license(",".join(self.license))
+            else:
+                gguf_writer.add_license(self.license)
         if self.license_name is not None:
             gguf_writer.add_license_name(self.license_name)
         if self.license_link is not None:

gguf-py/gguf/tensor_mapping.py

@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo_1124
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo_1124
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo_1124
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo2
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -145,7 +145,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -157,7 +157,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -170,7 +170,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -188,7 +188,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -215,7 +215,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo_1124
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2
         ),
 
         # Rotary embeddings
@@ -250,7 +250,7 @@ class TensorNameMap:
 
         # Post feed-forward norm
         MODEL_TENSOR.FFN_POST_NORM: (
-            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo_1124
+            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -273,7 +273,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.up_proj",                   # mpt
             "transformer.h.{bid}.mlp.dense_h_to_4h",                  # falcon
             "h.{bid}.mlp.dense_h_to_4h",                              # bloom
-            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo_1124
+            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo2
             "layers.{bid}.feed_forward.w3",                           # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
@@ -314,7 +314,7 @@ class TensorNameMap:
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo_1124
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo2
             "layers.{bid}.feed_forward.w1",               # llama-pth
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "transformer.h.{bid}.mlp.c_fc2",              # jais
@@ -346,7 +346,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
             "h.{bid}.mlp.dense_4h_to_h",                              # bloom
-            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo_1124
+            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo2
             "layers.{bid}.feed_forward.w2",                           # llama-pth
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
@@ -383,7 +383,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -392,7 +392,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm

include/llama-cpp.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include <memory>
+
+#include "llama.h"
+
+struct llama_model_deleter {
+    void operator()(llama_model * model) { llama_free_model(model); }
+};
+
+struct llama_context_deleter {
+    void operator()(llama_context * context) { llama_free(context); }
+};
+
+struct llama_sampler_deleter {
+    void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
+};
+
+typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
+typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
+typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;

include/llama.h

@@ -272,6 +272,9 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
+        ggml_backend_dev_t * devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 

pocs/CMakeLists.txt

@@ -8,5 +8,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(vdot)
+    if (NOT GGML_BACKEND_DL)
+        add_subdirectory(vdot)
+    endif()
 endif()

src/llama.cpp

@@ -179,7 +179,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
-    LLM_ARCH_OLMO_1124,
+    LLM_ARCH_OLMO2,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
@@ -233,7 +233,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
-    { LLM_ARCH_OLMO_1124,       "olmo_1124"    },
+    { LLM_ARCH_OLMO2,           "olmo2"        },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
@@ -1210,7 +1210,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_OLMO_1124,
+        LLM_ARCH_OLMO2,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
@@ -4866,7 +4866,9 @@ struct llama_model_loader {
             mappings.reserve(files.size());
             mmaps_used.reserve(files.size());
             for (const auto & file : files) {
-                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa()));
+                auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+                auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
+                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
                 mmaps_used.emplace_back(mapping->size, 0);
                 if (mlock_mmaps) {
                     std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
@@ -5898,7 +5900,7 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -7181,12 +7183,12 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ADD:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_add(ctx, a, w);
             } break;
         case GGML_OP_MUL:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_mul(ctx, a, w);
             } break;
         case GGML_OP_DIV:
@@ -8591,7 +8593,7 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
-            case LLM_ARCH_OLMO_1124:
+            case LLM_ARCH_OLMO2:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -9190,7 +9192,7 @@ static bool llm_load_tensors(
         ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
         if (!dev) {
             // FIXME: workaround for CPU backend buft having a NULL device
-            dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0);
+            dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
         }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
@@ -14481,7 +14483,7 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_olmo_1124() {
+    struct ggml_cgraph * build_olmo2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
@@ -16797,9 +16799,9 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
-                result = llm.build_olmo_1124();
+                result = llm.build_olmo2();
             } break;
         case LLM_ARCH_OLMOE:
             {
@@ -17443,8 +17445,9 @@ static enum ggml_status llama_graph_compute(
                     int   n_threads,
         ggml_threadpool * threadpool) {
     if (lctx.backend_cpu != nullptr) {
-        ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
-        ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu));
+        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
+        set_threadpool_fn(lctx.backend_cpu, threadpool);
     }
 
     // set the number of threads for all the backends
@@ -19361,6 +19364,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
 //
 struct llama_model_params llama_model_default_params() {
     struct llama_model_params result = {
+        /*.devices                     =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
@@ -19478,7 +19482,11 @@ void llama_backend_init(void) {
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
     if (numa != GGML_NUMA_STRATEGY_DISABLED) {
-        ggml_numa_init(numa);
+        auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        GGML_ASSERT(dev && "CPU backend is not loaded");
+        auto * reg = ggml_backend_dev_backend_reg(dev);
+        auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init");
+        numa_init_fn(numa);
     }
 }
 
@@ -19569,19 +19577,24 @@ struct llama_model * llama_load_model_from_file(
     }
 
     // create list of devices to use with this model
-    // currently, we use all available devices
-    // TODO: rework API to give user more control over device selection
-    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        switch (ggml_backend_dev_type(dev)) {
-            case GGML_BACKEND_DEVICE_TYPE_CPU:
-            case GGML_BACKEND_DEVICE_TYPE_ACCEL:
-                // skip CPU backends since they are handled separately
-                break;
+    if (params.devices) {
+        for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
+            model->devices.push_back(*dev);
+        }
+    } else {
+        // use all available devices
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            switch (ggml_backend_dev_type(dev)) {
+                case GGML_BACKEND_DEVICE_TYPE_CPU:
+                case GGML_BACKEND_DEVICE_TYPE_ACCEL:
+                    // skip CPU backends since they are handled separately
+                    break;
 
-            case GGML_BACKEND_DEVICE_TYPE_GPU:
-                model->devices.push_back(dev);
-                break;
+                case GGML_BACKEND_DEVICE_TYPE_GPU:
+                    model->devices.push_back(dev);
+                    break;
+            }
         }
     }
 
@@ -19752,9 +19765,6 @@ struct llama_context * llama_new_context_with_model(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    ctx->abort_callback      = params.abort_callback;
-    ctx->abort_callback_data = params.abort_callback_data;
-
     ctx->logits_all = params.logits_all;
 
     // build worst-case graph for encoder if a model contains encoder
@@ -19803,7 +19813,7 @@ struct llama_context * llama_new_context_with_model(
         }
 
         // add CPU backend
-        ctx->backend_cpu = ggml_backend_cpu_init();
+        ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
         if (ctx->backend_cpu == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
             llama_free(ctx);
@@ -19823,6 +19833,8 @@ struct llama_context * llama_new_context_with_model(
             }
         }
 
+        llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data);
+
         if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -19868,7 +19880,8 @@ struct llama_context * llama_new_context_with_model(
             std::vector<ggml_backend_t> backend_ptrs;
             for (auto & backend : ctx->backends) {
                 auto * buft = ggml_backend_get_default_buffer_type(backend.get());
-                if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) {
+                auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) {
                     // use the host buffer of the first device CPU for faster transfer of the intermediate state
                     auto * dev = model->devices[0];
                     auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
@@ -19896,7 +19909,8 @@ struct llama_context * llama_new_context_with_model(
             // pipeline parallelism requires support for async compute and events in all devices
             if (pipeline_parallel) {
                 for (auto & backend : ctx->backends) {
-                    if (ggml_backend_is_cpu(backend.get())) {
+                    auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                    if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
                         // ignore CPU backend
                         continue;
                     }
@@ -20070,7 +20084,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
@@ -21450,6 +21464,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) {
 void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
     ctx->abort_callback      = abort_callback;
     ctx->abort_callback_data = abort_callback_data;
+
+    for (auto & backend : ctx->backends) {
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
+        auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
+        if (set_abort_callback_fn) {
+            set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data);
+        }
+    }
 }
 
 void llama_set_embeddings(struct llama_context * ctx, bool embeddings) {
@@ -22191,32 +22213,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int
 }
 
 const char * llama_print_system_info(void) {
-    ggml_cpu_init(); // some ARM features are detected at runtime
-
     static std::string s;
 
-    s  = "";
-    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
-    s += "AVX_VNNI = "    + std::to_string(ggml_cpu_has_avx_vnni())    + " | ";
-    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
-    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
-    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
-    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
-    s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
-    s += "AMX_INT8 = "    + std::to_string(ggml_cpu_has_amx_int8())    + " | ";
-    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
-    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
-    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
-    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
-    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
-    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
-    s += "RISCV_VECT = "  + std::to_string(ggml_cpu_has_riscv_v())     + " | ";
-    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
-    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
-    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
-    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
-    s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
-    s += "LLAMAFILE = "   + std::to_string(ggml_cpu_has_llamafile())   + " | ";
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        auto * reg = ggml_backend_reg_get(i);
+        auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
+        if (get_features_fn) {
+            ggml_backend_feature * features = get_features_fn(reg);
+            s += ggml_backend_reg_name(reg);
+            s += " : ";
+            for (; features->name; features++) {
+                s += features->name;
+                s += " = ";
+                s += features->value;
+                s += " | ";
+            }
+        }
+    }
 
     return s.c_str();
 }

tests/CMakeLists.txt

@@ -110,23 +110,26 @@ llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CU
 # llama_target_and_test(test-double-float.cpp) # SLOW
 llama_target_and_test(test-log.cpp)
 llama_target_and_test(test-arg-parser.cpp)
-llama_target_and_test(test-quantize-fns.cpp)
-llama_target_and_test(test-quantize-perf.cpp)
 llama_target_and_test(test-sampling.cpp)
 llama_target_and_test(test-chat-template.cpp)
 
 llama_target_and_test(test-grammar-parser.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
 llama_target_and_test(test-llama-grammar.cpp)
-llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)
 
-llama_target_and_test(test-rope.cpp)
-
 llama_target_and_test(test-model-load-cancel.cpp  LABEL "model")
 llama_target_and_test(test-autorelease.cpp        LABEL "model")
 
+if (NOT GGML_BACKEND_DL)
+    # these tests use the backends directly and cannot be built with dynamic loading
+    llama_target_and_test(test-barrier.cpp)
+    llama_target_and_test(test-quantize-fns.cpp)
+    llama_target_and_test(test-quantize-perf.cpp)
+    llama_target_and_test(test-rope.cpp)
+endif()
+
 # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
 if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
     llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)

tests/test-arg-parser.cpp

@@ -70,7 +70,7 @@ int main(void) {
 
     // non-existence arg in specific example (--draft cannot be used outside llama-speculative)
     argv = {"binary_name", "--draft", "123"};
-    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SERVER));
+    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_EMBEDDING));
 
 
     printf("test-arg-parser: test valid usage\n\n");
@@ -96,7 +96,7 @@ int main(void) {
     // --draft cannot be used outside llama-speculative
     argv = {"binary_name", "--draft", "123"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE));
-    assert(params.n_draft == 123);
+    assert(params.speculative.n_max == 123);
 
 // skip this part on windows, because setenv is not supported
 #ifdef _WIN32

tests/test-backend-ops.cpp

@@ -16,7 +16,6 @@
 
 
 #include <ggml.h>
-#include <ggml-cpu.h>
 #include <ggml-alloc.h>
 #include <ggml-backend.h>
 
@@ -26,7 +25,6 @@
 #include <cstdint>
 #include <cstring>
 #include <cinttypes>
-#include <functional>
 #include <memory>
 #include <random>
 #include <stdio.h>
@@ -639,19 +637,20 @@ struct test_case {
 
         // determine number of runs
         int n_runs;
+        bool is_cpu = ggml_backend_dev_type(ggml_backend_get_device(backend)) == GGML_BACKEND_DEVICE_TYPE_CPU;
         if (op_flops(out) > 0) {
             // based on flops
             const uint64_t GFLOP = 1000 * 1000 * 1000;
             const uint64_t target_flops_cpu =   8ULL * GFLOP;
             const uint64_t target_flops_gpu = 100ULL * GFLOP;
-            uint64_t target_flops = ggml_backend_is_cpu(backend) ? target_flops_cpu : target_flops_gpu;
+            uint64_t target_flops = is_cpu ? target_flops_cpu : target_flops_gpu;
             n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_flops / op_flops(out)) + 1;
         } else {
             // based on memory size
             const size_t GB = 1ULL << 30;
             const size_t target_size_cpu =  8 * GB;
             const size_t target_size_gpu = 32 * GB;
-            size_t target_size = ggml_backend_is_cpu(backend) ? target_size_cpu : target_size_gpu;
+            size_t target_size = is_cpu ? target_size_cpu : target_size_gpu;
             n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
         }
 
@@ -3873,7 +3872,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
 static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     if (mode == MODE_TEST) {
         auto test_cases = make_test_cases_eval();
-        ggml_backend_t backend_cpu = ggml_backend_cpu_init();
+        ggml_backend_t backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, NULL);
+        if (backend_cpu == NULL) {
+            printf("  Failed to initialize CPU backend\n");
+            return false;
+        }
 
         size_t n_ok = 0;
         for (auto & test : test_cases) {
@@ -3953,7 +3956,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    // enumerate backends
+    // load and enumerate backends
+    ggml_backend_load_all();
+
     printf("Testing %zu devices\n\n", ggml_backend_dev_count());
 
     size_t n_ok = 0;
@@ -3969,16 +3974,15 @@ int main(int argc, char ** argv) {
             continue;
         }
 
-        ggml_backend_t backend = ggml_backend_dev_init(dev, NULL);
-        GGML_ASSERT(backend != NULL);
-
-        if (backend_filter == NULL && ggml_backend_is_cpu(backend) && mode != MODE_GRAD) {
+        if (backend_filter == NULL && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU && mode != MODE_GRAD) {
             printf("  Skipping CPU backend\n");
-            ggml_backend_free(backend);
             n_ok++;
             continue;
         }
 
+        ggml_backend_t backend = ggml_backend_dev_init(dev, NULL);
+        GGML_ASSERT(backend != NULL);
+
         ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
         auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
         if (ggml_backend_set_n_threads_fn) {

tests/test-quantize-fns.cpp

@@ -79,9 +79,9 @@ static float dot_product(const float * a1, const float * a2, size_t test_size) {
 }
 
 // Total dot product error
-static float dot_product_error(
-    const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data1, const float *test_data2
-) {
+static float dot_product_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data1, const float * test_data2) {
+    GGML_UNUSED(qfns);
+
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);