ci : close inactive issue, increase operations per run (#6270)

* sampling: remove duplicated code for probability distribution access

* free original_logits

* fix original_logits allocation

* fixes based on review @cebtenzzre

* change function name to `llama_sampling_prepare`

.github/workflows/build.yml

@@ -15,14 +15,133 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
 jobs:
+  macOS-latest-cmake-arm64:
+    runs-on: macos-14
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+
+  macOS-latest-cmake-x64:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+
   ubuntu-focal-make:
     runs-on: ubuntu-20.04
+    env:
+      LLAMA_NODE_AVAILABLE: true
+      LLAMA_PYTHON_AVAILABLE: true
 
     steps:
       - name: Clone
@@ -35,6 +154,14 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential gcc-8
 
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: "3.11"
+
       - name: Build
         id: make_build
         env:
@@ -48,6 +175,28 @@ jobs:
           CC=gcc-8 make tests -j $(nproc)
           make test -j $(nproc)
 
+  ubuntu-focal-make-curl:
+    runs-on: ubuntu-20.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
+
+      - name: Build
+        id: make_build
+        env:
+          LLAMA_FATAL_WARNINGS: 1
+          LLAMA_CURL: 1
+        run: |
+          CC=gcc-8 make -j $(nproc)
+
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
 
@@ -76,40 +225,51 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-  ubuntu-latest-cmake-sanitizer:
-    runs-on: ubuntu-latest
-
-    continue-on-error: true
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Test
-        id: cmake_test
+      - name: Test llama2c conversion
+        id: llama2c_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v3
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+#
+#      - name: Test
+#        id: cmake_test
+#        run: |
+#          cd build
+#          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -640,6 +800,7 @@ jobs:
 
   windows-latest-cmake-sycl:
     runs-on: windows-latest
+
     defaults:
       run:
         shell: bash
@@ -648,7 +809,6 @@ jobs:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
 
-
     steps:
       - name: Clone
         id: checkout
@@ -663,6 +823,32 @@ jobs:
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v3
+        with:
+          path: |
+            llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+
   ios-xcode-build:
     runs-on: macos-latest
 
@@ -726,6 +912,8 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cublas
+      - macOS-latest-cmake-arm64
+      - macOS-latest-cmake-x64
 
     steps:
       - name: Clone

.github/workflows/close-issue.yml

@@ -0,0 +1,23 @@
+name: Close inactive issues
+on:
+  schedule:
+    - cron: "42 0 * * *"
+
+jobs:
+  close-issues:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v5
+        with:
+          exempt-issue-labels: "refactor,help wanted,good first issue,research"
+          days-before-issue-stale: 30
+          days-before-issue-close: 14
+          stale-issue-label: "stale"
+          close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+          operations-per-run: 10000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/code-coverage.yml

@@ -5,6 +5,10 @@ env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   run:
     runs-on: ubuntu-20.04

.github/workflows/docker.yml

@@ -15,6 +15,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub

.github/workflows/editorconfig.yml

@@ -14,6 +14,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   editorconfig:
     runs-on: ubuntu-latest

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

@@ -17,6 +17,10 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/*.nix', 'flake.lock']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -8,6 +8,10 @@ on:
   pull_request:
     types: [opened, synchronize, reopened]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   nix-eval:
     strategy:

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

@@ -16,6 +16,10 @@ on:
       - 'requirements.txt'
       - 'requirements/*.txt'
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   python-check-requirements:
     runs-on: ubuntu-latest

.github/workflows/python-lint.yml

@@ -2,6 +2,10 @@ name: flake8 Lint
 
 on: [push, pull_request]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   flake8-lint:
     runs-on: ubuntu-latest

.github/workflows/server.yml

@@ -18,20 +18,23 @@ on:
   schedule:
     -  cron: '0 0 * * *'
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   server:
     runs-on: ubuntu-latest
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
         build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
-          - build_type: Debug
-            sanitizer: THREAD
-            disabled_on_pr: true
       fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     container:
@@ -57,7 +60,8 @@ jobs:
             cmake \
             python3-pip \
             wget \
-            language-pack-en
+            language-pack-en \
+            libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -67,6 +71,7 @@ jobs:
           cmake .. \
               -DLLAMA_NATIVE=OFF \
               -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
               -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
@@ -101,12 +106,21 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: libCURL
+        id: get_libcurl
+        env:
+          CURL_VERSION: 8.6.0_6
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+          mkdir $env:RUNNER_TEMP/libcurl
+          tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
+
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake ..  -DLLAMA_BUILD_SERVER=ON -DCMAKE_BUILD_TYPE=Release ;
+          cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
 
       - name: Python setup
@@ -120,6 +134,11 @@ jobs:
         run: |
           pip install -r examples/server/tests/requirements.txt
 
+      - name: Copy Libcurl
+        id: prepare_libcurl
+        run: |
+          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+
       - name: Tests
         id: server_integration_tests
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}

.github/workflows/zig-build.yml

@@ -6,6 +6,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
 jobs:
   build:
     strategy:

.gitignore

@@ -11,7 +11,10 @@
 *.gcda
 *.dot
 *.bat
+*.tmp
 *.metallib
+*.etag
+*.lastModified
 .DS_Store
 .build/
 .cache/
@@ -47,6 +50,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/gguf-split
 /gritlm
 /imatrix
 /infill
@@ -55,6 +59,9 @@ models-mnt
 /llava-cli
 /lookahead
 /lookup
+/lookup-create
+/lookup-merge
+/lookup-stats
 /main
 /metal
 /passkey

CMakeLists.txt

@@ -99,6 +99,8 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
+option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
+option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
@@ -386,6 +388,9 @@ if (LLAMA_CUBLAS)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
+        if (LLAMA_CUDA_NO_PEER_COPY)
+            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
@@ -530,6 +535,10 @@ if (LLAMA_HIPBLAS)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
 
+    if (LLAMA_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
     add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
     add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
     add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})

Makefile

@@ -1,7 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search  \
 	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
@@ -9,7 +9,8 @@ TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease                                 \
+	tests/test-json-schema-to-grammar
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -451,9 +452,9 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-#ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
-#endif # LLAMA_CUDA_CUBLAS
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
@@ -534,6 +535,9 @@ endif # LLAMA_HIP_UMA
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
@@ -553,7 +557,7 @@ endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h
+ggml-metal.o: ggml-metal.m ggml-metal.h ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 ifdef LLAMA_METAL_EMBED_LIBRARY
@@ -595,6 +599,11 @@ include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
+ifdef LLAMA_CURL
+override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
+endif
+
 #
 # Print build information
 #
@@ -661,9 +670,15 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 train.o: common/train.cpp common/train.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
@@ -671,7 +686,7 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -740,7 +755,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
@@ -748,6 +763,10 @@ gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -797,9 +816,15 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
 
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -856,6 +881,10 @@ tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

README-sycl.md

@@ -29,6 +29,8 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 ## News
 
 - 2024.3
+  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
+  - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
   - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
   - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
   - Support detecting all GPUs with level-zero and same top **Max compute units**.
@@ -81,7 +83,7 @@ For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, rec
 |-|-|-|
 |Ampere Series| Support| A100|
 
-### oneMKL
+### oneMKL for CUDA
 
 The current oneMKL release does not contain the oneMKL cuBlas backend.
 As a result for Nvidia GPU's oneMKL must be built from source.
@@ -114,7 +116,7 @@ You can choose between **F16** and **F32** build. F16 is faster for long-prompt
 # Or, for F32:
 docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
 
-# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example
+# Note: you can also use the ".devops/server-intel.Dockerfile", which compiles the "server" example
 ```
 
 ### Run
@@ -254,16 +256,16 @@ Run without parameter:
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
+found 6 SYCL devices:
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
 |Attribute|Note|
@@ -271,12 +273,35 @@ found 4 SYCL devices:
 |compute capability 1.3|Level-zero running time, recommended |
 |compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
 
-4. Set device ID and execute llama.cpp
+4. Device selection and execution of llama.cpp
 
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
+There are two device selection modes:
+
+- Single device: Use one device assigned by user.
+- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+
+|Device selection|Parameter|
+|-|-|
+|Single device|--split-mode none --main-gpu DEVICE_ID |
+|Multiple devices|--split-mode layer (default)|
+
+Examples:
+
+- Use device 0:
 
 ```sh
-GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
+```
+or run by script:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
+
+- Use multiple devices:
+
+```sh
+ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 or run by script:
 
@@ -289,12 +314,18 @@ Note:
 - By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
 
 
-5. Check the device ID in output
+5. Verify the device ID in output
+
+Verify to see if the selected GPU is shown in the output, like:
 
-Like:
 ```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
+Or
+```
+use 1 SYCL GPUs: [0] with Max compute units:512
+```
+
 
 ## Windows
 
@@ -355,7 +386,7 @@ a. Download & install cmake for Windows: https://cmake.org/download/
 
 b. Download & install mingw-w64 make for Windows provided by w64devkit
 
-- Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+- Download the 1.19.0 version of [w64devkit](https://github.com/skeeto/w64devkit/releases/download/v1.19.0/w64devkit-1.19.0.zip).
 
 - Extract `w64devkit` on your pc.
 
@@ -430,15 +461,16 @@ build\bin\main.exe
 Check the ID in startup log, like:
 
 ```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+found 6 SYCL devices:
+|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
+|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
+|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
+| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
+| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 
 ```
 
@@ -447,13 +479,31 @@ found 4 SYCL devices:
 |compute capability 1.3|Level-zero running time, recommended |
 |compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
 
-4. Set device ID and execute llama.cpp
 
-Set device ID = 0 by **set GGML_SYCL_DEVICE=0**
+4. Device selection and execution of llama.cpp
+
+There are two device selection modes:
+
+- Single device: Use one device assigned by user.
+- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+
+|Device selection|Parameter|
+|-|-|
+|Single device|--split-mode none --main-gpu DEVICE_ID |
+|Multiple devices|--split-mode layer (default)|
+
+Examples:
+
+- Use device 0:
 
 ```
-set GGML_SYCL_DEVICE=0
-build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0
+```
+
+- Use multiple devices:
+
+```
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
 or run by script:
 
@@ -466,11 +516,17 @@ Note:
 - By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
 
 
-5. Check the device ID in output
 
-Like:
+5. Verify the device ID in output
+
+Verify to see if the selected GPU is shown in the output, like:
+
 ```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
+detect 1 SYCL GPUs: [0] with top Max compute units:512
+```
+Or
+```
+use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
 ## Environment Variable
@@ -489,7 +545,6 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
 |Name|Value|Function|
 |-|-|-|
-|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
 |GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
 |ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
 
@@ -507,6 +562,9 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
 ## Q&A
 
+Note: please add prefix **[SYCL]** in issue title, so that we will check it as soon as possible.
+
+
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
 
   Miss to enable oneAPI running environment.
@@ -538,4 +596,4 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
 ## Todo
 
-- Support multiple cards.
+- Support row layer split for multiple card runs.

README.md

@@ -17,10 +17,12 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
+- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
 - Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
+- Support loading sharded model, using `gguf-split` CLI https://github.com/ggerganov/llama.cpp/pull/6187
 
 ----
 
@@ -112,6 +114,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
 - [x] [Gemma](https://ai.google.dev/gemma)
 - [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
 
 **Multimodal models:**
 
@@ -133,6 +136,7 @@ Typically finetunes of the base models below are supported as well.
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
+- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
@@ -163,6 +167,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
 - [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)

build.zig

@@ -122,6 +122,7 @@ pub fn build(b: *std.build.Builder) !void {
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
@@ -133,7 +134,7 @@ pub fn build(b: *std.build.Builder) !void {
     _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
     _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

common/CMakeLists.txt

@@ -47,6 +47,8 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+set(TARGET json-schema-to-grammar)
+add_library(${TARGET} OBJECT json-schema-to-grammar.cpp json-schema-to-grammar.h)
 
 set(TARGET common)
 
@@ -60,14 +62,28 @@ add_library(${TARGET} STATIC
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    json.hpp
     train.h
     train.cpp
+    ngram-cache.h
+    ngram-cache.cpp
     )
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+set(LLAMA_COMMON_EXTRA_LIBS build_info)
+
+# Use curl to download model url
+if (LLAMA_CURL)
+    find_package(CURL REQUIRED)
+    add_definitions(-DLLAMA_USE_CURL)
+    include_directories(${CURL_INCLUDE_DIRS})
+    find_library(CURL_LIBRARY curl REQUIRED)
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+endif ()
+
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)
+target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)

common/common.h

@@ -37,10 +37,13 @@ extern char const *LLAMA_COMMIT;
 extern char const *LLAMA_COMPILER;
 extern char const *LLAMA_BUILD_TARGET;
 
+struct llama_control_vector_load_info;
+
+int32_t get_num_physical_cores();
+
 //
 // CLI argument parsing
 //
-int32_t get_num_physical_cores();
 
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
@@ -85,17 +88,22 @@ struct gpt_params {
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
-    std::string model_draft       = "";                              // draft model for speculative decoding
-    std::string model_alias       = "unknown"; // model alias
-    std::string prompt            = "";
-    std::string prompt_file       = "";  // store the external prompt file name
-    std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
-    std::string input_prefix      = "";  // string to prefix user inputs with
-    std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string model                = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft          = "";  // draft model for speculative decoding
+    std::string model_alias          = "unknown"; // model alias
+    std::string model_url            = "";  // model url to download
+    std::string hf_repo              = "";  // HF repo
+    std::string hf_file              = "";  // HF file
+    std::string prompt               = "";
+    std::string prompt_file          = "";  // store the external prompt file name
+    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix         = "";  // string to prefix user inputs with
+    std::string input_suffix         = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
-    std::string logdir            = "";  // directory in which to save YAML log files
-    std::string logits_file       = "";  // file for saving *all* logits
+    std::string logdir               = "";  // directory in which to save YAML log files
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
+    std::string logits_file          = "";  // file for saving *all* logits
 
     std::vector<llama_model_kv_override> kv_overrides;
 
@@ -103,6 +111,11 @@ struct gpt_params {
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
 
+    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
+
+    int32_t control_vector_layer_start = -1; // layer range for control vector
+    int32_t control_vector_layer_end   = -1; // layer range for control vector
+
     int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                     //                                       (which is more convenient to use for plotting)
@@ -130,7 +143,7 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
-    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
+    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
@@ -183,6 +196,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -269,3 +285,31 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40
 void llama_embd_normalize(const float * inp, float * out, int n);
 
 float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
+
+//
+// Control vector utils
+//
+
+struct llama_control_vector_data {
+    int n_embd;
+
+    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
+    std::vector<float> data;
+};
+
+struct llama_control_vector_load_info {
+    float strength;
+
+    std::string fname;
+};
+
+// Load control vectors, scale each by strength, and add them together.
+// On error, returns {-1, empty}
+llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
+
+//
+// Split utils
+//
+static const char * const LLM_KV_SPLIT_NO            = "split.no";
+static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

common/json-schema-to-grammar.cpp

@@ -0,0 +1,721 @@
+#include "json-schema-to-grammar.h"
+#include <algorithm>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+const std::string SPACE_RULE = "\" \"?";
+
+std::unordered_map<std::string, std::string> PRIMITIVE_RULES = {
+    {"boolean", "(\"true\" | \"false\") space"},
+    {"number", "(\"-\"? ([0-9] | [1-9] [0-9]*)) (\".\" [0-9]+)? ([eE] [-+]? [0-9]+)? space"},
+    {"integer", "(\"-\"? ([0-9] | [1-9] [0-9]*)) space"},
+    {"value", "object | array | string | number | boolean"},
+    {"object", "\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space"},
+    {"array", "\"[\" space ( value (\",\" space value)* )? \"]\" space"},
+    {"uuid", "\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space"},
+    {"string", " \"\\\"\" (\n"
+               "        [^\"\\\\] |\n"
+               "        \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])\n"
+               "      )* \"\\\"\" space"},
+    {"null", "\"null\" space"}
+};
+std::vector<std::string> OBJECT_RULE_NAMES = {"object", "array", "string", "number", "boolean", "null", "value"};
+
+std::unordered_map<std::string, std::string> DATE_RULES = {
+    {"date", "[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )"},
+    {"time", "([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )"},
+    {"date-time", "date \"T\" time"},
+    {"date-string", "\"\\\"\" date \"\\\"\" space"},
+    {"time-string", "\"\\\"\" time \"\\\"\" space"},
+    {"date-time-string", "\"\\\"\" date-time \"\\\"\" space"}
+};
+
+static bool is_reserved_name(const std::string & name) {
+    static std::unordered_set<std::string> RESERVED_NAMES;
+    if (RESERVED_NAMES.empty()) {
+        RESERVED_NAMES.insert("root");
+        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
+        for (const auto &p : DATE_RULES) RESERVED_NAMES.insert(p.first);
+    }
+    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
+}
+
+std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
+std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+};
+
+std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
+
+template <typename Iterator>
+std::string join(Iterator begin, Iterator end, const std::string & separator) {
+    std::ostringstream result;
+    if (begin != end) {
+        result << *begin;
+        for (Iterator it = begin + 1; it != end; ++it) {
+            result << separator << *it;
+        }
+    }
+    return result.str();
+}
+
+static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        tokens.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    tokens.push_back(str.substr(start));
+
+    return tokens;
+}
+
+static std::string repeat(const std::string & str, size_t n) {
+    if (n == 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(str.length() * n);
+
+    for (size_t i = 0; i < n; ++i) {
+        result += str;
+    }
+
+    return result;
+}
+
+static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
+    std::smatch match;
+    std::string result;
+
+    std::string::const_iterator searchStart(input.cbegin());
+    std::string::const_iterator searchEnd(input.cend());
+
+    while (std::regex_search(searchStart, searchEnd, match, regex)) {
+        result.append(searchStart, searchStart + match.position());
+        result.append(replacement(match));
+        searchStart = match.suffix().first;
+    }
+
+    result.append(searchStart, searchEnd);
+
+    return result;
+}
+
+static std::string format_literal(const std::string & literal) {
+    std::string escaped = replacePattern(literal, GRAMMAR_LITERAL_ESCAPE_RE, [&](const std::smatch & match) {
+        char c = match.str()[0];
+        return GRAMMAR_LITERAL_ESCAPES.at(c);
+    });
+    return "\"" + escaped + "\"";
+}
+
+
+class SchemaConverter {
+private:
+    std::function<json(const std::string &)> _fetch_json;
+    bool _dotall;
+    std::map<std::string, std::string> _rules;
+    std::unordered_map<std::string, json> _refs;
+    std::unordered_set<std::string> _refs_being_resolved;
+    std::vector<std::string> _errors;
+    std::vector<std::string> _warnings;
+
+    std::string _add_rule(const std::string & name, const std::string & rule) {
+        std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
+        if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
+            _rules[esc_name] = rule;
+            return esc_name;
+        } else {
+            int i = 0;
+            while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
+                i++;
+            }
+            std::string key = esc_name + std::to_string(i);
+            _rules[key] = rule;
+            return key;
+        }
+    }
+
+    std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
+        std::vector<std::string> rules;
+        for (size_t i = 0; i < alt_schemas.size(); i++) {
+            rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
+        }
+        return join(rules.begin(), rules.end(), " | ");
+    }
+
+    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
+        if (!(pattern.front() == '^' && pattern.back() == '$')) {
+            _errors.push_back("Pattern must start with '^' and end with '$'");
+            return "";
+        }
+        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
+        std::unordered_map<std::string, std::string> sub_rule_ids;
+
+        size_t i = 0;
+        size_t length = sub_pattern.length();
+
+        using literal_or_rule = std::pair<std::string, bool>;
+        auto to_rule = [&](const literal_or_rule & ls) {
+            auto is_literal = ls.second;
+            auto s = ls.first;
+            return is_literal ? "\"" + s + "\"" : s;
+        };
+        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
+            size_t start = i;
+            std::vector<literal_or_rule> seq;
+
+            auto get_dot = [&]() {
+                std::string rule;
+                if (_dotall) {
+                    rule = "[\\U00000000-\\U0010FFFF]";
+                } else {
+                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
+                }
+                return _add_rule("dot", rule);
+            };
+
+            // Joins the sequence, merging consecutive literals together.
+            auto join_seq = [&]() {
+                std::vector<literal_or_rule> ret;
+
+                std::string literal;
+                auto flush_literal = [&]() {
+                    if (literal.empty()) {
+                        return false;
+                    }
+                    ret.push_back(std::make_pair(literal, true));
+                    literal.clear();
+                    return true;
+                };
+
+                for (const auto & item : seq) {
+                    auto is_literal = item.second;
+                    if (is_literal) {
+                        literal += item.first;
+                    } else {
+                        flush_literal();
+                        ret.push_back(item);
+                    }
+                }
+                flush_literal();
+
+                std::vector<std::string> results;
+                for (const auto & item : ret) {
+                    results.push_back(to_rule(item));
+                }
+                return std::make_pair(join(results.begin(), results.end(), " "), false);
+            };
+
+            while (i < length) {
+                char c = sub_pattern[i];
+                if (c == '.') {
+                    seq.push_back(std::make_pair(get_dot(), false));
+                    i++;
+                } else if (c == '(') {
+                    i++;
+                    if (i < length) {
+                        if (sub_pattern[i] == '?') {
+                            _warnings.push_back("Unsupported pattern syntax");
+                        }
+                    }
+                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
+                } else if (c == ')') {
+                    i++;
+                    if (start > 0 && sub_pattern[start - 1] != '(') {
+                        _errors.push_back("Unbalanced parentheses");
+                    }
+                    return join_seq();
+                } else if (c == '[') {
+                    std::string square_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != ']') {
+                        if (sub_pattern[i] == '\\') {
+                            square_brackets += sub_pattern.substr(i, 2);
+                            i += 2;
+                        } else {
+                            square_brackets += sub_pattern[i];
+                            i++;
+                        }
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced square brackets");
+                    }
+                    square_brackets += ']';
+                    i++;
+                    seq.push_back(std::make_pair(square_brackets, false));
+                } else if (c == '|') {
+                    seq.push_back(std::make_pair("|", false));
+                    i++;
+                } else if (c == '*' || c == '+' || c == '?') {
+                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
+                    i++;
+                } else if (c == '{') {
+                    std::string curly_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != '}') {
+                        curly_brackets += sub_pattern[i];
+                        i++;
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced curly brackets");
+                    }
+                    curly_brackets += '}';
+                    i++;
+                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
+                    int min_times = 0;
+                    int max_times = std::numeric_limits<int>::max();
+                    try {
+                        if (nums.size() == 1) {
+                            min_times = max_times = std::stoi(nums[0]);
+                        } else if (nums.size() != 2) {
+                            _errors.push_back("Wrong number of values in curly brackets");
+                        } else {
+                            if (!nums[0].empty()) {
+                                min_times = std::stoi(nums[0]);
+                            }
+                            if (!nums[1].empty()) {
+                                max_times = std::stoi(nums[1]);
+                            }
+                        }
+                    } catch (const std::invalid_argument & e) {
+                        _errors.push_back("Invalid number in curly brackets");
+                        return std::make_pair("", false);
+                    }
+                    auto &last = seq.back();
+                    auto &sub = last.first;
+                    auto sub_is_literal = last.second;
+
+                    if (min_times == 0 && max_times == std::numeric_limits<int>::max()) {
+                        sub += "*";
+                    } else if (min_times == 0 && max_times == 1) {
+                        sub += "?";
+                    } else if (min_times == 1 && max_times == std::numeric_limits<int>::max()) {
+                        sub += "+";
+                    } else {
+                        if (!sub_is_literal) {
+                            std::string & sub_id = sub_rule_ids[sub];
+                            if (sub_id.empty()) {
+                                sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub);
+                            }
+                            sub = sub_id;
+                        }
+                        std::string result;
+                        if (sub_is_literal && min_times > 0) {
+                            result = "\"" + repeat(sub.substr(1, sub.length() - 2), min_times) + "\"";
+                        } else {
+                            for (int j = 0; j < min_times; j++) {
+                                if (j > 0) {
+                                    result += " ";
+                                }
+                                result += sub;
+                            }
+                        }
+                        if (min_times > 0 && min_times < max_times) {
+                            result += " ";
+                        }
+                        if (max_times == std::numeric_limits<int>::max()) {
+                            result += sub + "*";
+                        } else {
+                            for (int j = min_times; j < max_times; j++) {
+                                if (j > min_times) {
+                                    result += " ";
+                                }
+                                result += sub + "?";
+                            }
+                        }
+                        seq.back().first = result;
+                        seq.back().second = false;
+                    }
+                } else {
+                    std::string literal;
+                    auto is_non_literal = [&](char c) {
+                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
+                    };
+                    while (i < length) {
+                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                            char next = sub_pattern[i + 1];
+                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
+                                i++;
+                                literal += sub_pattern[i];
+                                i++;
+                            } else {
+                                literal += sub_pattern.substr(i, 2);
+                                i += 2;
+                            }
+                        } else if (sub_pattern[i] == '"') {
+                            literal += "\\\"";
+                            i++;
+                        } else if (!is_non_literal(sub_pattern[i]) &&
+                                (i == length - 1 || literal.empty() || sub_pattern[i + 1] == '.' || !is_non_literal(sub_pattern[i + 1]))) {
+                            literal += sub_pattern[i];
+                            i++;
+                        } else {
+                            break;
+                        }
+                    }
+                    if (!literal.empty()) {
+                        seq.push_back(std::make_pair(literal, true));
+                    }
+                }
+            }
+            return join_seq();
+        };
+        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
+    }
+
+    std::string _resolve_ref(const std::string & ref) {
+        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
+            _refs_being_resolved.insert(ref);
+            json resolved = _refs[ref];
+            ref_name = visit(resolved, ref_name);
+            _refs_being_resolved.erase(ref);
+        }
+        return ref_name;
+    }
+
+    std::string _build_object_rule(
+        const std::vector<std::pair<std::string, json>> & properties,
+        const std::unordered_set<std::string> & required,
+        const std::string & name,
+        const json & additional_properties)
+    {
+        std::vector<std::string> required_props;
+        std::vector<std::string> optional_props;
+        std::unordered_map<std::string, std::string> prop_kv_rule_names;
+        for (const auto & kv : properties) {
+            const auto &prop_name = kv.first;
+            const auto &prop_schema = kv.second;
+
+            std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name);
+            prop_kv_rule_names[prop_name] = _add_rule(
+                name + (name.empty() ? "" : "-") + prop_name + "-kv",
+                format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name
+            );
+            if (required.find(prop_name) != required.end()) {
+                required_props.push_back(prop_name);
+            } else {
+                optional_props.push_back(prop_name);
+            }
+        }
+        if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
+            std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
+            std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
+            std::string kv_rule = _add_rule(sub_name + "-kv", _add_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            prop_kv_rule_names["*"] = kv_rule;
+            optional_props.push_back("*");
+        }
+
+        std::string rule = "\"{\" space ";
+        for (size_t i = 0; i < required_props.size(); i++) {
+            if (i > 0) {
+                rule += " \",\" space ";
+            }
+            rule += prop_kv_rule_names[required_props[i]];
+        }
+
+        if (!optional_props.empty()) {
+            rule += " (";
+            if (!required_props.empty()) {
+                rule += " \",\" space ( ";
+            }
+
+            std::function<std::string(const std::vector<std::string> &, bool)> get_recursive_refs = [&](const std::vector<std::string> & ks, bool first_is_optional) {
+                std::string res;
+                if (ks.empty()) {
+                    return res;
+                }
+                std::string k = ks[0];
+                std::string kv_rule_name = prop_kv_rule_names[k];
+                if (k == "*") {
+                    res = _add_rule(
+                        name + (name.empty() ? "" : "-") + "additional-kvs",
+                        kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
+                    );
+                } else if (first_is_optional) {
+                    res = "( \",\" space " + kv_rule_name + " )?";
+                } else {
+                    res = kv_rule_name;
+                }
+                if (ks.size() > 1) {
+                    res += " " + _add_rule(
+                        name + (name.empty() ? "" : "-") + k + "-rest",
+                        get_recursive_refs(std::vector<std::string>(ks.begin() + 1, ks.end()), true)
+                    );
+                }
+                return res;
+            };
+
+            for (size_t i = 0; i < optional_props.size(); i++) {
+                if (i > 0) {
+                    rule += " | ";
+                }
+                rule += get_recursive_refs(std::vector<std::string>(optional_props.begin() + i, optional_props.end()), false);
+            }
+            if (!required_props.empty()) {
+                rule += " )";
+            }
+            rule += " )?";
+        }
+
+        rule += " \"}\" space";
+
+        return rule;
+    }
+
+public:
+    SchemaConverter(
+        const std::function<json(const std::string &)> & fetch_json,
+        bool dotall)
+          : _fetch_json(fetch_json), _dotall(dotall)
+    {
+        _rules["space"] = SPACE_RULE;
+    }
+
+    void resolve_refs(json & schema, const std::string & url) {
+        /*
+        * Resolves all $ref fields in the given schema, fetching any remote schemas,
+        * replacing each $ref with absolute reference URL and populates _refs with the
+        * respective referenced (sub)schema dictionaries.
+        */
+        std::function<void(json &)> visit_refs = [&](json & n) {
+            if (n.is_array()) {
+                for (auto & x : n) {
+                    visit_refs(x);
+                }
+            } else if (n.is_object()) {
+                if (n.contains("$ref")) {
+                    std::string ref = n["$ref"];
+                    if (_refs.find(ref) == _refs.end()) {
+                        json target;
+                        if (ref.find("https://") == 0) {
+                            std::string base_url = ref.substr(0, ref.find('#'));
+                            auto it = _refs.find(base_url);
+                            if (it != _refs.end()) {
+                                target = it->second;
+                            } else {
+                                // Fetch the referenced schema and resolve its refs
+                                auto referenced = _fetch_json(ref);
+                                resolve_refs(referenced, base_url);
+                                _refs[base_url] = referenced;
+                            }
+                            if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) {
+                                return;
+                            }
+                        } else if (ref.find("#/") == 0) {
+                            target = schema;
+                            n["$ref"] = url + ref;
+                            ref = url + ref;
+                        } else {
+                            _errors.push_back("Unsupported ref: " + ref);
+                            return;
+                        }
+                        std::string pointer = ref.substr(ref.find('#') + 1);
+                        std::vector<std::string> tokens = split(pointer, "/");
+                        for (size_t i = 1; i < tokens.size(); ++i) {
+                            std::string sel = tokens[i];
+                            if (target.is_null() || !target.contains(sel)) {
+                                _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                return;
+                            }
+                            target = target[sel];
+                        }
+                        _refs[ref] = target;
+                    }
+                } else {
+                    for (auto & kv : n.items()) {
+                        visit_refs(kv.value());
+                    }
+                }
+            }
+        };
+
+        visit_refs(schema);
+    }
+
+    std::string _generate_constant_rule(const json & value) {
+        return format_literal(value.dump());
+    }
+
+    std::string visit(const json & schema, const std::string & name) {
+        json schema_type = schema.contains("type") ? schema["type"] : json();
+        std::string schema_format = schema.contains("format") ? schema["format"].get<std::string>() : "";
+        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
+
+        if (schema.contains("$ref")) {
+            return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
+        } else if (schema.contains("oneOf") || schema.contains("anyOf")) {
+            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
+        } else if (schema_type.is_array()) {
+            std::vector<json> schema_types;
+            for (const auto & t : schema_type) {
+                schema_types.push_back({{"type", t}});
+            }
+            return _add_rule(rule_name, _generate_union_rule(name, schema_types));
+        } else if (schema.contains("const")) {
+            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
+        } else if (schema.contains("enum")) {
+            std::vector<std::string> enum_values;
+            for (const auto & v : schema["enum"]) {
+                enum_values.push_back(_generate_constant_rule(v));
+            }
+            return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
+        } else if ((schema_type.is_null() || schema_type == "object")
+                && (schema.contains("properties") ||
+                    (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
+            std::unordered_set<std::string> required;
+            if (schema.contains("required") && schema["required"].is_array()) {
+                for (const auto & item : schema["required"]) {
+                    if (item.is_string()) {
+                        required.insert(item.get<std::string>());
+                    }
+                }
+            }
+            std::vector<std::pair<std::string, json>> properties;
+            if (schema.contains("properties")) {
+                for (const auto & prop : schema["properties"].items()) {
+                    properties.emplace_back(prop.key(), prop.value());
+                }
+            }
+            return _add_rule(rule_name,
+                _build_object_rule(
+                    properties, required, name,
+                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
+        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+            std::unordered_set<std::string> required;
+            std::vector<std::pair<std::string, json>> properties;
+            std::string hybrid_name = name;
+            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
+                if (comp_schema.contains("$ref")) {
+                    add_component(_refs[comp_schema["$ref"]], is_required);
+                } else if (comp_schema.contains("properties")) {
+                    for (const auto & prop : comp_schema["properties"].items()) {
+                        properties.emplace_back(prop.key(), prop.value());
+                        if (is_required) {
+                            required.insert(prop.key());
+                        }
+                    }
+                } else {
+                  // todo warning
+                }
+            };
+            for (auto & t : schema["allOf"]) {
+                if (t.contains("anyOf")) {
+                    for (auto & tt : t["anyOf"]) {
+                        add_component(tt, false);
+                    }
+                } else {
+                    add_component(t, true);
+                }
+            }
+            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
+        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
+            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
+            if (items.is_array()) {
+                std::string rule = "\"[\" space ";
+                for (size_t i = 0; i < items.size(); i++) {
+                    if (i > 0) {
+                        rule += " \",\" space ";
+                    }
+                    rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i));
+                }
+                rule += " \"]\" space";
+                return _add_rule(rule_name, rule);
+            } else {
+                std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
+                std::string list_item_operator = "( \",\" space " + item_rule_name + " )";
+                std::string successive_items;
+                int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
+                json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
+                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : -1;
+                if (min_items > 0) {
+                    successive_items += repeat(list_item_operator, min_items - 1);
+                    min_items--;
+                }
+                if (max_items >= 0 && max_items > min_items) {
+                    successive_items += repeat(list_item_operator + "?", max_items - min_items - 1);
+                } else {
+                    successive_items += list_item_operator + "*";
+                }
+                std::string rule;
+                if (min_items == 0) {
+                    rule =  "\"[\" space ( " + item_rule_name + " " + successive_items + " )? \"]\" space";
+                } else {
+                    rule =  "\"[\" space " + item_rule_name + " " + successive_items + " \"]\" space";
+                }
+                return _add_rule(rule_name, rule);
+            }
+        } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
+            return _visit_pattern(schema["pattern"], rule_name);
+        } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
+            return _add_rule(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
+        } else if ((schema_type.is_null() || schema_type == "string") && DATE_RULES.find(schema_format) != DATE_RULES.end()) {
+            for (const auto & kv : DATE_RULES) {
+                _add_rule(kv.first, kv.second);
+            }
+            return schema_format + "-string";
+        } else if (schema.empty() || schema_type == "object") {
+            for (const auto & n : OBJECT_RULE_NAMES) {
+                _add_rule(n, PRIMITIVE_RULES.at(n));
+            }
+            return _add_rule(rule_name, "object");
+        } else {
+            if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
+                _errors.push_back("Unrecognized schema: " + schema.dump());
+                return "";
+            }
+            // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+            return _add_rule(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+        }
+    }
+
+    void check_errors() {
+        if (!_errors.empty()) {
+            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
+        }
+        if (!_warnings.empty()) {
+            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
+        }
+    }
+
+    std::string format_grammar() {
+        std::stringstream ss;
+        for (const auto & kv : _rules) {
+            ss << kv.first << " ::= " << kv.second << std::endl;
+        }
+        return ss.str();
+    }
+};
+
+std::string json_schema_to_grammar(const json & schema) {
+    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
+    auto copy = schema;
+    converter.resolve_refs(copy, "input");
+    converter.visit(copy, "");
+    converter.check_errors();
+    return converter.format_grammar();
+}

common/json-schema-to-grammar.h

@@ -0,0 +1,4 @@
+#pragma once
+#include "json.hpp"
+
+std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);

common/ngram-cache.cpp

@@ -0,0 +1,280 @@
+#include "ngram-cache.h"
+#include "log.h"
+
+#include <fstream>
+
+void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
+                              std::vector<llama_token> & inp, int nnew, bool print_progress) {
+    const int64_t t_start_ms = ggml_time_ms();
+    const int64_t inp_size = inp.size();
+
+    const int64_t n_todo = inp_size * (ngram_max - ngram_min + 1);
+    int64_t n_done = 0;
+
+    for (int64_t ngram_size = ngram_min; ngram_size <= ngram_max; ++ngram_size) {
+        const int64_t i_start = std::max(inp_size - nnew, ngram_size);
+        for (int64_t i = i_start; i < inp_size; ++i) {
+            const int64_t ngram_start = i - ngram_size;
+            llama_ngram ngram(&inp[ngram_start], ngram_size);
+            const llama_token token = inp[i];
+
+            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
+            if (part_it == ngram_cache.end()) {
+                llama_ngram_cache_part part;
+                part.emplace(token, 1);
+                ngram_cache.emplace(ngram, part);
+            } else {
+                llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
+                if (token_count_it == part_it->second.end()) {
+                    part_it->second.emplace(token, 1);
+                } else {
+                    token_count_it->second++;
+                }
+            }
+            ++n_done;
+
+            if (print_progress && n_done % 10000000 == 0) {
+                const int64_t t_now_ms = ggml_time_ms();
+                const int64_t eta_ms   = (inp_size*(ngram_max-ngram_min+1) - n_done) * (t_now_ms - t_start_ms) / n_done;
+                const int64_t eta_min  = eta_ms / (60*1000);
+                const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+                fprintf(stderr, "%s: %" PRId64 "/%" PRId64 " done, ETA: %02" PRId64 ":%02" PRId64 "\n", __func__, n_done, n_todo, eta_min, eta_s);
+            }
+        }
+    }
+}
+
+// Helper function to get a token from the combined, speculative sequence of inp and draft.
+static llama_token get_token(const std::vector<llama_token> & inp, const std::vector<llama_token> & draft, const size_t i) {
+    return i < inp.size() ? inp[i] : draft[1 + i - inp.size()];
+}
+
+// If sample size or percentage are below these thresholds the draft is aborted early:
+constexpr int    draft_min_sample_size_lax[LLAMA_NGRAM_MAX] = { 2,  2,  1,  1};
+constexpr int        draft_min_percent_lax[LLAMA_NGRAM_MAX] = {66, 50, 50, 50};
+constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4,  3,  2,  2};
+constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
+
+// Helper function that tries to draft a token from only the static ngram cache:
+static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
+    llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+    if (part_static_it == nc_static.end()) {
+        return -1;
+    }
+    const llama_ngram_cache_part part_static = part_static_it->second;
+
+    int max_count_static  = 0;
+    int sum_count_static  = 0;
+    llama_token max_token = -1;
+
+    for (std::pair<llama_token, int> token_count_static : part_static) {
+        const llama_token token = token_count_static.first;
+        const int32_t count_static  = token_count_static.second;
+
+        if (count_static > max_count_static) {
+            max_token        = token;
+            max_count_static = count_static;
+        }
+        sum_count_static += count_static;
+    }
+
+    if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
+        return -1;
+    }
+    if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
+        return -1;
+    }
+    return max_token;
+}
+
+// Try to draft a token from primary cache (context/dynamic), validate with static cache:
+static llama_token try_draft(
+    llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
+    const int * min_sample_size, const int * min_percent) {
+
+    llama_token drafted_token = -1;
+
+    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
+        const llama_ngram ngram_primary = ngrams_primary[i];
+
+        llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
+        if (part_primary_it == nc_primary.end()) {
+            continue;
+        }
+        const llama_ngram_cache_part part_primary = part_primary_it->second;
+
+        int max_count_primary = 0;
+        int max_count_static  = 0;
+        int sum_count_primary = 0;
+        llama_token max_token = -1;
+
+        for (std::pair<llama_token, int> token_count_primary : part_primary) {
+            const llama_token token = token_count_primary.first;
+
+            llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
+
+            const int32_t count_primary = token_count_primary.second;
+            const int32_t count_static  = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
+
+            if (count_primary*count_static > max_count_primary*max_count_static) {
+                max_token         = token;
+                max_count_primary = count_primary;
+                max_count_static  = count_static;
+            }
+            sum_count_primary += count_primary;
+        }
+
+        if (sum_count_primary < min_sample_size[i]) {
+            continue;
+        }
+        if (100*max_count_primary < min_percent[i]*sum_count_primary) {
+            continue;;
+        }
+        drafted_token = max_token;
+    }
+
+    return drafted_token;
+}
+
+void llama_ngram_cache_draft(
+    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
+    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
+) {
+    GGML_ASSERT(draft.size() == 1);
+    const int inp_size = inp.size();
+
+    if (inp_size < LLAMA_NGRAM_STATIC) {
+        return;
+    }
+
+    while ((int) draft.size()-1 < n_draft) {
+        llama_token drafted_token = -1;
+
+        const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
+        llama_ngram ngram_static;
+        for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
+            ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
+        }
+        llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+        llama_ngram_cache_part part_static;
+        if (part_static_it != nc_static.end()) {
+            part_static = part_static_it->second;
+        }
+
+        // cd = context + dynamic
+        std::vector<llama_ngram> ngrams_cd;
+        for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
+            const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
+            llama_ngram ngram_cd;
+            for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
+                ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
+            }
+            ngrams_cd.push_back(ngram_cd);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_static, ngram_static);
+        }
+
+        if (drafted_token == -1) {
+            break;
+        }
+
+        LOG(" - draft candidate: token=%d\n", drafted_token);
+        draft.push_back(drafted_token);
+    }
+}
+
+void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
+    std::ofstream file_out(filename, std::ios::binary);
+    for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
+        const llama_ngram      ngram        = item.first;
+        llama_ngram_cache_part token_counts = item.second;
+        GGML_ASSERT(!token_counts.empty());
+        const int32_t ntokens = token_counts.size();
+        GGML_ASSERT(ntokens > 0);
+
+        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(llama_ngram));
+        file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
+        for (std::pair<llama_token, int32_t> item2 : token_counts) {
+            const llama_token token = item2.first;
+            const int32_t     count = item2.second;
+            GGML_ASSERT(count > 0);
+
+            file_out.write(reinterpret_cast<const char *>(&token), sizeof(llama_token));
+            file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
+        }
+    }
+
+}
+
+llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
+    std::ifstream hashmap_file(filename, std::ios::binary);
+    if (!hashmap_file) {
+        throw std::ifstream::failure("Unable to open file " + filename);
+    }
+    llama_ngram_cache ngram_cache;
+
+    llama_ngram ngram;
+    int32_t     ntokens;
+    llama_token token;
+    int32_t     count;
+
+    char * ngramc   = reinterpret_cast<char*>(&ngram);
+    char * ntokensc = reinterpret_cast<char*>(&ntokens);
+    char * tokenc   = reinterpret_cast<char*>(&token);
+    char * countc   = reinterpret_cast<char*>(&count);
+    while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
+        GGML_ASSERT(!hashmap_file.eof());
+        GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
+        GGML_ASSERT(ntokens > 0);
+        llama_ngram_cache_part token_counts;
+
+        for (int i = 0; i < ntokens; ++i) {
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(tokenc, sizeof(llama_token)));
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(countc, sizeof(int32_t)));
+            GGML_ASSERT(count > 0);
+            token_counts.emplace(token, count);
+        }
+
+        ngram_cache.emplace(ngram, token_counts);
+    }
+    GGML_ASSERT(hashmap_file.eof());
+
+    return ngram_cache;
+}
+
+void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
+    for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
+        const llama_ngram      ngram = ngram_part.first;
+        llama_ngram_cache_part  part = ngram_part.second;
+
+        llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
+        if (part_merged_it == ngram_cache_target.end()) {
+            ngram_cache_target.emplace(ngram, part);
+            continue;
+        }
+
+        for (std::pair<llama_token, int32_t> token_count : part) {
+            const llama_token token = token_count.first;
+            const int32_t     count = token_count.second;
+            GGML_ASSERT(count > 0);
+
+            llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
+            if (token_count_merged_it == part_merged_it->second.end()) {
+                part_merged_it->second.emplace(token, count);
+                continue;
+            }
+
+            token_count_merged_it->second += count;
+        }
+    }
+}

common/ngram-cache.h

@@ -0,0 +1,94 @@
+#pragma once
+
+#include "llama.h"
+
+#include <unordered_map>
+#include <string>
+#include <vector>
+
+#define LLAMA_NGRAM_MIN    1
+#define LLAMA_NGRAM_MAX    4
+#define LLAMA_NGRAM_STATIC 2
+
+// Data structures to map n-grams to empirical token probabilities:
+
+struct llama_ngram {
+    llama_token tokens[LLAMA_NGRAM_MAX];
+
+    llama_ngram() {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            tokens[i] = -1;
+        }
+    }
+
+    llama_ngram(const llama_token * input, const int ngram_size) {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            tokens[i] = i < ngram_size ? input[i] : -1;
+        }
+    }
+
+    bool operator==(const llama_ngram & other) const {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            if (tokens[i] != other.tokens[i]) {
+                return false;
+            }
+        }
+        return true;
+    }
+};
+
+struct llama_ngram_hash_function {
+    size_t operator()(const llama_ngram & ngram) const {
+        size_t hash = 0;
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
+        }
+        return hash;
+    }
+};
+
+// token -> number of times token has been seen
+typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
+
+// n-gram -> empirical distribution of following tokens
+typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
+
+
+// Update an ngram cache with tokens.
+// ngram_cache:         the cache to modify.
+// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data.
+// inp_data:            the token sequence with which to update ngram_cache.
+// nnew:                how many new tokens have been appended to inp_data since the last call to this function.
+// print_progress:      whether to print progress to stderr.
+//
+// In order to get correct results inp_data can ONLY BE APPENDED TO.
+// Changes in the middle need a complete rebuild.
+void llama_ngram_cache_update(
+    llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
+
+// Try to draft tokens from ngram caches.
+// inp:                the tokens generated so far.
+// draft:              the token sequence to draft. Expected to initially contain the previously sampled token.
+// n_draft:            maximum number of tokens to add to draft.
+// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic.
+// nc_context:         ngram cache based on current context.
+// nc_dynamic:         ngram cache based on previous user generations.
+// nc_static:          ngram cache generated from a large text corpus, used for validation.
+void llama_ngram_cache_draft(
+    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
+    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
+
+// Save an ngram cache to a file.
+// ngram_cache: the ngram cache to save.
+// filename:    the path under which to save the ngram cache.
+void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
+
+// Load an ngram cache saved with llama_ngram_cache_save.
+// filename: the path from which to load the ngram cache.
+// returns:  an ngram cache containing the information saved to filename.
+llama_ngram_cache llama_ngram_cache_load(std::string & filename);
+
+// Merge two ngram caches.
+// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
+// ngram_cache_add:    the ngram cache to add to ngram_cache_target.
+void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);

common/sampling.cpp

@@ -168,77 +168,20 @@ static llama_token llama_sampling_sample_impl(
                   bool is_resampling) {  // Add a parameter to indicate if we are resampling
     const llama_sampling_params & params = ctx_sampling->params;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
     const float   temp            = params.temp;
-    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
-    const float   penalty_repeat  = params.penalty_repeat;
-    const float   penalty_freq    = params.penalty_freq;
-    const float   penalty_present = params.penalty_present;
     const int     mirostat        = params.mirostat;
     const float   mirostat_tau    = params.mirostat_tau;
     const float   mirostat_eta    = params.mirostat_eta;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
 
+    std::vector<float> original_logits;
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
+    if (!is_resampling) {
+        GGML_ASSERT(!original_logits.empty());
+    }
     llama_token id = 0;
-
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
-
-    if (!is_resampling) {
-        // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
-        original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
-    }
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    if (ctx_cfg) {
-        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
-        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
-    }
-
-    cur.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    // apply penalties
-    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
-    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
-    if (penalty_tokens_used_size) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
-                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    // If we are in the resampling phase, apply grammar checks before sampling logic
-    if (is_resampling && ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
     if (temp < 0.0) {
         // greedy sampling, with probs
         llama_sample_softmax(ctx_main, &cur_p);
@@ -302,11 +245,13 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
-static llama_token_data_array llama_sample_probability_distribution_impl(
+static llama_token_data_array llama_sampling_prepare_impl(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
-                  const int idx) {
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
     const llama_sampling_params & params = ctx_sampling->params;
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
@@ -315,6 +260,7 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
     const float   penalty_repeat  = params.penalty_repeat;
     const float   penalty_freq    = params.penalty_freq;
     const float   penalty_present = params.penalty_present;
+
     const bool    penalize_nl     = params.penalize_nl;
 
     auto & prev = ctx_sampling->prev;
@@ -323,8 +269,10 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
+    if (apply_grammar && original_logits != NULL) {
+        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
+        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+    }
 
     // apply params.logit_bias map
     for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
@@ -364,12 +312,11 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
         }
     }
 
-    // apply grammar checks
-    if (ctx_sampling->grammar != NULL) {
+    // apply grammar checks before sampling logic
+    if (apply_grammar && ctx_sampling->grammar != NULL) {
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
-    llama_sample_softmax(ctx_main, &cur_p);
     return cur_p;
 }
 
@@ -382,12 +329,14 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
-llama_token_data_array llama_sampling_probability_distribution(
+llama_token_data_array llama_sampling_prepare(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
-                  const int idx) {
-    return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx);
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
+    return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
 }
 
 void llama_sampling_accept(

common/sampling.h

@@ -32,13 +32,13 @@ typedef struct llama_sampling_params {
     float       dynatemp_range        = 0.00f;    // 0.0 = disabled
     float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
     int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_repeat        = 1.00f;    // 1.0 = disabled
     float       penalty_freq          = 0.00f;    // 0.0 = disabled
     float       penalty_present       = 0.00f;    // 0.0 = disabled
     int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float       mirostat_tau          = 5.00f;    // target entropy
     float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = true;     // consider newlines as a repeatable token
+    bool        penalize_nl           = false;     // consider newlines as a repeatable token
 
     std::vector<llama_sampler_type> samplers_sequence = {
         llama_sampler_type::TOP_K,
@@ -131,12 +131,14 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
-// returns the probability that token of given id will be sampled
-llama_token_data_array llama_sampling_probability_distribution(
+// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
+llama_token_data_array llama_sampling_prepare(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,
         struct llama_context * ctx_cfg,
-        int idx = 0);
+        int idx = 0,
+        bool apply_grammar = true,
+        std::vector<float> * original_logits = nullptr);
 
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,

convert-hf-to-gguf.py

@@ -93,31 +93,42 @@ class Model(ABC):
 
         if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
+            print(f"gguf: context length = {n_ctx}")
 
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         self.gguf_writer.add_embedding_length(n_embd)
+        print(f"gguf: embedding length = {n_embd}")
 
         if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
+            print(f"gguf: feed forward length = {n_ff}")
 
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         self.gguf_writer.add_head_count(n_head)
+        print(f"gguf: head count = {n_head}")
 
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
+            print(f"gguf: key-value head count = {n_head_kv}")
 
         if (rope_theta := self.hparams.get("rope_theta")) is not None:
             self.gguf_writer.add_rope_freq_base(rope_theta)
+            print(f"gguf: rope theta = {rope_theta}")
         if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+            print(f"gguf: rms norm epsilon = {f_rms_eps}")
         if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+            print(f"gguf: layer norm epsilon = {f_norm_eps}")
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
+            print(f"gguf: expert count = {n_experts}")
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
+            print(f"gguf: experts used count = {n_experts_used}")
 
         self.gguf_writer.add_file_type(self.ftype)
+        print(f"gguf: file type = {self.ftype}")
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -1051,6 +1062,21 @@ class MixtralModel(Model):
         self._set_vocab_sentencepiece()
 
 
+@Model.register("GrokForCausalLM")
+class GrokModel(Model):
+    model_arch = gguf.MODEL_ARCH.GROK
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_name("Grok")
+
+
 @Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
     model_arch = gguf.MODEL_ARCH.MINICPM
@@ -1634,7 +1660,7 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
-@Model.register("BertModel")
+@Model.register("BertModel", "CamembertModel")
 class BertModel(Model):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -1965,6 +1991,23 @@ class MambaModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
+@Model.register("CohereForCausalLM")
+class CommandR2Model(Model):
+    model_arch = gguf.MODEL_ARCH.COMMAND_R
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # max_position_embeddings = 8192 in config.json but model was actually
+        # trained on 128k context length
+        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+
 ###### CONVERSION LOGIC ######
 
 

convert.py

@@ -332,6 +332,9 @@ class Params:
 #
 
 class BpeVocab:
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
         if isinstance(self.bpe_tokenizer.get('model'), dict):
@@ -390,6 +393,9 @@ class BpeVocab:
 
 
 class SentencePieceVocab:
+    tokenizer_model = "llama"
+    name = "spm"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
         added_tokens: dict[str, int]
@@ -453,6 +459,9 @@ class SentencePieceVocab:
 
 
 class HfVocab:
+    tokenizer_model = "llama"
+    name = "hfft"
+
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
         try:
             from transformers import AutoTokenizer
@@ -553,7 +562,15 @@ class HfVocab:
         return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
+class NoVocab:
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"
 
 
 #
@@ -935,8 +952,10 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
+            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
         )
+    if isinstance(vocab, NoVocab):
+        return  # model has no vocab
 
     # Check for a vocab size mismatch
     if params.n_vocab == vocab.vocab_size:
@@ -977,6 +996,7 @@ class OutputFile:
             name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
+        self.gguf.add_vocab_size          (params.n_vocab)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -1013,21 +1033,9 @@ class OutputFile:
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
-    def handle_tokenizer_model(self, vocab: Vocab) -> str:
-        # Map the vocab types to the supported tokenizer models
-        tokenizer_model = {
-            SentencePieceVocab: "llama",
-            HfVocab: "llama",
-            BpeVocab: "gpt2",
-        }.get(type(vocab))
-
-        # Block if vocab type is not predefined
-        if tokenizer_model is None:
-            raise ValueError("Unknown vocab type: Not supported")
-
-        return tokenizer_model
-
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+        assert not isinstance(vocab, NoVocab)
+
         tokens = []
         scores = []
         toktypes = []
@@ -1043,11 +1051,8 @@ class OutputFile:
         return tokens, scores, toktypes
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
-        # Handle the tokenizer model
-        tokenizer_model = self.handle_tokenizer_model(vocab)
-
         # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(tokenizer_model)
+        self.gguf.add_tokenizer_model(vocab.tokenizer_model)
 
         # Extract model vocabulary for model conversion
         tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
@@ -1074,6 +1079,26 @@ class OutputFile:
     def write_tensor_info(self) -> None:
         self.gguf.write_ti_data_to_file()
 
+    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(
+                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            self.gguf.write_tensor_data(ndarray)
+
     def close(self) -> None:
         self.gguf.close()
 
@@ -1082,7 +1107,7 @@ class OutputFile:
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
         endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
     ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1120,8 +1145,11 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        of.add_meta_vocab(vocab)
-        of.add_meta_special_vocab(svocab)
+        if isinstance(vocab, NoVocab):
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
+        else:
+            of.add_meta_vocab(vocab)
+            of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1131,24 +1159,7 @@ class OutputFile:
         of.write_tensor_info()
 
         # tensor data
-        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
-        if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
-        else:
-            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
-
-        start = time.time()
-        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            elapsed = time.time() - start
-            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
-            padi = len(str(len(model)))
-            print(
-                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
-            of.gguf.write_tensor_data(ndarray)
+        of.write_tensor_data(ftype, model, concurrency)
 
         of.close()
 
@@ -1156,9 +1167,9 @@ class OutputFile:
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
     wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
-    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
+    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
         return GGMLFileType.AllF32
-    if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
+    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
         return GGMLFileType.MostlyF16
     if output_type_str == "q8_0":
         return GGMLFileType.MostlyQ8_0
@@ -1309,8 +1320,8 @@ class VocabFactory:
                 return vtype, path
         raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
 
-    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocabtype == "bpe"
+    def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
+        load_merges = vocab.name == "bpe"
         n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
         return gguf.SpecialVocab(
             model_parent_path,
@@ -1319,30 +1330,34 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
         vocab_type, path = self._select_file(vocab_types)
         print(f"Loading vocab file {path!r}, type {vocab_type!r}")
 
         added_tokens_path = path.parent / "added_tokens.json"
-        vocab: Vocab
         if vocab_type == "bpe":
-            vocab = BpeVocab(
+            return BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "spm":
-            vocab = SentencePieceVocab(
+        if vocab_type == "spm":
+            return SentencePieceVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "hfft":
-            vocab = HfVocab(
+        if vocab_type == "hfft":
+            return HfVocab(
                 path.parent, added_tokens_path if added_tokens_path.exists() else None
             )
+        raise ValueError(vocab_type)
+
+    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        vocab: Vocab
+        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+            vocab = NoVocab()
         else:
-            raise ValueError(vocab_type)
+            vocab = self._create_vocab_by_path(vocab_types)
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocab_type,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1380,6 +1395,7 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--no-vocab",     action="store_true",    help="store model without the vocab")
     parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
     parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
@@ -1392,6 +1408,10 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
+    if args.no_vocab:
+        if args.vocab_only:
+            raise ValueError("no need to specify --vocab-only if using --no-vocab")
+        args.vocab_type = "no_vocab"
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1442,7 +1462,7 @@ def main(args_in: list[str] | None = None) -> None:
         print(f"Wrote {outfile}")
         return
 
-    if model_plus.vocab is not None and args.vocab_dir is None:
+    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
         vocab = model_plus.vocab
 
     print(f"Vocab info: {vocab}")

examples/CMakeLists.txt

@@ -21,6 +21,7 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(finetune)
     add_subdirectory(gritlm)
+    add_subdirectory(gguf-split)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)

examples/batched/batched.cpp

@@ -48,6 +48,8 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    process_escapes(params.prompt);
+
     // init LLM
 
     llama_backend_init();
@@ -78,7 +80,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx = n_kv_req;
+    ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
     ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;

examples/convert-llama2c-to-ggml/README.md

@@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface.
 
 `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
+Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
+
 Now you can use the model with a command like:
 
 `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

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

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "log.h"
 
 #include <unordered_map>
 #include <vector>
@@ -78,111 +79,101 @@ typedef struct {
 
 struct TransformerWeights {
     // token embedding table
-    float* token_embedding_table;    // (vocab_size, dim)
+    std::vector<float> token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
-    float* rms_att_weight; // (layer, dim) rmsnorm weights
-    float* rms_ffn_weight; // (layer, dim)
+    std::vector<float> rms_att_weight; // (layer, dim) rmsnorm weights
+    std::vector<float> rms_ffn_weight; // (layer, dim)
     // weights for matmuls
-    float* wq; // (layer, dim, dim)
-    float* wk; // (layer, dim, dim)
-    float* wv; // (layer, dim, dim)
-    float* wo; // (layer, dim, dim)
+    std::vector<float> wq; // (layer, dim, dim)
+    std::vector<float> wk; // (layer, dim, dim)
+    std::vector<float> wv; // (layer, dim, dim)
+    std::vector<float> wo; // (layer, dim, dim)
     // weights for ffn
-    float* w1; // (layer, hidden_dim, dim)
-    float* w2; // (layer, dim, hidden_dim)
-    float* w3; // (layer, hidden_dim, dim)
+    std::vector<float> w1; // (layer, hidden_dim, dim)
+    std::vector<float> w2; // (layer, dim, hidden_dim)
+    std::vector<float> w3; // (layer, hidden_dim, dim)
     // final rmsnorm
-    float* rms_final_weight; // (dim,)
+    std::vector<float> rms_final_weight; // (dim,)
     // freq_cis for RoPE relatively positional embeddings
-    // float* freq_cis_real; // (seq_len, dim/2)
-    // float* freq_cis_imag; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_real; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    float* wcls;
-
-    ~TransformerWeights() {
-        delete[] token_embedding_table;
-        delete[] rms_att_weight;
-        delete[] rms_ffn_weight;
-        delete[] wq;
-        delete[] wk;
-        delete[] wv;
-        delete[] wo;
-        delete[] w1;
-        delete[] w2;
-        delete[] w3;
-        delete[] rms_final_weight;
-        delete[] wcls;
-    }
+    std::vector<float> wcls;
 };
 
-static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
-    // we calloc instead of malloc to keep valgrind happy
-    w->token_embedding_table = new float[p->vocab_size * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) {
+    const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
+    try {
+        w->token_embedding_table.resize(p->vocab_size * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
-    w->rms_att_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        w->rms_att_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
-    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        w->rms_ffn_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
-    w->wq = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wq.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->wk = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wv = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wo = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wo.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->rms_final_weight = new float[p->dim]();
-    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        w->rms_final_weight.resize(p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
-    if (shared_weights) {
-        w->wcls = NULL;
-    } else {
-        w->wcls = new float[p->vocab_size * p->dim]();
-        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        if (shared_weights) {
+            w->wcls = {};
+        } else {
+            w->wcls.resize(p->vocab_size * p->dim);
+            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        }
+    }
+    catch (std::length_error &) {
+        die("Invalid configuration. Failed to allocate memory for weights");
     }
 }
 
-static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
-    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
-    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
-    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) {
+    if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1;
+    if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1;
+    if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1;
+    if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1;
+    if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1;
+    if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1;
+    if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1;
+    if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1;
+    if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1;
+    if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1;
+    if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1;
 
     // Skip freq_cis_real & freq_cis_imag
     int head_size = p->dim / p->n_heads;
     fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
 
-    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1;
 
     // Check we didn't forget to read anything
     auto curr = ftell(f);
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    printf("----- Quick print of first of the weight vales of all the variables\n");
-    printf("%f\n", w->token_embedding_table[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    printf("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    printf("%f\n", w->wq[0]);
-    printf("%f\n", w->wk[0]);
-    printf("%f\n", w->wv[0]);
-    printf("%f\n", w->wo[0]);
-    printf("%f\n", w->w1[0]);
-    printf("%f\n", w->w2[0]);
-    printf("%f\n", w->w3[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    if (w->wcls) printf("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -225,14 +216,16 @@ struct llama_vocab {
 };
 
 struct my_llama_hparams {
-    uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
-    uint32_t n_ff    = 11008;
-    uint32_t n_mult  = 4;
-    uint32_t n_head  = 32;
-    uint32_t n_layer = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_vocab   = 32000;
+    uint32_t n_ctx     = 512;   // this is provided as user input?
+    uint32_t n_embd    = 4096;
+    uint32_t n_ff      = 11008;
+    uint32_t n_mult    = 4;
+    uint32_t n_head    = 32;
+    uint32_t n_head_kv = 32;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
     bool operator!=(const my_llama_hparams& other) const {
         return memcmp(this, &other, sizeof(my_llama_hparams));
     }
@@ -325,14 +318,30 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
-    printf("%s: n_head:  %u\n", __func__, params->n_head);
-    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
-    printf("%s: n_layer: %u\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+}
+
+static void print_tensor_info(const struct ggml_context * ctx) {
+    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        LOG("%s: Allocating ", __func__);
+        int64_t total = 1;
+        int i = 0;
+        for (; i < ggml_n_dims(t); ++i) {
+            if (i > 0) LOG("x ");
+            LOG("[%" PRId64 "] ", t->ne[i]);
+            total *= t->ne[i];
+        }
+        if (i > 1) LOG("= [%" PRId64 "] ", total);
+        LOG("float space for %s\n", ggml_get_name(t));
+    }
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
+    const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv;
+
     const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
@@ -350,25 +361,8 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
-
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
-
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
-
-    // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");
@@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) {
         layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
         layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
         layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
@@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) {
         ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
         ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
     }
+
+    print_tensor_info(ctx);
 }
 
 static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
@@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
 static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
-        printf(" %f", p);
+        LOG(" %f", p);
     }
-    printf("\n");
+    LOG("\n");
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
@@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
+            LOG(" %.2f", p);
         }
-        printf("\n");
+        LOG("\n");
     }
 }
 
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
@@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
 
         const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+        if (n_vocab != static_cast<uint32_t>(config->vocab_size)) {
+            die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size);
+        }
 
         vocab->id_to_token.resize(n_vocab);
 
@@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 }
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
-    int ct;
-    switch (ggml_n_dims(gg_weights)) {
-        case 1:
-            ct = 0;
-            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
-                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
-                *ptr = karpathy_weights[ct];
-                ct++;
-            }
-            break;
-        case 2:
-            ct = 0;
-            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
-                    *ptr = karpathy_weights[ct];
-                    ct++;
-                }
-            }
-            break;
-        case 3:
-            ct = 0;
-            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
-                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
-                        *ptr = karpathy_weights[ct];
-                        ct++;
-                    }
-                }
-            }
-            break;
+    int size = 1;
+    for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) {
+        size *= gg_weights->ne[dim];
+    }
+    for (int ct = 0; ct < size; ++ct) {
+        int64_t i0 = 0; int64_t i1 = 0;
+        int64_t i2 = 0; int64_t i3 = 0;
+        ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3);
+        ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]);
     }
 }
 
@@ -679,16 +635,18 @@ static void save_as_llama_model(
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
-    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data());
+    convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data());
 
-    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data());
     //print_row(model->norm, 0);
 
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     int n_ff = model->hparams.n_ff;
 
+    const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv;
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
@@ -697,9 +655,10 @@ static void save_as_llama_model(
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
         convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
         convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length/n_multiqueries]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length/n_multiqueries]);
 
         convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
         convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
@@ -736,8 +695,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
-    // n_head_kv is optional, default to n_head
-    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv);
     gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
     gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
     gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
@@ -789,12 +748,12 @@ static void save_as_llama_model(
 
 static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_vocab_model          = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
-    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+    params.fn_train_data           = "shakespeare.txt";
+    params.fn_checkpoint_in        = "checkpoint.bin";
+    params.fn_checkpoint_out       = "checkpoint.bin";
+    params.fn_model_out            = "ggml-checkpoint-f32.bin";
 
     params.seed       =   -1;
 
@@ -829,8 +788,8 @@ static struct train_params get_default_train_params() {
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
 
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
+    params.mem_model_gb    = 2;
+    params.mem_compute_gb  = 24;
     params.mem_compute0_gb = 8;
     params.mem_compute1_gb = 2;
 
@@ -916,19 +875,30 @@ int main(int argc, char ** argv) {
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        FILE *file = fopen(params.fn_llama2c_model, "rb");
-        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        FILE *file = fopen(params.fn_llama2c_model, "r");
+        if (!file) {
+            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            return 1;
+        }
         // read in the config header
-        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        if (fread(&config, sizeof(Config), 1, file) != 1) {
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         auto shared_weights = config.vocab_size > 0;
         config.vocab_size = abs(config.vocab_size);
 
         // read in the Transformer weights
-        malloc_weights(&weights, &config, shared_weights);
-        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        alloc_weights(&weights, &config, shared_weights);
+        if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
+            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         fclose(file);
     }
 
@@ -936,15 +906,18 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
-    model.hparams.n_embd  = config.dim; //params.n_embd;
-    model.hparams.n_ff    = config.hidden_dim;
-    model.hparams.n_mult  = 32;//params.n_mult;
-    model.hparams.n_head  = config.n_heads; //params.n_head;
-    model.hparams.n_layer = config.n_layers; //params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx     = params.n_ctx;
+    model.hparams.n_embd    = config.dim; //params.n_embd;
+    model.hparams.n_ff      = config.hidden_dim;
+    model.hparams.n_mult    = 32;//params.n_mult;
+    model.hparams.n_head    = config.n_heads; //params.n_head;
+    model.hparams.n_head_kv = config.n_kv_heads;
+    model.hparams.n_layer   = config.n_layers; //params.n_layer;
+    model.hparams.n_rot     = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+
     print_params(&model.hparams);
+
     struct ggml_init_params lcparams;
     lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
     lcparams.mem_buffer = NULL;
@@ -956,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/embedding/embedding.cpp

@@ -112,13 +112,20 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true);
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
         if (inp.size() > n_batch) {
             inp.resize(n_batch);
         }
         inputs.push_back(inp);
     }
 
+    // add eos if not present
+    for (auto & inp : inputs) {
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+    }
+
     // tokenization stats
     if (params.verbose_prompt) {
         for (int i = 0; i < (int) inputs.size(); i++) {
@@ -172,7 +179,7 @@ int main(int argc, char ** argv) {
     for (int j = 0; j < n_prompts; j++) {
         fprintf(stdout, "embedding %d: ", j);
         for (int i = 0; i < std::min(16, n_embd); i++) {
-            fprintf(stdout, "%f ", emb[j * n_embd + i]);
+            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
         }
         fprintf(stdout, "\n");
     }

examples/gguf-split/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET gguf-split)
+add_executable(${TARGET} gguf-split.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/gguf-split/README.md

@@ -0,0 +1,9 @@
+## GGUF split Example
+
+CLI to split / merge GGUF files.
+
+**Command line options:**
+
+- `--split`: split GGUF to multiple GGUF, default operation.
+- `--split-max-tensors`: maximum tensors in each split: default(128)
+- `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -0,0 +1,468 @@
+#include "llama.h"
+#include "common.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include <stdio.h>
+#include <string.h>
+#include <climits>
+#include <stdexcept>
+
+#if defined(_WIN32)
+    #include <windows.h>
+    #ifndef PATH_MAX
+        #define PATH_MAX MAX_PATH
+    #endif
+    #include <io.h>
+#endif
+
+enum split_operation : uint8_t {
+    SPLIT_OP_SPLIT,
+    SPLIT_OP_MERGE,
+};
+
+struct split_params {
+    split_operation operation = SPLIT_OP_SPLIT;
+    int n_split_tensors = 128;
+    std::string input;
+    std::string output;
+};
+
+static void split_print_usage(const char * executable) {
+    const split_params default_params;
+    printf("\n");
+    printf("usage: %s [options] GGUF_IN GGUF_OUT\n", executable);
+    printf("\n");
+    printf("Apply a GGUF operation on IN to OUT.");
+    printf("\n");
+    printf("options:\n");
+    printf("  -h, --help            show this help message and exit\n");
+    printf("  --version             show version and build info\n");
+    printf("  --split               split GGUF to multiple GGUF (default)\n");
+    printf("  --split-max-tensors   max tensors in each split: default(%d)\n", default_params.n_split_tensors);
+    printf("  --merge               merge multiple GGUF to a single GGUF\n");
+    printf("\n");
+}
+
+static bool split_params_parse_ex(int argc, const char ** argv, split_params & params) {
+    std::string arg;
+    const std::string arg_prefix = "--";
+    bool invalid_param = false;
+
+    int arg_idx = 1;
+    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
+        arg = argv[arg_idx];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        bool arg_found = false;
+        if (arg == "-h" || arg == "--help") {
+            split_print_usage(argv[0]);
+            exit(0);
+        }
+        if (arg == "--version") {
+            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            exit(0);
+        }
+
+        if (arg == "--merge") {
+            arg_found = true;
+            params.operation = SPLIT_OP_MERGE;
+        }
+        if (arg == "--split") {
+            arg_found = true;
+            params.operation = SPLIT_OP_SPLIT;
+        }
+        if (arg == "--split-max-tensors") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            params.n_split_tensors = atoi(argv[arg_idx]);
+        }
+
+        if (!arg_found) {
+            throw std::invalid_argument("error: unknown argument: " + arg);
+        }
+    }
+
+    if (invalid_param) {
+        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
+    }
+
+    if (argc - arg_idx < 2) {
+        printf("%s: bad arguments\n", argv[0]);
+        split_print_usage(argv[0]);
+        return false;
+    }
+
+    params.input = argv[arg_idx++];
+    params.output = argv[arg_idx++];
+
+    return true;
+}
+
+static bool split_params_parse(int argc, const char ** argv, split_params & params) {
+    bool result = true;
+    try {
+        if (!split_params_parse_ex(argc, argv, params)) {
+            split_print_usage(argv[0]);
+            exit(EXIT_FAILURE);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        split_print_usage(argv[0]);
+        exit(EXIT_FAILURE);
+    }
+    return result;
+}
+
+static void zeros(std::ofstream & file, size_t n) {
+    char zero = 0;
+    for (size_t i = 0; i < n; ++i) {
+        file.write(&zero, 1);
+    }
+}
+
+struct split_strategy {
+    const split_params params;
+    std::ifstream & f_input;
+    struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_meta = NULL;
+    const int n_tensors;
+
+    const int n_split;
+    int i_split = 0;
+
+    int i_tensor = 0;
+
+    std::vector<uint8_t> read_data;
+
+    struct gguf_context * ctx_out;
+    std::ofstream fout;
+
+    split_strategy(const split_params & params,
+            std::ifstream & f_input,
+            struct gguf_context * ctx_gguf,
+            struct ggml_context * ctx_meta) :
+        params(params),
+        f_input(f_input),
+        ctx_gguf(ctx_gguf),
+        ctx_meta(ctx_meta),
+        n_tensors(gguf_get_n_tensors(ctx_gguf)),
+        n_split(std::ceil(1. * n_tensors / params.n_split_tensors)) {
+        }
+
+    bool should_split() const {
+        return i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
+    }
+
+    void split_start() {
+        ctx_out = gguf_init_empty();
+
+        // Save all metadata in first split only
+        if (i_split == 0) {
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
+        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, n_split);
+        gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
+
+        // populate the original tensors, so we get an initial metadata
+        for (int i = i_split * params.n_split_tensors; i < n_tensors && i < (i_split + 1) * params.n_split_tensors; ++i) {
+            struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+            gguf_add_tensor(ctx_out, meta);
+        }
+
+        char split_path[PATH_MAX] = {0};
+        llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
+
+        fprintf(stderr, "%s: %s ...", __func__, split_path);
+        fout = std::ofstream(split_path, std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+        auto meta_size = gguf_get_meta_size(ctx_out);
+
+        // placeholder for the meta data
+        ::zeros(fout, meta_size);
+
+        i_split++;
+    }
+
+    void next_tensor() {
+        const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+        struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+        auto n_bytes = ggml_nbytes(t);
+
+        if (read_data.size() < n_bytes) {
+            read_data.resize(n_bytes);
+        }
+
+        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
+        f_input.seekg(offset);
+        f_input.read((char *)read_data.data(), n_bytes);
+
+        t->data = read_data.data();
+
+        // write tensor data + padding
+        fout.write((const char *)t->data, n_bytes);
+        zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+
+        i_tensor++;
+    }
+
+    void split_end() {
+        // go back to beginning of file and write the updated metadata
+        fout.seekp(0);
+        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+        gguf_get_meta_data(ctx_out, data.data());
+        fout.write((const char *)data.data(), data.size());
+
+        fout.close();
+        gguf_free(ctx_out);
+
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+};
+
+static void gguf_split(const split_params & split_params) {
+    struct ggml_context * ctx_meta = NULL;
+
+    struct gguf_init_params params = {
+        /*.no_alloc = */ true,
+        /*.ctx      = */ &ctx_meta,
+    };
+
+    std::ifstream f_input(split_params.input.c_str(), std::ios::binary);
+    if (!f_input.is_open()) {
+        fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(EXIT_FAILURE);
+    }
+
+    auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+        exit(EXIT_FAILURE);
+    }
+
+    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+
+    char first_split_path[PATH_MAX] = {0};
+    llama_split_path(first_split_path, sizeof(first_split_path),
+                     split_params.output.c_str(), strategy.i_split, strategy.n_split);
+    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
+            __func__, split_params.input.c_str(),
+            first_split_path,
+            split_params.n_split_tensors);
+
+    strategy.split_start();
+
+    while (strategy.i_tensor < strategy.n_tensors) {
+        strategy.next_tensor();
+        if (strategy.should_split()) {
+            strategy.split_end();
+            strategy.split_start();
+        }
+    }
+    strategy.split_end();
+
+    gguf_free(ctx_gguf);
+    f_input.close();
+
+    fprintf(stderr, "%s: %d gguf split written with a total of %d tensors.\n",
+            __func__, strategy.n_split, strategy.n_tensors);
+}
+
+static void gguf_merge(const split_params & split_params) {
+    fprintf(stderr, "%s: %s -> %s\n",
+            __func__, split_params.input.c_str(),
+            split_params.output.c_str());
+    int n_split = 1;
+    int total_tensors = 0;
+
+    auto * ctx_out = gguf_init_empty();
+    std::ofstream fout(split_params.output.c_str(), std::ios::binary);
+    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+    std::vector<uint8_t> read_data;
+    std::vector<ggml_context *> ctx_metas;
+    std::vector<gguf_context *> ctx_ggufs;
+
+    char split_path[PATH_MAX] = {0};
+    strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1);
+    char split_prefix[PATH_MAX] = {0};
+
+    // First pass to find KV and tensors metadata
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        struct ggml_context * ctx_meta = NULL;
+
+        struct gguf_init_params params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ &ctx_meta,
+        };
+
+        if (i_split > 0) {
+            llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        }
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path);
+
+        auto * ctx_gguf = gguf_init_from_file(split_path, params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+            exit(EXIT_FAILURE);
+        }
+        ctx_ggufs.push_back(ctx_gguf);
+        ctx_metas.push_back(ctx_meta);
+
+        if (i_split == 0) {
+            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+            if (key_n_split < 0) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain %s metadata\n",
+                        __func__,
+                        LLM_KV_SPLIT_COUNT);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+            if (n_split < 1) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain a valid split count %d\n",
+                        __func__,
+                        n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            // Verify the file naming and extract split_prefix
+            if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d"
+                                " n_split=%d\n", __func__,
+                        split_path, i_split, n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
+            }
+
+            // Do not trigger merge if we try to merge again the output
+            gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0);
+
+            // Set metadata from the first split
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+            gguf_add_tensor(ctx_out, t);
+        }
+        total_tensors += n_tensors;
+
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+
+    // placeholder for the meta data
+    {
+        auto meta_size = gguf_get_meta_size(ctx_out);
+        ::zeros(fout, meta_size);
+    }
+
+    // Write tensors data
+    for (int i_split = 0; i_split < n_split; i_split++) {
+        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        std::ifstream f_input(split_path, std::ios::binary);
+        if (!f_input.is_open()) {
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_path);
+            for (uint32_t i = 0; i < ctx_ggufs.size(); i++) {
+                gguf_free(ctx_ggufs[i]);
+                ggml_free(ctx_metas[i]);
+            }
+            gguf_free(ctx_out);
+            fout.close();
+            exit(EXIT_FAILURE);
+        }
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
+
+        auto * ctx_gguf = ctx_ggufs[i_split];
+        auto * ctx_meta = ctx_metas[i_split];
+
+        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
+        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
+            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+
+            auto n_bytes = ggml_nbytes(t);
+
+            if (read_data.size() < n_bytes) {
+                read_data.resize(n_bytes);
+            }
+
+            auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
+            f_input.seekg(offset);
+            f_input.read((char *)read_data.data(), n_bytes);
+
+            // write tensor data + padding
+            fout.write((const char *)read_data.data(), n_bytes);
+            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+        }
+
+        gguf_free(ctx_gguf);
+        ggml_free(ctx_meta);
+        f_input.close();
+        fprintf(stderr, "\033[3Ddone\n");
+    }
+
+    {
+        // go back to beginning of file and write the updated metadata
+        fout.seekp(0);
+        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+        gguf_get_meta_data(ctx_out, data.data());
+        fout.write((const char *)data.data(), data.size());
+
+        fout.close();
+        gguf_free(ctx_out);
+    }
+
+    fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
+            __func__, split_params.output.c_str(), n_split, total_tensors);
+}
+
+int main(int argc, const char ** argv) {
+    if (argc < 3) {
+        split_print_usage(argv[0]);
+    }
+
+    split_params params;
+    split_params_parse(argc, argv, params);
+
+    switch (params.operation) {
+        case SPLIT_OP_SPLIT: gguf_split(params);
+            break;
+        case SPLIT_OP_MERGE: gguf_merge(params);
+            break;
+        default: split_print_usage(argv[0]);
+            exit(EXIT_FAILURE);
+    }
+
+    return 0;
+}

examples/gguf/gguf.cpp

@@ -211,6 +211,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
                         fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
+                        gguf_free(ctx);
                         return false;
                     }
                 }

examples/gritlm/README.md

@@ -0,0 +1,62 @@
+## Generative Representational Instruction Tuning (GRIT) Example
+[gritlm] a model which can generate embeddings as well as "normal" text
+generation depending on the instructions in the prompt.
+
+* Paper: https://arxiv.org/pdf/2402.09906.pdf
+
+### Retrieval-Augmented Generation (RAG) use case
+One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
+that we take documents that we want to use as context, to ground the large
+language model (LLM), and we create token embeddings for them. We then store
+these token embeddings in a vector database.
+
+When we perform a query, prompt the LLM, we will first create token embeddings
+for the query and then search the vector database to retrieve the most
+similar vectors, and return those documents so they can be passed to the LLM as
+context. Then the query and the context will be passed to the LLM which will
+have to _again_ create token embeddings for the query. But because gritlm is used
+the first query can be cached and the second query tokenization generation does
+not have to be performed at all.
+
+### Running the example
+Download a Grit model:
+```console
+$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
+```
+
+Run the example using the downloaded model:
+```console
+$ ./gritlm -m gritlm-7b_q4_1.gguf
+
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
+Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
+Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+When shadows lurk and ghosts do roam,
+And darkness reigns, a fearsome sight.
+
+Thou didst set out, with heart aglow,
+To conquer this mountain, so high,
+And reach the summit, where the stars do glow,
+And the moon shines bright, up in the sky.
+
+Through the mist and fog, thou didst press on,
+With steadfast courage, and a steadfast will,
+Through the darkness, thou didst not be gone,
+But didst climb on, with a steadfast skill.
+
+At last, thou didst reach the summit's crest,
+And gazed upon the world below,
+And saw the beauty of the night's best,
+And felt the peace, that only nature knows.
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+Thou art a hero, in the eyes of all,
+For thou didst conquer this mountain, so bright.
+```
+
+[gritlm]: https://github.com/ContextualAI/gritlm

examples/imatrix/imatrix.cpp

@@ -56,13 +56,31 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
 
+    std::string wname;
+    {
+        // remove any prefix and suffixes from the name
+        // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+        const char * p = strchr(src0->name, '#');
+        if (p != NULL) {
+            p = p + 1;
+            const char * q = strchr(p, '#');
+            if (q != NULL) {
+                wname = std::string(p, q - p);
+            } else {
+                wname = p;
+            }
+        } else {
+            wname = src0->name;
+        }
+    }
+
     // when ask is true, the scheduler wants to know if we are interested in data from this tensor
     // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
     if (ask) {
         if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
         if (t->op != GGML_OP_MUL_MAT) return false;
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
         return true;
     }
 
@@ -94,12 +112,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         // this is necessary to guarantee equal number of "ncall" for each tensor
         for (int ex = 0; ex < n_as; ++ex) {
             src0 = t->src[2 + ex];
-            auto& e = m_stats[src0->name];
+            auto& e = m_stats[wname];
             if (e.values.empty()) {
                 e.values.resize(src1->ne[0], 0);
             }
             else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
                 exit(1); //GGML_ASSERT(false);
             }
             // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
@@ -107,7 +125,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
             ++e.ncall;
             if (m_params.verbosity > 1) {
-                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
             }
             for (int row = 0; row < (int)src1->ne[1]; ++row) {
                 const int excur = m_ids[row*n_as + idx];
@@ -129,17 +147,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             }
         }
     } else {
-        auto& e = m_stats[src0->name];
+        auto& e = m_stats[wname];
         if (e.values.empty()) {
             e.values.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
             exit(1); //GGML_ASSERT(false);
         }
         ++e.ncall;
         if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
         }
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];

examples/json-schema-pydantic-example.py

@@ -0,0 +1,74 @@
+# Usage:
+#! ./server -m some-model.gguf &
+#! pip install pydantic
+#! python json-schema-pydantic-example.py
+
+from pydantic import BaseModel, TypeAdapter
+from annotated_types import MinLen
+from typing import Annotated, List, Optional
+import json, requests
+
+if True:
+
+    def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs):
+        '''
+        Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support
+        (llama.cpp server, llama-cpp-python, Anyscale / Together...)
+
+        The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below)
+        '''
+        if response_model:
+            type_adapter = TypeAdapter(response_model)
+            schema = type_adapter.json_schema()
+            messages = [{
+                "role": "system",
+                "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}"
+            }] + messages
+            response_format={"type": "json_object", "schema": schema}
+
+        data = requests.post(endpoint, headers={"Content-Type": "application/json"},
+                             json=dict(messages=messages, response_format=response_format, **kwargs)).json()
+        if 'error' in data:
+            raise Exception(data['error']['message'])
+
+        content = data["choices"][0]["message"]["content"]
+        return type_adapter.validate_json(content) if type_adapter else content
+
+else:
+
+    # This alternative branch uses Instructor + OpenAI client lib.
+    # Instructor support streamed iterable responses, retry & more.
+    # (see https://python.useinstructor.com/)
+    #! pip install instructor openai
+    import instructor, openai
+    client = instructor.patch(
+        openai.OpenAI(api_key="123", base_url="http://localhost:8080"),
+        mode=instructor.Mode.JSON_SCHEMA)
+    create_completion = client.chat.completions.create
+
+
+if __name__ == '__main__':
+
+    class QAPair(BaseModel):
+        question: str
+        concise_answer: str
+        justification: str
+
+    class PyramidalSummary(BaseModel):
+        title: str
+        summary: str
+        question_answers: Annotated[List[QAPair], MinLen(2)]
+        sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]]
+
+    print("# Summary\n", create_completion(
+        model="...",
+        response_model=PyramidalSummary,
+        messages=[{
+            "role": "user",
+            "content": f"""
+                You are a highly efficient corporate document summarizer.
+                Create a pyramidal summary of an imaginary internal document about our company processes
+                (starting high-level, going down to each sub sections).
+                Keep questions short, and answers even shorter (trivia / quizz style).
+            """
+        }]))

examples/json-schema-to-grammar.py

@@ -1,8 +1,10 @@
 #!/usr/bin/env python3
 import argparse
+import itertools
 import json
 import re
 import sys
+from typing import Any, Dict, List, Set, Tuple, Union
 
 # whitespace is constrained to a single space char to prevent model "running away" in
 # whitespace. Also maybe improves generation quality?
@@ -12,26 +14,54 @@ PRIMITIVE_RULES = {
     'boolean': '("true" | "false") space',
     'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
     'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
+    'value'  : 'object | array | string | number | boolean',
+    'object' : '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
+    'array'  : '"[" space ( value ("," space value)* )? "]" space',
+    'uuid'   : '"\\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + ' "\\"" space',
     'string': r''' "\"" (
         [^"\\] |
         "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-      )* "\"" space ''',
+      )* "\"" space''',
     'null': '"null" space',
 }
+OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']
+
+# TODO: support "uri", "email" string formats
+DATE_RULES = {
+    'date'   : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )',
+    'time'   : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )',
+    'date-time': 'date "T" time',
+    'date-string': '"\\"" date "\\"" space',
+    'time-string': '"\\"" time "\\"" space',
+    'date-time-string': '"\\"" date-time "\\"" space',
+}
+
+RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
 GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}
+GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
 
+NON_LITERAL_SET = set('|.()[]{}*+?')
+ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
+
+DATE_PATTERN = '[0-9]{4}-(0[1-9]|1[0-2])-([0-2][0-9]|3[0-1])'
+TIME_PATTERN = '([01][0-9]|2[0-3])(:[0-5][0-9]){2}(\\.[0-9]{1,3})?(Z|[+-](([01][0-9]|2[0-3]):[0-5][0-9]))' # Cap millisecond precision w/ 3 digits
 
 class SchemaConverter:
-    def __init__(self, prop_order):
+    def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
         self._prop_order = prop_order
+        self._allow_fetch = allow_fetch
+        self._dotall = dotall
+        self._raw_pattern = raw_pattern
         self._rules = {'space': SPACE_RULE}
+        self._refs = {}
+        self._refs_being_resolved = set()
 
     def _format_literal(self, literal):
         escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal
         )
         return f'"{escaped}"'
 
@@ -41,78 +71,420 @@ class SchemaConverter:
             key = esc_name
         else:
             i = 0
-            while f'{esc_name}{i}' in self._rules:
+            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
                 i += 1
             key = f'{esc_name}{i}'
         self._rules[key] = rule
         return key
 
+    def resolve_refs(self, schema: dict, url: str):
+        '''
+            Resolves all $ref fields in the given schema, fetching any remote schemas,
+            replacing $ref with absolute reference URL and populating self._refs with the
+            respective referenced (sub)schema dictionaries.
+        '''
+        def visit(n: dict):
+            if isinstance(n, list):
+                return [visit(x) for x in n]
+            elif isinstance(n, dict):
+                ref = n.get('$ref')
+                if ref is not None and ref not in self._refs:
+                    if ref.startswith('https://'):
+                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
+                        import requests
+
+                        frag_split = ref.split('#')
+                        base_url = frag_split[0]
+
+                        target = self._refs.get(base_url)
+                        if target is None:
+                            target = self.resolve_refs(requests.get(ref).json(), base_url)
+                            self._refs[base_url] = target
+
+                        if len(frag_split) == 1 or frag_split[-1] == '':
+                            return target
+                    elif ref.startswith('#/'):
+                        target = schema
+                        ref = f'{url}{ref}'
+                        n['$ref'] = ref
+                    else:
+                        raise ValueError(f'Unsupported ref {ref}')
+
+                    for sel in ref.split('#')[-1].split('/')[1:]:
+                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                        target = target[sel]
+
+                    self._refs[ref] = target
+                else:
+                    for v in n.values():
+                        visit(v)
+
+            return n
+        return visit(schema)
+
+    def _generate_union_rule(self, name, alt_schemas):
+        return ' | '.join((
+            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
+            for i, alt_schema in enumerate(alt_schemas)
+        ))
+
+    def _visit_pattern(self, pattern, name):
+        '''
+            Transforms a regular expression pattern into a GBNF rule.
+
+            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
+            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
+
+            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
+
+            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
+            we define sub-rules to keep the output lean.
+        '''
+
+        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
+        pattern = pattern[1:-1]
+        sub_rule_ids = {}
+
+        i = 0
+        length = len(pattern)
+
+        def to_rule(s: Tuple[str, bool]) -> str:
+            (txt, is_literal) = s
+            return "\"" + txt + "\"" if is_literal else txt
+
+        def transform() -> Tuple[str, bool]:
+            '''
+                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
+            '''
+            nonlocal i
+            nonlocal pattern
+            nonlocal sub_rule_ids
+
+            start = i
+            # For each component of this sequence, store its string representation and whether it's a literal.
+            # We only need a flat structure here to apply repetition operators to the last item, and
+            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
+            # (GBNF's syntax is luckily very close to regular expressions!)
+            seq: list[Tuple[str, bool]] = []
+
+            def get_dot():
+                if self._dotall:
+                    rule = '[\\U00000000-\\U0010FFFF]'
+                else:
+                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
+                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
+                return self._add_rule(f'dot', rule)
+
+            def join_seq():
+                nonlocal seq
+                ret = []
+                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
+                    if is_literal:
+                        ret.append((''.join(x[0] for x in g), True))
+                    else:
+                        ret.extend(g)
+                if len(ret) == 1:
+                    return ret[0]
+                return (' '.join(to_rule(x) for x in seq), False)
+
+            while i < length:
+                c = pattern[i]
+                if c == '.':
+                    seq.append((get_dot(), False))
+                    i += 1
+                elif c == '(':
+                    i += 1
+                    if i < length:
+                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
+                    seq.append((f'({to_rule(transform())})', False))
+                elif c == ')':
+                    i += 1
+                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
+                    return join_seq()
+                elif c == '[':
+                    square_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != ']':
+                        if pattern[i] == '\\':
+                            square_brackets += pattern[i:i+2]
+                            i += 2
+                        else:
+                            square_brackets += pattern[i]
+                            i += 1
+                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    square_brackets += ']'
+                    i += 1
+                    seq.append((square_brackets, False))
+                elif c == '|':
+                    seq.append(('|', False))
+                    i += 1
+                elif c in ('*', '+', '?'):
+                    seq[-1] = (to_rule(seq[-1]) + c, False)
+                    i += 1
+                elif c == '{':
+                    curly_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != '}':
+                        curly_brackets += pattern[i]
+                        i += 1
+                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    curly_brackets += '}'
+                    i += 1
+                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
+                    min_times = 0
+                    max_times = None
+                    try:
+                        if len(nums) == 1:
+                            min_times = int(nums[0])
+                            max_times = min_times
+                        else:
+                            assert len(nums) == 2
+                            min_times = int(nums[0]) if nums[0] else 0
+                            max_times = int(nums[1]) if nums[1] else None
+                    except ValueError:
+                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
+
+                    (sub, sub_is_literal) = seq[-1]
+
+                    if min_times == 0 and max_times is None:
+                        seq[-1] = (f'{sub}*', False)
+                    elif min_times == 0 and max_times == 1:
+                        seq[-1] = (f'{sub}?', False)
+                    elif min_times == 1 and max_times is None:
+                        seq[-1] = (f'{sub}+', False)
+                    else:
+                        if not sub_is_literal:
+                            id = sub_rule_ids.get(sub)
+                            if id is None:
+                                id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
+                                sub_rule_ids[sub] = id
+                            sub = id
+
+                        seq[-1] = (
+                            ' '.join(
+                                ([f'"{sub[1:-1] * min_times}"'] if sub_is_literal else [sub] * min_times) +
+                                ([f'{sub}?'] * (max_times - min_times) if max_times is not None else [f'{sub}*'])),
+                            False
+                        )
+                else:
+                    literal = ''
+                    while i < length:
+                        if pattern[i] == '\\' and i < length - 1:
+                            next = pattern[i + 1]
+                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
+                                i += 1
+                                literal += pattern[i]
+                                i += 1
+                            else:
+                                literal += pattern[i:i+2]
+                                i += 2
+                        elif pattern[i] == '"' and not self._raw_pattern:
+                            literal += '\\"'
+                            i += 1
+                        elif pattern[i] not in NON_LITERAL_SET and \
+                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
+                            literal += pattern[i]
+                            i += 1
+                        else:
+                            break
+                    if literal:
+                        seq.append((literal, True))
+
+            return join_seq()
+
+        return self._add_rule(
+            name,
+            to_rule(transform()) if self._raw_pattern \
+                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+
+
+    def _resolve_ref(self, ref):
+        ref_name = ref.split('/')[-1]
+        if ref_name not in self._rules and ref not in self._refs_being_resolved:
+            self._refs_being_resolved.add(ref)
+            resolved = self._refs[ref]
+            ref_name = self.visit(resolved, ref_name)
+            self._refs_being_resolved.remove(ref)
+        return ref_name
+
+    def _generate_constant_rule(self, value):
+        return self._format_literal(json.dumps(value))
+
     def visit(self, schema, name):
         schema_type = schema.get('type')
-        rule_name = name or 'root'
+        schema_format = schema.get('format')
+        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
 
-        if 'oneOf' in schema or 'anyOf' in schema:
-            rule = ' | '.join((
-                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
-                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
-            ))
-            return self._add_rule(rule_name, rule)
+        if (ref := schema.get('$ref')) is not None:
+            return self._add_rule(rule_name, self._resolve_ref(ref))
+
+        elif 'oneOf' in schema or 'anyOf' in schema:
+            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
+
+        elif isinstance(schema_type, list):
+            return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type]))
 
         elif 'const' in schema:
-            return self._add_rule(rule_name, self._format_literal(schema['const']))
+            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
 
         elif 'enum' in schema:
-            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
+            rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum']))
             return self._add_rule(rule_name, rule)
 
-        elif schema_type == 'object' and 'properties' in schema:
-            # TODO: `required` keyword
-            prop_order = self._prop_order
-            prop_pairs = sorted(
-                schema['properties'].items(),
-                # sort by position in prop_order (if specified) then by key
-                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
+        elif schema_type in (None, 'object') and \
+             ('properties' in schema or \
+              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
+            required = set(schema.get('required', []))
+            properties = list(schema.get('properties', {}).items())
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
+
+        elif schema_type in (None, 'object') and 'allOf' in schema:
+            required = set()
+            properties = []
+            hybrid_name = name
+            def add_component(comp_schema, is_required):
+                if (ref := comp_schema.get('$ref')) is not None:
+                    comp_schema = self._refs[ref]
+
+                if 'properties' in comp_schema:
+                    for prop_name, prop_schema in comp_schema['properties'].items():
+                        properties.append((prop_name, prop_schema))
+                        if is_required:
+                            required.add(prop_name)
+
+            for t in schema['allOf']:
+                if 'anyOf' in t:
+                    for tt in t['anyOf']:
+                        add_component(tt, is_required=False)
+                else:
+                    add_component(t, is_required=True)
+
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[]))
+
+        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
+            items = schema.get('items') or schema['prefixItems']
+            if isinstance(items, list):
+                return self._add_rule(
+                    rule_name,
+                    '"[" space ' +
+                    ' "," space '.join(
+                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
+                        for i, item in enumerate(items)) +
+                    ' "]" space')
+            else:
+                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
+                list_item_operator = f'( "," space {item_rule_name} )'
+                successive_items = ""
+                min_items = schema.get("minItems", 0)
+                max_items = schema.get("maxItems")
+                if min_items > 0:
+                    successive_items = list_item_operator * (min_items - 1)
+                    min_items -= 1
+                if max_items is not None and max_items > min_items:
+                    successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
+                else:
+                    successive_items += list_item_operator + "*"
+                if min_items == 0:
+                    rule = f'"[" space ( {item_rule_name} {successive_items} )? "]" space'
+                else:
+                    rule = f'"[" space {item_rule_name} {successive_items} "]" space'
+                return self._add_rule(rule_name, rule)
+
+        elif schema_type in (None, 'string') and 'pattern' in schema:
+            return self._visit_pattern(schema['pattern'], rule_name)
+
+        elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
+            return self._add_rule(
+                'root' if rule_name == 'root' else schema_format,
+                PRIMITIVE_RULES['uuid']
             )
 
-            rule = '"{" space'
-            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
-                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
-                if i > 0:
-                    rule += ' "," space'
-                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
-            rule += ' "}" space'
+        elif schema_type in (None, 'string') and schema_format in DATE_RULES:
+            for t, r in DATE_RULES.items():
+                self._add_rule(t, r)
+            return schema_format + '-string'
 
-            return self._add_rule(rule_name, rule)
-
-        elif schema_type == 'array' and 'items' in schema:
-            # TODO `prefixItems` keyword
-            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
-            return self._add_rule(rule_name, rule)
+        elif (schema_type == 'object') or (len(schema) == 0):
+            for n in OBJECT_RULE_NAMES:
+                self._add_rule(n, PRIMITIVE_RULES[n])
+            return self._add_rule(rule_name, 'object')
 
         else:
             assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
             return self._add_rule(
                 'root' if rule_name == 'root' else schema_type,
                 PRIMITIVE_RULES[schema_type]
             )
 
+    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
+        prop_order = self._prop_order
+        # sort by position in prop_order (if specified) then by original order
+        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
+
+        prop_kv_rule_names = {}
+        for prop_name, prop_schema in properties:
+            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+            prop_kv_rule_names[prop_name] = self._add_rule(
+                f'{name}{"-" if name else ""}{prop_name}-kv',
+                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
+            )
+        required_props = [k for k in sorted_props if k in required]
+        optional_props = [k for k in sorted_props if k not in required]
+
+        if additional_properties == True or isinstance(additional_properties, dict):
+            sub_name = f'{name}{"-" if name else ""}additional'
+            value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
+            prop_kv_rule_names["*"] = self._add_rule(
+                f'{sub_name}-kv',
+                self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+            )
+            optional_props.append("*")
+
+        rule = '"{" space '
+        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
+
+        if optional_props:
+            rule += ' ('
+            if required_props:
+                rule += ' "," space ( '
+
+            def get_recursive_refs(ks, first_is_optional):
+                [k, *rest] = ks
+                kv_rule_name = prop_kv_rule_names[k]
+                if k == '*':
+                    res = self._add_rule(
+                        f'{name}{"-" if name else ""}additional-kvs',
+                        f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*'
+                    )
+                elif first_is_optional:
+                    res = f'( "," space {kv_rule_name} )?'
+                else:
+                    res = kv_rule_name
+                if len(rest) > 0:
+                    res += ' ' + self._add_rule(
+                        f'{name}{"-" if name else ""}{k}-rest',
+                        get_recursive_refs(rest, first_is_optional=True)
+                    )
+                return res
+
+            rule += ' | '.join(
+                get_recursive_refs(optional_props[i:], first_is_optional=False)
+                for i in range(len(optional_props))
+            )
+            if required_props:
+                rule += ' )'
+            rule += ' )?'
+
+        rule += ' "}" space'
+
+        return rule
+
     def format_grammar(self):
-        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
+        return '\n'.join(
+            f'{name} ::= {rule}'
+            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
+        )
 
 
 def main(args_in = None):
@@ -129,16 +501,47 @@ def main(args_in = None):
         type=lambda s: s.split(','),
         help='''
             comma-separated property names defining the order of precedence for object properties;
-            properties not specified here are given lower precedence than those that are, and are
-            sorted alphabetically
+            properties not specified here are given lower precedence than those that are, and
+            are kept in their original order from the schema. Required properties are always
+            given precedence over optional properties.
         '''
     )
+    parser.add_argument(
+        '--allow-fetch',
+        action='store_true',
+        default=False,
+        help='Whether to allow fetching referenced schemas over HTTPS')
+    parser.add_argument(
+        '--dotall',
+        action='store_true',
+        default=False,
+        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
+    parser.add_argument(
+        '--raw-pattern',
+        action='store_true',
+        default=False,
+        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
+
     parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
     args = parser.parse_args(args_in)
 
-    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
-    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
-    converter = SchemaConverter(prop_order)
+    if args.schema.startswith('https://'):
+        url = args.schema
+        import requests
+        schema = requests.get(url).json()
+    elif args.schema == '-':
+        url = 'stdin'
+        schema = json.load(sys.stdin)
+    else:
+        url = f'file://{args.schema}'
+        with open(args.schema) as f:
+            schema = json.load(f)
+    converter = SchemaConverter(
+        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
+        allow_fetch=args.allow_fetch,
+        dotall=args.dotall,
+        raw_pattern=args.raw_pattern)
+    schema = converter.resolve_refs(schema, url)
     converter.visit(schema, '')
     print(converter.format_grammar())
 

examples/llama-bench/llama-bench.cpp

@@ -8,6 +8,7 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
@@ -103,6 +104,7 @@ static std::string get_cpu_info() {
                 }
             }
         }
+        fclose(f);
     }
 #endif
     // TODO: other platforms
@@ -112,10 +114,10 @@ static std::string get_cpu_info() {
 static std::string get_gpu_info() {
     std::string id;
 #ifdef GGML_USE_CUBLAS
-    int count = ggml_cuda_get_device_count();
+    int count = ggml_backend_cuda_get_device_count();
     for (int i = 0; i < count; i++) {
         char buf[128];
-        ggml_cuda_get_device_description(i, buf, sizeof(buf));
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
         id += buf;
         if (i < count - 1) {
             id += "/";
@@ -247,6 +249,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "q5_1") {
         return GGML_TYPE_Q5_1;
     }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
 
     return GGML_TYPE_COUNT;
 }
@@ -1122,15 +1127,19 @@ struct sql_printer : public printer {
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
-    //std::vector<llama_token> tokens(n_prompt, llama_token_bos(llama_get_model(ctx)));
-    //llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt, n_past, 0));
-    //GGML_UNUSED(n_batch);
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token> tokens(n_batch);
 
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
+        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        for (int i = 1; i < n_tokens; i++) {
+            tokens[i] = std::rand() % n_vocab;
+        }
         llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
@@ -1141,11 +1150,15 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
-    llama_token token = llama_token_bos(llama_get_model(ctx));
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
         llama_synchronize(ctx);
+        token = std::rand() % n_vocab;
     }
 }
 

examples/llava/MobileVLM-README.md

@@ -1,11 +1,13 @@
 # MobileVLM
 
-Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants.
+Currently this implementation supports [MobileVLM-1.7B](https://huggingface.co/mtgv/MobileVLM-1.7B) / [MobileVLM_V2-1.7B](https://huggingface.co/mtgv/MobileVLM_V2-1.7B) variants.
 
 for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM)
 
 The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
 
+Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion  is a little different. Therefore, using MobiVLM as an example, the different conversion step will be shown.
+
 ## Usage
 Build with cmake or run `make llava-cli` to build it.
 
@@ -34,7 +36,7 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF:
+3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` (for **V2** the arg is `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf \
@@ -44,6 +46,14 @@ python ./examples/llava/convert-image-encoder-to-gguf \
     --projector-type ldp
 ```
 
+```sh
+python ./examples/llava/convert-image-encoder-to-gguf \
+    -m path/to/clip-vit-large-patch14-336 \
+    --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
+    --output-dir path/to/MobileVLM-1.7B_V2 \
+    --projector-type ldpv2
+```
+
 4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh

examples/llava/README.md

@@ -63,12 +63,20 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+
+2) Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
 python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+
+4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
 ```console
 mkdir vit
 cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@@ -76,18 +84,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
 curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
 ```
 
-4) Create the visual gguf model:
+5) Create the visual gguf model:
 ```console
 python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 
-5) Then convert the model to gguf format:
+6) Then convert the model to gguf format:
 ```console
 python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```
 
-6) And finally we can run the llava-cli using the 1.6 model version:
+7) And finally we can run the llava-cli using the 1.6 model version:
 ```console
 ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```

examples/llava/clip.cpp

@@ -119,6 +119,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
 
@@ -126,12 +127,14 @@ enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
 static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
+    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
 };
 
 
@@ -475,6 +478,14 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_block_2_block_2_0_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_b;
+
+    // MobileVLM_V2 projection
+    struct ggml_tensor * mm_model_mlp_0_w;
+    struct ggml_tensor * mm_model_mlp_0_b;
+    struct ggml_tensor * mm_model_mlp_2_w;
+    struct ggml_tensor * mm_model_mlp_2_b;
+    struct ggml_tensor * mm_model_peg_0_w;
+    struct ggml_tensor * mm_model_peg_0_b;
 };
 
 struct clip_ctx {
@@ -497,7 +508,6 @@ struct clip_ctx {
 
     // memory buffers to evaluate the model
     ggml_backend_buffer_t params_buffer  = NULL;
-    ggml_backend_buffer_t compute_buffer = NULL;
 
     ggml_backend_t backend       = NULL;
     ggml_gallocr_t compute_alloc = NULL;
@@ -808,6 +818,29 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             }
             embeddings = block_1;
         }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
         else {
             GGML_ASSERT(false);
         }
@@ -995,6 +1028,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!new_clip->ctx_data) {
             fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1002,6 +1036,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         if (!fin) {
             printf("cannot open model file for loading tensors\n");
             clip_free(new_clip);
+            gguf_free(ctx);
             return nullptr;
         }
 
@@ -1023,6 +1058,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             if (!fin) {
                 printf("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
+                gguf_free(ctx);
                 return nullptr;
             }
             int num_bytes = ggml_nbytes(cur);
@@ -1175,7 +1211,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_block_2_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
             vision_model.mm_model_block_2_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
             vision_model.mm_model_block_2_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
-        } else {
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            // MobilVLM_V2 projection
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight"));
+            vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias"));
+            vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight"));
+            vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias"));
+            vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
+            vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
+        }
+        else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
         }
@@ -1232,16 +1279,16 @@ struct clip_image_f32 * clip_image_f32_init() {
 
 void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
 void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
     }
 }
-void clip_image_f32_batch_free(struct clip_image_f32_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
     }
 }
 
@@ -1494,7 +1541,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
 
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
-bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
@@ -1506,11 +1553,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
         pad_to_square = false;
     }
     // free the previous res_imgs if any set
-    if (res_imgs.size > 0) {
+    if (res_imgs->size > 0) {
         clip_image_f32_batch_free(res_imgs);
     }
-    res_imgs.data = nullptr;
-    res_imgs.size = 0;
+    res_imgs->data = nullptr;
+    res_imgs->size = 0;
 
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
@@ -1565,11 +1612,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
             patches.insert(patches.begin(), image_original_resize);
             // clip_image_f32_batch_init(patches.size());
-            res_imgs.size = patches.size();
-            res_imgs.data = new clip_image_f32[res_imgs.size];
+            res_imgs->size = patches.size();
+            res_imgs->data = new clip_image_f32[res_imgs->size];
             int num=0;
             for (auto& patch : patches) {
-                normalize_image_u8_to_f32(patch, &res_imgs.data[num], ctx->image_mean, ctx->image_std);
+                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
                 num++;
             }
 
@@ -1657,9 +1704,9 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
     // }
     // res_imgs.push_back(res);
 
-    res_imgs.size = 1;
-    res_imgs.data = new clip_image_f32[res_imgs.size];
-    res_imgs.data[0] = *res;
+    res_imgs->size = 1;
+    res_imgs->data = new clip_image_f32[res_imgs->size];
+    res_imgs->data[0] = *res;
     clip_image_f32_free(res);
 
     return true;
@@ -1673,6 +1720,9 @@ void clip_free(clip_ctx * ctx) {
     ggml_free(ctx->ctx_data);
     gguf_free(ctx->ctx_gguf);
 
+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
     delete ctx;
 }
 
@@ -1908,6 +1958,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 break;
             default:
                 printf("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
                 return false;
             }
 
@@ -1960,6 +2011,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
         return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        return ctx->vision_model.mm_model_peg_0_b->ne[0];
+    }
     if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
     }

examples/llava/clip.h

@@ -60,8 +60,8 @@ CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
-CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  & batch);
-CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
@@ -69,7 +69,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
 /** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
 
 CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
 

examples/llava/convert-image-encoder-to-gguf.py

@@ -1,6 +1,7 @@
 import argparse
 import os
 import json
+import re
 
 import torch
 import numpy as np
@@ -38,9 +39,11 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b
 def get_tensor_name(name: str) -> str:
     if "projection" in name:
         return name
-
     if "mm_projector" in name:
-        return name.replace("model.mm_projector", "mm")
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
 
     return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
 
@@ -83,7 +86,7 @@ ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
 ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
                 help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
-ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
 # Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5

examples/llava/llava.cpp

@@ -223,7 +223,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     clip_image_f32_batch img_res_v;
     img_res_v.size = 0;
     img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
         delete[] img_res_v.data;
         return false;

examples/llava/llava.h

@@ -29,9 +29,9 @@ struct llava_image_embed {
 };
 
 /** sanity check for clip <-> llava embed size match */
-LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
+LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
 
-LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
 
 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);

examples/lookup/CMakeLists.txt

@@ -3,3 +3,21 @@ add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-create)
+add_executable(${TARGET} lookup-create.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-merge)
+add_executable(${TARGET} lookup-merge.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-stats)
+add_executable(${TARGET} lookup-stats.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/lookup/lookup-create.cpp

@@ -0,0 +1,43 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    GGML_ASSERT(model != nullptr);
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+    fprintf(stderr, "%s: tokenization done\n", __func__);
+
+
+    llama_ngram_cache ngram_cache;
+    llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
+
+    llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+}

examples/lookup/lookup-merge.cpp

@@ -0,0 +1,47 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static void print_usage() {
+    fprintf(stderr, "Merges multiple lookup cache files into a single one.\n");
+    fprintf(stderr, "Usage: lookup-merge [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n");
+}
+
+int main(int argc, char ** argv){
+    if (argc < 3) {
+        print_usage();
+        exit(1);
+    }
+
+    std::vector<std::string> args;
+    args.resize(argc-1);
+    for (int i = 0; i < argc-1; ++i) {
+        args[i] = argv[i+1];
+        if (args[i] == "-h" || args[i] == "--help") {
+            print_usage();
+            exit(0);
+        }
+    }
+
+    fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
+    llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
+
+    for (size_t i = 1; i < args.size()-1; ++i) {
+        fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
+        llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
+
+        llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
+    }
+
+    fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
+    llama_ngram_cache_save(ngram_cache_merged, args.back());
+}

examples/lookup/lookup-stats.cpp

@@ -0,0 +1,163 @@
+#include "ggml.h"
+#include "common.h"
+#include "llama.h"
+#include "log.h"
+#include "ngram-cache.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    const int n_draft = params.n_draft;
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
+
+    // tokenize the prompt
+    const bool add_bos = llama_should_add_bos_token(model);
+    LOG("add_bos tgt: %d\n", add_bos);
+
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        const int64_t t_start_draft_us = ggml_time_us();
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
+    const int n_input = inp.size();
+    const int n_ctx = params.n_ctx;
+
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    const int64_t t_start_ms = ggml_time_ms();
+
+    // Iterate over input tokens in chunks of size n_ctx.
+    // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
+    for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
+        const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
+        std::vector<llama_token> pseudo_output;
+        pseudo_output.push_back(inp_slice[0]);
+
+        while ((int) pseudo_output.size() < n_ctx) {
+            // Simulate drafting and decoding from draft:
+            std::vector<llama_token> draft;
+            draft.push_back(pseudo_output.back());
+
+            {
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
+
+            n_drafted += draft.size() - 1;
+
+            for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
+                const llama_token ground_truth = inp_slice[pseudo_output.size()];
+                const llama_token drafted = draft[j];
+
+                if (ground_truth != drafted) {
+                    break;
+                }
+
+                ++n_accept;
+                pseudo_output.push_back(ground_truth);
+
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            // After each simulated batch decoding simulate the sampling of a single token:
+            if ((int) pseudo_output.size() < n_ctx) {
+                pseudo_output.push_back(inp_slice[pseudo_output.size()]);
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            draft.erase(draft.begin());
+
+        }
+        if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
+            const int64_t t_now_ms = ggml_time_ms();
+            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
+            const int64_t eta_min  = eta_ms / (60*1000);
+            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+            LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
+        }
+
+        // After each chunk, update the dynamic ngram cache with the context ngram cache:
+        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+        ngram_cache_context.clear();
+    }
+
+    LOG_TEE("\n");
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_input - n_input % n_ctx);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/lookup/lookup.cpp

@@ -1,12 +1,15 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
 
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
+#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
@@ -15,11 +18,7 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    // max/min n-grams size to search for in prompt
-    const int ngram_max = 4;
-    const int ngram_min = 1;
-
-    // length of the candidate / draft sequence, if match is found
+    // max. number of additional tokens to draft if match is found
     const int n_draft = params.n_draft;
 
     const bool dump_kv_cache = params.dump_kv_cache;
@@ -39,6 +38,8 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
     const bool add_bos = llama_should_add_bos_token(model);
@@ -47,6 +48,35 @@ int main(int argc, char ** argv){
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
 
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        // Fill up context ngram cache with tokens from user input:
+        const int64_t t_start_draft_us = ggml_time_us();
+        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
     const int max_context_size     = llama_n_ctx(ctx);
     const int max_tokens_list_size = max_context_size - 4;
 
@@ -76,8 +106,6 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
-    int64_t t_draft_us = 0;
-
     int n_past = inp.size();
 
     bool has_eos = false;
@@ -129,6 +157,12 @@ int main(int argc, char ** argv){
                 ++n_past;
                 ++i_dft;
                 inp.push_back(id);
+                {
+                    // Update context ngram cache with the newly accepted token:
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
 
                 if (params.use_color) {
                     // color accepted draft token
@@ -149,6 +183,12 @@ int main(int argc, char ** argv){
             draft.clear();
             draft.push_back(id);
             inp.push_back(id);
+            {
+                // Update context ngram cache with the newly accepted token:
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
             break;
         }
 
@@ -163,44 +203,19 @@ int main(int argc, char ** argv){
         llama_batch_clear(batch_tgt);
         llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
 
-        // generate n_pred tokens through prompt lookup
-        auto prompt_lookup = [&]() -> void {
-            const int inp_size = inp.size();
-            for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
-                const llama_token * ngram = &inp[inp_size - ngram_size];
-
-                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
-                    bool match = true;
-                    for (int j = 0; j < ngram_size; ++j) {
-                        if (inp[i + j] != ngram[j]) {
-                            match = false;
-                            break;
-                        }
-                    }
-
-                    if (match) {
-                        const int startIdx = i + ngram_size;
-                        const int endIdx = startIdx + n_draft;
-                        if (endIdx < inp_size) {
-                            for (int j = startIdx; j < endIdx; ++j) {
-                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
-                                draft.push_back(inp[j]);
-                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
-                                ++n_drafted;
-                            }
-                            return;
-                        }
-                    }
-                }
-            }
-            return;
-        };
-
+        // Draft already contains a single token sampled from the model:
+        GGML_ASSERT(draft.size() == 1);
+        GGML_ASSERT(draft[0] == inp.back());
         const int64_t t_start_draft_us = ggml_time_us();
 
-        prompt_lookup();
+        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+
+        for (size_t i = 1; i < draft.size(); ++i) {
+            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+        }
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
+        n_drafted += draft.size() - 1;
 
         llama_decode(ctx, batch_tgt);
         ++n_past;
@@ -210,19 +225,24 @@ int main(int argc, char ** argv){
 
     auto t_dec_end = ggml_time_us();
 
+    // Update dynamic ngram cache with context ngram cache and save it to disk:
+    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+
     LOG_TEE("\n\n");
 
     LOG_TEE("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_TEE("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_TEE("\n");
-    LOG_TEE("n_draft   = %d\n", n_draft);
-    LOG_TEE("n_predict = %d\n", n_predict);
-    LOG_TEE("n_drafted = %d\n", n_drafted);
-    LOG_TEE("t_draft   = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_predict);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
             t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
-    LOG_TEE("n_accept  = %d\n", n_accept);
-    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx);

examples/main/README.md

@@ -67,6 +67,7 @@ main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt
 In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
 
 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.

examples/quantize/quantize.cpp

@@ -189,6 +189,18 @@ static void prepare_imatrix(const std::string& imatrix_file,
     }
 }
 
+static ggml_type parse_ggml_type(const char * arg) {
+    ggml_type result = GGML_TYPE_COUNT;
+    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
+        auto type = ggml_type(j);
+        const auto * name = ggml_type_name(type);
+        if (name && strcmp(arg, name) == 0) {
+            result = type; break;
+        }
+    }
+    return result;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -203,6 +215,18 @@ int main(int argc, char ** argv) {
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
             params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {

examples/regex-to-grammar.py

@@ -0,0 +1,20 @@
+import json, subprocess, sys, os
+
+assert len(sys.argv) >= 2
+[_, pattern, *rest] = sys.argv
+
+print(subprocess.check_output(
+    [
+        "python",
+        os.path.join(
+        os.path.dirname(os.path.realpath(__file__)),
+        "json-schema-to-grammar.py"),
+        *rest,
+        "-",
+        "--raw-pattern",
+    ],
+    text=True,
+    input=json.dumps({
+        "type": "string",
+        "pattern": pattern,
+    }, indent=2)))

examples/server/CMakeLists.txt

@@ -2,12 +2,16 @@ set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET}
+    server.cpp
+    utils.hpp
+    httplib.h
+)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common json-schema-to-grammar ${CMAKE_THREAD_LIBS_INIT})
 if (LLAMA_SERVER_SSL)
     find_package(OpenSSL REQUIRED)
     target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto)

examples/server/README.md

@@ -16,16 +16,20 @@ The project is under active development, and we are [looking for feedback and co
 
 **Command line options:**
 
-- `--threads N`, `-t N`: Set the number of threads to use during generation.
-- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
+- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching, this parameter is used only if one token is to be processed on CPU backend.
+- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
 - `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`)
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
+- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (default: unused).
+- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository (default: unused).
+- `-hff FILE, --hf-file FILE`: Hugging Face model file (default: unused).
 - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 - `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
 - `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 - `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 - `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
-- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
+- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048`.
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. Default: `512`.
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
@@ -56,7 +60,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
 - `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
 - `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, default: enabled.
+- `--log-disable`: Output logs to stdout only, not to `llama.log`. default: enabled.
 - `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
 
 **If compiled with `LLAMA_SERVER_SSL=ON`**
@@ -259,7 +263,7 @@ node index.js
 
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
-    `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
+    `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 previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
 

examples/server/chat.mjs

@@ -26,8 +26,9 @@ const propOrder = grammarJsonSchemaPropOrder
 
 let grammar = null
 if (grammarJsonSchemaFile) {
-    const schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
-    const converter = new SchemaConverter(propOrder)
+    let schema = JSON.parse(readFileSync(grammarJsonSchemaFile, 'utf-8'))
+    const converter = new SchemaConverter({prop_order: propOrder, allow_fetch: true})
+    schema = await converter.resolveRefs(schema, grammarJsonSchemaFile)
     converter.visit(schema, '')
     grammar = converter.formatGrammar()
 }

examples/server/completion.js.hpp

@@ -483,4 +483,4 @@ unsigned char completion_js[] = {
   0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
   0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x7d, 0x0a
 };
-unsigned int completion_js_len = 5796;
+size_t completion_js_len = 5796;

examples/server/public/index.html

@@ -630,14 +630,16 @@
 
       const grammarJsonSchemaPropOrder = signal('')
       const updateGrammarJsonSchemaPropOrder = (el) => grammarJsonSchemaPropOrder.value = el.target.value
-      const convertJSONSchemaGrammar = () => {
+      const convertJSONSchemaGrammar = async () => {
         try {
-          const schema = JSON.parse(params.value.grammar)
-          const converter = new SchemaConverter(
-            grammarJsonSchemaPropOrder.value
+          let schema = JSON.parse(params.value.grammar)
+          const converter = new SchemaConverter({
+            prop_order: grammarJsonSchemaPropOrder.value
               .split(',')
-              .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {})
-          )
+              .reduce((acc, cur, i) => ({ ...acc, [cur.trim()]: i }), {}),
+            allow_fetch: true,
+          })
+          schema = await converter.resolveRefs(schema, 'input')
           converter.visit(schema, '')
           params.value = {
             ...params.value,

examples/server/public/json-schema-to-grammar.mjs

@@ -1,112 +1,538 @@
+// WARNING: This file was ported from json-schema-to-grammar.py, please fix bugs / add features there first.
 const SPACE_RULE = '" "?';
 
 const PRIMITIVE_RULES = {
   boolean: '("true" | "false") space',
   number: '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
   integer: '("-"? ([0-9] | [1-9] [0-9]*)) space',
+  value: 'object | array | string | number | boolean',
+  object: '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
+  array: '"[" space ( value ("," space value)* )? "]" space',
+  uuid: '"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space',
   string: ` "\\"" (
         [^"\\\\] |
         "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
       )* "\\"" space`,
   null: '"null" space',
 };
+const OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value'];
+
+// TODO: support "uri", "email" string formats
+const DATE_RULES = {
+    'date'   : '[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )',
+    'time'   : '([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )',
+    'date-time': 'date "T" time',
+    'date-string': '"\\"" date "\\"" space',
+    'time-string': '"\\"" time "\\"" space',
+    'date-time-string': '"\\"" date-time "\\"" space',
+};
+
+const RESERVED_NAMES = {'root': true, ...PRIMITIVE_RULES, ...DATE_RULES};
 
 const INVALID_RULE_CHARS_RE = /[^\dA-Za-z-]+/g;
 const GRAMMAR_LITERAL_ESCAPE_RE = /[\n\r"]/g;
-const GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'};
+const GRAMMAR_RANGE_LITERAL_ESCAPE_RE = /[\n\r"\]\-\\]/g;
+const GRAMMAR_LITERAL_ESCAPES = { '\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]' };
+
+const NON_LITERAL_SET = new Set('|.()[]{}*+?');
+const ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = new Set('[]()|{}*+?');
 
 export class SchemaConverter {
-  constructor(propOrder) {
-    this._propOrder = propOrder || {};
-    this._rules = new Map();
-    this._rules.set('space', SPACE_RULE);
+  constructor(options) {
+    this._propOrder = options.prop_order || {};
+    this._allowFetch = options.allow_fetch || false;
+    this._dotall = options.dotall || false;
+    this._rules = {'space': SPACE_RULE};
+    this._refs = {};
+    this._refsBeingResolved = new Set();
   }
 
   _formatLiteral(literal) {
-    const escaped = JSON.stringify(literal).replace(
+    const escaped = literal.replace(
       GRAMMAR_LITERAL_ESCAPE_RE,
       m => GRAMMAR_LITERAL_ESCAPES[m]
     );
     return `"${escaped}"`;
   }
 
+  _formatRangeChar(literal) {
+    return JSON.stringify(literal).slice(1, -1).replace(
+      GRAMMAR_RANGE_LITERAL_ESCAPE_RE,
+      m => GRAMMAR_LITERAL_ESCAPES[m]
+    );
+  }
+
   _addRule(name, rule) {
     let escName = name.replace(INVALID_RULE_CHARS_RE, '-');
     let key = escName;
 
-    if (this._rules.has(escName)) {
-      if (this._rules.get(escName) === rule) {
+    if (escName in this._rules) {
+      if (this._rules[escName] === rule) {
         return key;
       }
 
       let i = 0;
-      while (this._rules.has(`${escName}${i}`)) {
+      while ((`${escName}${i}` in this._rules) && (this._rules[`${escName}${i}`] !== rule)) {
         i += 1;
       }
       key = `${escName}${i}`;
     }
 
-    this._rules.set(key, rule);
+    this._rules[key] = rule;
     return key;
   }
 
+  async resolveRefs(schema, url) {
+    const visit = async (n) => {
+      if (Array.isArray(n)) {
+        return Promise.all(n.map(visit));
+      } else if (typeof n === 'object' && n !== null) {
+        let ref = n.$ref;
+        let target;
+        if (ref !== undefined && !this._refs[ref]) {
+          if (ref.startsWith('https://')) {
+            if (!this._allowFetch) {
+              throw new Error('Fetching remote schemas is not allowed (use --allow-fetch for force)');
+            }
+            const fetch = (await import('node-fetch')).default;
+
+            const fragSplit = ref.split('#');
+            const baseUrl = fragSplit[0];
+
+            target = this._refs[baseUrl];
+            if (!target) {
+              target = await this.resolveRefs(await fetch(ref).then(res => res.json()), baseUrl);
+              this._refs[baseUrl] = target;
+            }
+
+            if (fragSplit.length === 1 || fragSplit[fragSplit.length - 1] === '') {
+              return target;
+            }
+          } else if (ref.startsWith('#/')) {
+            target = schema;
+            ref = `${url}${ref}`;
+            n.$ref = ref;
+          } else {
+            throw new Error(`Unsupported ref ${ref}`);
+          }
+
+          const selectors = ref.split('#')[1].split('/').slice(1);
+          for (const sel of selectors) {
+            if (!target || !(sel in target)) {
+              throw new Error(`Error resolving ref ${ref}: ${sel} not in ${JSON.stringify(target)}`);
+            }
+            target = target[sel];
+          }
+
+          this._refs[ref] = target;
+        } else {
+          await Promise.all(Object.values(n).map(visit));
+        }
+      }
+
+      return n;
+    };
+
+    return visit(schema);
+  }
+
+  _generateUnionRule(name, altSchemas) {
+    return altSchemas
+      .map((altSchema, i) => this.visit(altSchema, `${name ?? ''}${name ? '-' : 'alternative-'}${i}`))
+      .join(' | ');
+  }
+
+  _visitPattern(pattern, name) {
+    if (!pattern.startsWith('^') || !pattern.endsWith('$')) {
+      throw new Error('Pattern must start with "^" and end with "$"');
+    }
+    pattern = pattern.slice(1, -1);
+    const subRuleIds = {};
+
+    let i = 0;
+    const length = pattern.length;
+
+    const getDot = () => {
+      let rule;
+      if (this._dotall) {
+        rule = '[\\U00000000-\\U0010FFFF]';
+      } else {
+        // Accept any character... except \n and \r line break chars (\x0A and \xOD)
+        rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]';
+      }
+      return this._addRule('dot', rule);
+    };
+
+
+    const toRule = ([s, isLiteral]) => isLiteral ? "\"" + s + "\"" : s;
+
+    const transform = () => {
+      const start = i;
+      // For each component of this sequence, store its string representation and whether it's a literal.
+      // We only need a flat structure here to apply repetition operators to the last item, and
+      // to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
+      // (GBNF's syntax is luckily very close to regular expressions!)
+      const seq = [];
+
+      const joinSeq = () => {
+        const ret = [];
+        for (const [isLiteral, g] of groupBy(seq, x => x[1])) {
+          if (isLiteral) {
+            ret.push([[...g].map(x => x[0]).join(''), true]);
+          } else {
+            ret.push(...g);
+          }
+        }
+        if (ret.length === 1) {
+          return ret[0];
+        }
+        return [ret.map(x => toRule(x)).join(' '), false];
+      };
+
+      while (i < length) {
+        const c = pattern[i];
+        if (c === '.') {
+          seq.push([getDot(), false]);
+          i += 1;
+        } else if (c === '(') {
+          i += 1;
+          if (i < length) {
+            if (pattern[i] === '?') {
+              throw new Error(`Unsupported pattern syntax "${pattern[i]}" at index ${i} of /${pattern}/`);
+            }
+          }
+          seq.push([`(${toRule(transform())})`, false]);
+        } else if (c === ')') {
+          i += 1;
+          if (start <= 0 || pattern[start - 1] !== '(') {
+            throw new Error(`Unbalanced parentheses; start = ${start}, i = ${i}, pattern = ${pattern}`);
+          }
+          return joinSeq();
+        } else if (c === '[') {
+          let squareBrackets = c;
+          i += 1;
+          while (i < length && pattern[i] !== ']') {
+            if (pattern[i] === '\\') {
+              squareBrackets += pattern.slice(i, i + 2);
+              i += 2;
+            } else {
+              squareBrackets += pattern[i];
+              i += 1;
+            }
+          }
+          if (i >= length) {
+            throw new Error(`Unbalanced square brackets; start = ${start}, i = ${i}, pattern = ${pattern}`);
+          }
+          squareBrackets += ']';
+          i += 1;
+          seq.push([squareBrackets, false]);
+        } else if (c === '|') {
+          seq.push(['|', false]);
+          i += 1;
+        } else if (c === '*' || c === '+' || c === '?') {
+          seq[seq.length - 1] = [toRule(seq[seq.length - 1]) + c, false];
+          i += 1;
+        } else if (c === '{') {
+          let curlyBrackets = c;
+          i += 1;
+          while (i < length && pattern[i] !== '}') {
+            curlyBrackets += pattern[i];
+            i += 1;
+          }
+          if (i >= length) {
+            throw new Error(`Unbalanced curly brackets; start = ${start}, i = ${i}, pattern = ${pattern}`);
+          }
+          curlyBrackets += '}';
+          i += 1;
+          const nums = curlyBrackets.slice(1, -1).split(',').map(s => s.trim());
+          let minTimes, maxTimes;
+          if (nums.length === 1) {
+            minTimes = parseInt(nums[0], 10);
+            maxTimes = minTimes;
+          } else {
+            if (nums.length !== 2) {
+              throw new Error(`Invalid quantifier ${curlyBrackets}`);
+            }
+            minTimes = nums[0] ? parseInt(nums[0], 10) : 0;
+            maxTimes = nums[1] ? parseInt(nums[1], 10) : Infinity;
+          }
+
+          let [sub, subIsLiteral] = seq[seq.length - 1];
+
+          if (minTimes === 0 && maxTimes === Infinity) {
+            seq[seq.length - 1] = [`${sub}*`, false];
+          } else if (minTimes === 0 && maxTimes === 1) {
+            seq[seq.length - 1] = [`${sub}?`, false];
+          } else if (minTimes === 1 && maxTimes === Infinity) {
+            seq[seq.length - 1] = [`${sub}+`, false];
+          } else {
+            if (!subIsLiteral) {
+              let id = subRuleIds[sub];
+              if (id === undefined) {
+                id = this._addRule(`${name}-${Object.keys(subRuleIds).length + 1}`, sub);
+                subRuleIds[sub] = id;
+              }
+              sub = id;
+            }
+
+            const repeatedSub = Array.from({ length: minTimes }, () => subIsLiteral ? `"${sub.slice(1, -1).repeat(minTimes)}"` : sub);
+            const optionalSub = maxTimes !== undefined ? Array.from({ length: maxTimes - minTimes }, () => `${sub}?`) : [`${sub}*`];
+            seq[seq.length - 1] = [repeatedSub.concat(optionalSub).join(' '), false];
+          }
+        } else {
+          let literal = '';
+          while (i < length) {
+            if (pattern[i] === '\\' && i < length - 1) {
+              const next = pattern[i + 1];
+              if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.has(next)) {
+                i += 1;
+                literal += pattern[i];
+                i += 1;
+              } else {
+                literal += pattern.slice(i, i + 2);
+                i += 2;
+              }
+            } else if (pattern[i] === '"') {
+              literal += '\\"';
+              i += 1;
+            } else if (!NON_LITERAL_SET.has(pattern[i]) &&
+                (i === length - 1 || literal === '' || pattern[i + 1] === '.' || !NON_LITERAL_SET.has(pattern[i+1]))) {
+              literal += pattern[i];
+              i += 1;
+            } else {
+              break;
+            }
+          }
+          if (literal !== '') {
+            seq.push([literal, true]);
+          }
+        }
+      }
+
+      return joinSeq();
+    };
+
+    return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space")
+  }
+
+  _resolveRef(ref) {
+    let refName = ref.split('/').pop();
+    if (!(refName in this._rules) && !this._refsBeingResolved.has(ref)) {
+      this._refsBeingResolved.add(ref);
+      const resolved = this._refs[ref];
+      refName = this.visit(resolved, refName);
+      this._refsBeingResolved.delete(ref);
+    }
+    return refName;
+  }
+
+  _generateConstantRule(value) {
+    return this._formatLiteral(JSON.stringify(value));
+  }
+
   visit(schema, name) {
     const schemaType = schema.type;
-    const ruleName = name || 'root';
+    const schemaFormat = schema.format;
+    const ruleName = name in RESERVED_NAMES ? name + '-' : name == '' ? 'root' : name;
 
-    if (schema.oneOf || schema.anyOf) {
-      const rule = (schema.oneOf || schema.anyOf).map((altSchema, i) =>
-        this.visit(altSchema, `${name}${name ? "-" : ""}${i}`)
-      ).join(' | ');
-
-      return this._addRule(ruleName, rule);
+    const ref = schema.$ref;
+    if (ref !== undefined) {
+      return this._addRule(ruleName, this._resolveRef(ref));
+    } else if (schema.oneOf || schema.anyOf) {
+      return this._addRule(ruleName, this._generateUnionRule(name, schema.oneOf || schema.anyOf));
+    } else if (Array.isArray(schemaType)) {
+      return this._addRule(ruleName, this._generateUnionRule(name, schemaType.map(t => ({ type: t }))));
     } else if ('const' in schema) {
-      return this._addRule(ruleName, this._formatLiteral(schema.const));
+      return this._addRule(ruleName, this._generateConstantRule(schema.const));
     } else if ('enum' in schema) {
-      const rule = schema.enum.map(v => this._formatLiteral(v)).join(' | ');
+      const rule = schema.enum.map(v => this._generateConstantRule(v)).join(' | ');
       return this._addRule(ruleName, rule);
-    } else if (schemaType === 'object' && 'properties' in schema) {
-      // TODO: `required` keyword (from python implementation)
-      const propOrder = this._propOrder;
-      const propPairs = Object.entries(schema.properties).sort((a, b) => {
-        // sort by position in prop_order (if specified) then by key
-        const orderA = typeof propOrder[a[0]] === 'number' ? propOrder[a[0]] : Infinity;
-        const orderB = typeof propOrder[b[0]] === 'number' ? propOrder[b[0]] : Infinity;
-        return orderA - orderB || a[0].localeCompare(b[0]);
-      });
-
-      let rule = '"{" space';
-      propPairs.forEach(([propName, propSchema], i) => {
-        const propRuleName = this.visit(propSchema, `${name}${name ? "-" : ""}${propName}`);
-        if (i > 0) {
-          rule += ' "," space';
+    } else if ((schemaType === undefined || schemaType === 'object') &&
+               ('properties' in schema ||
+                ('additionalProperties' in schema && schema.additionalProperties !== true))) {
+      const required = new Set(schema.required || []);
+      const properties = Object.entries(schema.properties ?? {});
+      return this._addRule(ruleName, this._buildObjectRule(properties, required, name, schema.additionalProperties));
+    } else if ((schemaType === undefined || schemaType === 'object') && 'allOf' in schema) {
+      const required = new Set();
+      const properties = [];
+      const addComponent = (compSchema, isRequired) => {
+        const ref = compSchema.$ref;
+        if (ref !== undefined) {
+          compSchema = this._refs[ref];
         }
-        rule += ` ${this._formatLiteral(propName)} space ":" space ${propRuleName}`;
-      });
-      rule += ' "}" space';
 
-      return this._addRule(ruleName, rule);
-    } else if (schemaType === 'array' && 'items' in schema) {
-      // TODO `prefixItems` keyword (from python implementation)
-      const itemRuleName = this.visit(schema.items, `${name}${name ? "-" : ""}item`);
-      const rule = `"[" space (${itemRuleName} ("," space ${itemRuleName})*)? "]" space`;
-      return this._addRule(ruleName, rule);
+        if ('properties' in compSchema) {
+          for (const [propName, propSchema] of Object.entries(compSchema.properties)) {
+            properties.push([propName, propSchema]);
+            if (isRequired) {
+              required.add(propName);
+            }
+          }
+        }
+      };
+
+      for (const t of schema.allOf) {
+        if ('anyOf' in t) {
+          for (const tt of t.anyOf) {
+            addComponent(tt, false);
+          }
+        } else {
+          addComponent(t, true);
+        }
+      }
+
+      return this._addRule(ruleName, this._buildObjectRule(properties, required, name, /* additionalProperties= */ false));
+    } else if ((schemaType === undefined || schemaType === 'array') && ('items' in schema || 'prefixItems' in schema)) {
+      const items = schema.items ?? schema.prefixItems;
+      if (Array.isArray(items)) {
+        return this._addRule(
+          ruleName,
+          '"[" space ' +
+            items.map((item, i) => this.visit(item, `${name ?? ''}${name ? '-' : ''}tuple-${i}`)).join(' "," space ') +
+            ' "]" space'
+        );
+      } else {
+        const itemRuleName = this.visit(items, `${name ?? ''}${name ? '-' : ''}item`);
+        const listItemOperator = `( "," space ${itemRuleName} )`;
+        let successiveItems = '';
+        let minItems = schema.minItems || 0;
+        const maxItems = schema.maxItems;
+        if (minItems > 0) {
+          successiveItems = listItemOperator.repeat(minItems - 1);
+          minItems--;
+        }
+        if (maxItems !== undefined && maxItems > minItems) {
+          successiveItems += `${listItemOperator}?`.repeat(maxItems - minItems - 1);
+        } else {
+          successiveItems += `${listItemOperator}*`;
+        }
+        const rule = minItems === 0
+          ? `"[" space ( ${itemRuleName} ${successiveItems} )? "]" space`
+          : `"[" space ${itemRuleName} ${successiveItems} "]" space`;
+        return this._addRule(ruleName, rule);
+      }
+    } else if ((schemaType === undefined || schemaType === 'string') && 'pattern' in schema) {
+      return this._visitPattern(schema.pattern, ruleName);
+    } else if ((schemaType === undefined || schemaType === 'string') && /^uuid[1-5]?$/.test(schema.format || '')) {
+      return this._addRule(
+          ruleName === 'root' ? 'root' : schemaFormat,
+          PRIMITIVE_RULES['uuid'])
+    } else if ((schemaType === undefined || schemaType === 'string') && schema.format in DATE_RULES) {
+      for (const [t, r] of Object.entries(DATE_RULES)) {
+        this._addRule(t, r);
+      }
+      return schemaFormat + '-string';
+    } else if ((schemaType === 'object') || (Object.keys(schema).length === 0)) {
+      for (const n of OBJECT_RULE_NAMES) {
+        this._addRule(n, PRIMITIVE_RULES[n]);
+      }
+      return this._addRule(ruleName, 'object');
     } else {
-      if (!PRIMITIVE_RULES[schemaType]) {
+      if (!(schemaType in PRIMITIVE_RULES)) {
         throw new Error(`Unrecognized schema: ${JSON.stringify(schema)}`);
       }
-      return this._addRule(
-        ruleName === 'root' ? 'root' : schemaType,
-        PRIMITIVE_RULES[schemaType]
+      // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+      return this._addRule(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]);
+    }
+  }
+
+  _buildObjectRule(properties, required, name, additionalProperties) {
+    const propOrder = this._propOrder;
+    // sort by position in prop_order (if specified) then by original order
+    const sortedProps = properties.map(([k]) => k).sort((a, b) => {
+      const orderA = propOrder[a] || Infinity;
+      const orderB = propOrder[b] || Infinity;
+      return orderA - orderB || properties.findIndex(([k]) => k === a) - properties.findIndex(([k]) => k === b);
+    });
+
+    const propKvRuleNames = {};
+    for (const [propName, propSchema] of properties) {
+      const propRuleName = this.visit(propSchema, `${name ?? ''}${name ? '-' : ''}${propName}`);
+      propKvRuleNames[propName] = this._addRule(
+        `${name ?? ''}${name ? '-' : ''}${propName}-kv`,
+        `${this._formatLiteral(JSON.stringify(propName))} space ":" space ${propRuleName}`
       );
     }
+    const requiredProps = sortedProps.filter(k => required.has(k));
+    const optionalProps = sortedProps.filter(k => !required.has(k));
+
+    if (typeof additionalProperties === 'object' || additionalProperties === true) {
+      const subName = `${name ?? ''}${name ? '-' : ''}additional`;
+      const valueRule = this.visit(additionalProperties === true ? {} : additionalProperties, `${subName}-value`);
+      propKvRuleNames['*'] = this._addRule(
+        `${subName}-kv`,
+        `${this._addRule('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
+      optionalProps.push('*');
+    }
+
+    let rule = '"{" space ';
+    rule += requiredProps.map(k => propKvRuleNames[k]).join(' "," space ');
+
+    if (optionalProps.length > 0) {
+      rule += ' (';
+      if (requiredProps.length > 0) {
+        rule += ' "," space ( ';
+      }
+
+      const getRecursiveRefs = (ks, firstIsOptional) => {
+        const [k, ...rest] = ks;
+        const kvRuleName = propKvRuleNames[k];
+        let res;
+        if (k === '*') {
+            res = this._addRule(
+                `${name ?? ''}${name ? '-' : ''}additional-kvs`,
+                `${kvRuleName} ( "," space ` + kvRuleName + ` )*`
+            )
+        } else if (firstIsOptional) {
+          res = `( "," space ${kvRuleName} )?`;
+        } else {
+          res = kvRuleName;
+        }
+        if (rest.length > 0) {
+          res += ' ' + this._addRule(
+            `${name ?? ''}${name ? '-' : ''}${k}-rest`,
+            getRecursiveRefs(rest, true)
+          );
+        }
+        return res;
+      };
+
+      rule += optionalProps.map((_, i) => getRecursiveRefs(optionalProps.slice(i), false)).join(' | ');
+      if (requiredProps.length > 0) {
+        rule += ' )';
+      }
+      rule += ' )?';
+    }
+
+    rule += ' "}" space';
+
+    return rule;
   }
 
   formatGrammar() {
     let grammar = '';
-    this._rules.forEach((rule, name) => {
+    for (const [name, rule] of Object.entries(this._rules).sort(([a], [b]) => a.localeCompare(b))) {
       grammar += `${name} ::= ${rule}\n`;
-    });
+    }
     return grammar;
   }
 }
+
+// Helper function to group elements by a key function
+function* groupBy(iterable, keyFn) {
+  let lastKey = null;
+  let group = [];
+  for (const element of iterable) {
+    const key = keyFn(element);
+    if (lastKey !== null && key !== lastKey) {
+      yield [lastKey, group];
+      group = [];
+    }
+    group.push(element);
+    lastKey = key;
+  }
+  if (group.length > 0) {
+    yield [lastKey, group];
+  }
+}

examples/server/server.cpp

@@ -1,6 +1,7 @@
 #include "utils.hpp"
 
 #include "common.h"
+#include "json-schema-to-grammar.h"
 #include "llama.h"
 #include "grammar-parser.h"
 
@@ -29,7 +30,7 @@
 #include <signal.h>
 #include <memory>
 
-using json = nlohmann::json;
+using json = nlohmann::ordered_json;
 
 bool server_verbose = false;
 bool server_log_json = true;
@@ -178,6 +179,7 @@ struct server_slot {
     llama_token sampled;
     struct llama_sampling_params sparams;
     llama_sampling_context * ctx_sampling = nullptr;
+    json json_schema;
 
     int32_t ga_i = 0;   // group-attention state
     int32_t ga_n = 1;   // group-attention factor
@@ -845,7 +847,17 @@ struct server_context {
         slot.sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
         slot.params.n_keep             = json_value(data, "n_keep",            slot.params.n_keep);
         slot.params.seed               = json_value(data, "seed",              default_params.seed);
-        slot.sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
+        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);
+            } 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.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);
 
@@ -1235,7 +1247,7 @@ struct server_context {
             {"penalize_nl",               slot.sparams.penalize_nl},
             {"stop",                      slot.params.antiprompt},
             {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
-            {"n_keep",                    params.n_keep},
+            {"n_keep",                    slot.params.n_keep},
             {"ignore_eos",                ignore_eos},
             {"stream",                    slot.params.stream},
             {"logit_bias",                slot.sparams.logit_bias},
@@ -1746,7 +1758,7 @@ struct server_context {
         }
 
         // process in chunks of params.n_batch
-        int32_t n_batch = llama_n_batch(ctx);
+        int32_t n_batch  = llama_n_batch(ctx);
         int32_t n_ubatch = llama_n_ubatch(ctx);
 
         // next, batch any pending prompts without exceeding n_batch
@@ -2195,6 +2207,12 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     }
     printf("  -m FNAME, --model FNAME\n");
     printf("                            model path (default: %s)\n", params.model.c_str());
+    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
+    printf("                            model download url (default: unused)\n");
+    printf("  -hfr REPO, --hf-repo REPO\n");
+    printf("                            Hugging Face model repository (default: unused)\n");
+    printf("  -hff FILE, --hf-file FILE\n");
+    printf("                            Hugging Face model file (default: unused)\n");
     printf("  -a ALIAS, --alias ALIAS\n");
     printf("                            set an alias for the model, will be added as `model` field in completion response\n");
     printf("  --lora FNAME              apply LoRA adapter (implies --no-mmap)\n");
@@ -2211,7 +2229,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     printf("  -to N, --timeout N        server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embeddings              enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N       number of slots for process requests (default: %d)\n", params.n_parallel);
-    printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
+    printf("  -cb, --cont-batching      enable continuous batching (a.k.a dynamic batching) (default: enabled)\n");
     printf("  -spf FNAME, --system-prompt-file FNAME\n");
     printf("                            set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
     printf("  -ctk TYPE, --cache-type-k TYPE\n");
@@ -2317,6 +2335,24 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 break;
             }
             params.model = argv[i];
+        } else if (arg == "-mu" || arg == "--model-url") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_url = argv[i];
+        } else if (arg == "-hfr" || arg == "--hf-repo") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.hf_repo = argv[i];
+        } else if (arg == "-hff" || arg == "--hf-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.hf_file = argv[i];
         } else if (arg == "-a" || arg == "--alias") {
             if (++i >= argc) {
                 invalid_param = true;

examples/server/tests/README.md

@@ -57,7 +57,7 @@ Feature or Scenario must be annotated with `@llama.cpp` to be included in the de
 To run a scenario annotated with `@bug`, start:
 
 ```shell
-DEBUG=ON ./tests.sh --no-skipped --tags bug
+DEBUG=ON ./tests.sh --no-skipped --tags bug --stop
 ```
 
 After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.

examples/server/tests/features/embeddings.feature

@@ -4,7 +4,8 @@ Feature: llama.cpp server
 
   Background: Server startup
     Given a server listening on localhost:8080
-    And   a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
+    And   a model url https://huggingface.co/ggml-org/models/resolve/main/bert-bge-small/ggml-model-f16.gguf
+    And   a model file ggml-model-f16.gguf
     And   a model alias bert-bge-small
     And   42 as server seed
     And   2 slots

examples/server/tests/features/environment.py

@@ -1,17 +1,18 @@
-import errno
 import os
-import socket
-import subprocess
-import time
-from contextlib import closing
 import signal
+import socket
+import sys
+import time
+import traceback
+from contextlib import closing
+from subprocess import TimeoutExpired
 
 
 def before_scenario(context, scenario):
     context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
     if context.debug:
-        print("DEBUG=ON\n")
-    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m\n")
+        print("DEBUG=ON")
+    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m")
     port = 8080
     if 'PORT' in os.environ:
         port = int(os.environ['PORT'])
@@ -20,58 +21,45 @@ def before_scenario(context, scenario):
 
 
 def after_scenario(context, scenario):
-    if context.server_process is None:
-        return
-    if scenario.status == "failed":
-        if 'GITHUB_ACTIONS' in os.environ:
-            print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
-            if os.path.isfile('llama.log'):
-                with closing(open('llama.log', 'r')) as f:
-                    for line in f:
-                        print(line)
-        if not is_server_listening(context.server_fqdn, context.server_port):
-            print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
+    try:
+        if 'server_process' not in context or context.server_process is None:
+            return
+        if scenario.status == "failed":
+            if 'GITHUB_ACTIONS' in os.environ:
+                print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n")
+                if os.path.isfile('llama.log'):
+                    with closing(open('llama.log', 'r')) as f:
+                        for line in f:
+                            print(line)
+            if not is_server_listening(context.server_fqdn, context.server_port):
+                print("\x1b[33;101mERROR: Server stopped listening\x1b[0m")
 
-    if not pid_exists(context.server_process.pid):
-        assert False, f"Server not running pid={context.server_process.pid} ..."
+        if context.server_process.poll() is not None:
+            assert False, f"Server not running pid={context.server_process.pid} ..."
 
-    server_graceful_shutdown(context)
+        server_graceful_shutdown(context)  # SIGINT
 
-    # Wait few for socket to free up
-    time.sleep(0.05)
+        try:
+            context.server_process.wait(0.5)
+        except TimeoutExpired:
+            print(f"server still alive after 500ms, force-killing pid={context.server_process.pid} ...")
+            context.server_process.kill()  # SIGKILL
+            context.server_process.wait()
 
-    attempts = 0
-    while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
-        server_kill(context)
-        time.sleep(0.1)
-        attempts += 1
-        if attempts > 5:
-            server_kill_hard(context)
+        while is_server_listening(context.server_fqdn, context.server_port):
+            time.sleep(0.1)
+    except Exception:
+        print("ignoring error in after_scenario:")
+        traceback.print_exc(file=sys.stdout)
 
 
 def server_graceful_shutdown(context):
-    print(f"shutting down server pid={context.server_process.pid} ...\n")
+    print(f"shutting down server pid={context.server_process.pid} ...")
     if os.name == 'nt':
-        os.kill(context.server_process.pid, signal.CTRL_C_EVENT)
+        interrupt = signal.CTRL_C_EVENT
     else:
-        os.kill(context.server_process.pid, signal.SIGINT)
-
-
-def server_kill(context):
-    print(f"killing server pid={context.server_process.pid} ...\n")
-    context.server_process.kill()
-
-
-def server_kill_hard(context):
-    pid = context.server_process.pid
-    path = context.server_path
-
-    print(f"Server dangling exits, hard killing force {pid}={path}...\n")
-    if os.name == 'nt':
-        process = subprocess.check_output(['taskkill', '/F', '/pid', str(pid)]).decode()
-        print(process)
-    else:
-        os.kill(-pid, signal.SIGKILL)
+        interrupt = signal.SIGINT
+    context.server_process.send_signal(interrupt)
 
 
 def is_server_listening(server_fqdn, server_port):
@@ -79,22 +67,5 @@ def is_server_listening(server_fqdn, server_port):
         result = sock.connect_ex((server_fqdn, server_port))
         _is_server_listening = result == 0
         if _is_server_listening:
-            print(f"server is listening on {server_fqdn}:{server_port}...\n")
+            print(f"server is listening on {server_fqdn}:{server_port}...")
         return _is_server_listening
-
-
-def pid_exists(pid):
-    """Check whether pid exists in the current process table."""
-    if pid < 0:
-        return False
-    if os.name == 'nt':
-        output = subprocess.check_output(['TASKLIST', '/FI', f'pid eq {pid}']).decode()
-        print(output)
-        return "No tasks are running" not in output
-    else:
-        try:
-            os.kill(pid, 0)
-        except OSError as e:
-            return e.errno == errno.EPERM
-        else:
-            return True

examples/server/tests/features/parallel.feature

@@ -4,7 +4,8 @@ Feature: Parallel
 
   Background: Server startup
     Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
+    And   a model file test-model-00001-of-00003.gguf
     And   42 as server seed
     And   128 as batch size
     And   256 KV cache size

examples/server/tests/features/security.feature

@@ -37,6 +37,22 @@ Feature: Security
       | llama.cpp | no        |
       | hackme    | raised    |
 
+  Scenario Outline: OAI Compatibility (invalid response formats)
+    Given a system prompt test
+    And   a user prompt test
+    And   a response format <response_format>
+    And   a model test
+    And   2 max tokens to predict
+    And   streaming is disabled
+    Given an OAI compatible chat completions request with raised api error
+
+    Examples: Prompts
+      | response_format                                       |
+      | {"type": "sound"}                                     |
+      | {"type": "json_object", "schema": 123}                |
+      | {"type": "json_object", "schema": {"type": 123}}      |
+      | {"type": "json_object", "schema": {"type": "hiccup"}} |
+
 
   Scenario Outline: CORS Options
     Given a user api key llama.cpp

examples/server/tests/features/server.feature

@@ -5,6 +5,7 @@ Feature: llama.cpp server
   Background: Server startup
     Given a server listening on localhost:8080
     And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   a model file test-model.gguf
     And   a model alias tinyllama-2
     And   42 as server seed
       # KV Cache corresponds to the total amount of tokens
@@ -34,9 +35,9 @@ Feature: llama.cpp server
     And   metric llamacpp:tokens_predicted is <n_predicted>
 
     Examples: Prompts
-      | prompt                                                                    | n_predict | re_content                    | n_prompt | n_predicted | truncated |
-      | I believe the meaning of life is                                          | 8         | (read\|going)+                | 18       | 8           | not       |
-      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids)+ | 46       | 64          | not       |
+      | prompt                                                                    | n_predict | re_content                                  | n_prompt | n_predicted | truncated |
+      | I believe the meaning of life is                                          | 8         | (read\|going)+                              | 18       | 8           | not       |
+      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids\|Anna\|forest)+ | 46       | 64          | not       |
 
   Scenario: Completion prompt truncated
     Given a prompt:
@@ -47,7 +48,7 @@ Feature: llama.cpp server
     Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
     """
     And   a completion request with no api error
-    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry
+    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry|bowl
     And   the completion is  truncated
     And   109 prompt tokens are processed
 
@@ -64,9 +65,25 @@ Feature: llama.cpp server
     And   the completion is <truncated> truncated
 
     Examples: Prompts
-      | model        | system_prompt               | user_prompt                          | max_tokens | re_content             | n_prompt | n_predicted | enable_streaming | truncated |
-      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+          | 77       | 8           | disabled         | not       |
-      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird)+ | -1       | 64          | enabled          |           |
+      | model        | system_prompt               | user_prompt                          | max_tokens | re_content                        | n_prompt | n_predicted | enable_streaming | truncated |
+      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+                     | 77       | 8           | disabled         | not       |
+      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird\|Annabyear)+ | -1       | 64          | enabled          |           |
+
+
+  Scenario Outline: OAI Compatibility w/ response format
+    Given a model test
+    And   a system prompt test
+    And   a user prompt test
+    And   a response format <response_format>
+    And   10 max tokens to predict
+    Given an OAI compatible chat completions request with no api error
+    Then  <n_predicted> tokens are predicted matching <re_content>
+
+    Examples: Prompts
+      | response_format                                                     | n_predicted | re_content             |
+      | {"type": "json_object", "schema": {"const": "42"}}                  | 5           | "42"                   |
+      | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10          | \[ -300 \]             |
+      | {"type": "json_object"}                                             | 10          | \{ " Jacky.            |
 
 
   Scenario: Tokenize / Detokenize

examples/server/tests/features/steps/steps.py

@@ -5,6 +5,8 @@ import os
 import re
 import socket
 import subprocess
+import sys
+import threading
 import time
 from contextlib import closing
 from re import RegexFlag
@@ -14,7 +16,6 @@ import numpy as np
 import openai
 from behave import step
 from behave.api.async_step import async_run_until_complete
-from huggingface_hub import hf_hub_download
 from prometheus_client import parser
 
 
@@ -22,22 +23,29 @@ from prometheus_client import parser
 def step_server_config(context, server_fqdn, server_port):
     context.server_fqdn = server_fqdn
     context.server_port = int(server_port)
+    context.n_gpu_layer = None
     if 'PORT' in os.environ:
         context.server_port = int(os.environ['PORT'])
         print(f"$PORT set, overriding server port with to {context.server_port}")
     if 'FQDN' in os.environ:
         context.server_fqdn = os.environ['FQDN']
         print(f"$FQDN set, overriding server fqdn with to {context.server_fqdn}")
+    if 'N_GPU_LAYERS' in os.environ:
+        context.n_gpu_layer = int(os.environ['N_GPU_LAYERS'])
+        print(f"$N_GPU_LAYERS set, overriding n_gpu_layer with to {context.n_gpu_layer}")
 
     context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
 
     context.model_alias = None
+    context.model_file = None
+    context.model_hf_repo = None
+    context.model_hf_file = None
+    context.model_url = None
     context.n_batch = None
     context.n_ubatch = None
     context.n_ctx = None
     context.n_ga = None
     context.n_ga_w = None
-    context.n_gpu_layer = None
     context.n_predict = None
     context.n_prompts = 0
     context.n_server_predict = None
@@ -52,6 +60,7 @@ def step_server_config(context, server_fqdn, server_port):
     context.seed = None
     context.server_seed = None
     context.user_api_key = None
+    context.response_format = None
 
     context.tasks_result = []
     context.concurrent_tasks = []
@@ -60,9 +69,19 @@ def step_server_config(context, server_fqdn, server_port):
 
 @step('a model file {hf_file} from HF repo {hf_repo}')
 def step_download_hf_model(context, hf_file, hf_repo):
-    context.model_file = hf_hub_download(repo_id=hf_repo, filename=hf_file)
-    if context.debug:
-        print(f"model file: {context.model_file}\n")
+    context.model_hf_repo = hf_repo
+    context.model_hf_file = hf_file
+    context.model_file = os.path.basename(hf_file)
+
+
+@step('a model file {model_file}')
+def step_model_file(context, model_file):
+    context.model_file = model_file
+
+
+@step('a model url {model_url}')
+def step_model_url(context, model_url):
+    context.model_url = model_url
 
 
 @step('a model alias {model_alias}')
@@ -123,9 +142,12 @@ def step_start_server(context):
     if 'GITHUB_ACTIONS' in os.environ:
         max_attempts *= 2
 
+    addrs = socket.getaddrinfo(context.server_fqdn, context.server_port, type=socket.SOCK_STREAM)
+    family, typ, proto, _, sockaddr = addrs[0]
+
     while True:
-        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-            result = sock.connect_ex((context.server_fqdn, context.server_port))
+        with closing(socket.socket(family, typ, proto)) as sock:
+            result = sock.connect_ex(sockaddr)
             if result == 0:
                 print("\x1b[33;46mserver started!\x1b[0m")
                 return
@@ -141,7 +163,8 @@ def step_start_server(context):
 async def step_wait_for_the_server_to_be_started(context, expecting_status):
     match expecting_status:
         case 'healthy':
-            await wait_for_health_status(context, context.base_url, 200, 'ok')
+            await wait_for_health_status(context, context.base_url, 200, 'ok',
+                                         timeout=30)
 
         case 'ready' | 'idle':
             await wait_for_health_status(context, context.base_url, 200, 'ok',
@@ -194,7 +217,7 @@ async def step_request_completion(context, api_error):
                                           user_api_key=context.user_api_key)
     context.tasks_result.append(completion)
     if context.debug:
-        print(f"Completion response: {completion}\n")
+        print(f"Completion response: {completion}")
     if expect_api_error:
         assert completion == 401, f"completion must be an 401 status code: {completion}"
 
@@ -248,6 +271,11 @@ def step_max_tokens(context, max_tokens):
     context.n_predict = max_tokens
 
 
+@step('a response format {response_format}')
+def step_response_format(context, response_format):
+    context.response_format = json.loads(response_format)
+
+
 @step('streaming is {enable_streaming}')
 def step_streaming(context, enable_streaming):
     context.enable_streaming = enable_streaming == 'enabled'
@@ -339,7 +367,7 @@ def step_prompt_passkey(context, passkey, i_pos):
         prompt += context.prompt_junk_suffix
     if context.debug:
         passkey_highlight = "\x1b[33m" + passkey + "\x1b[0m"
-        print(f"Passkey challenge:\n```{prompt.replace(passkey, passkey_highlight)}```\n")
+        print(f"Passkey challenge:\n```{prompt.replace(passkey, passkey_highlight)}```")
     context.prompts.append(context.prompt_prefix + prompt + context.prompt_suffix)
     context.n_prompts = len(context.prompts)
 
@@ -348,7 +376,7 @@ def step_prompt_passkey(context, passkey, i_pos):
 @async_run_until_complete
 async def step_oai_chat_completions(context, api_error):
     if context.debug:
-        print(f"Submitting OAI compatible completions request...\n")
+        print(f"Submitting OAI compatible completions request...")
     expect_api_error = api_error == 'raised'
     completion = await oai_chat_completions(context.prompts.pop(),
                                             context.system_prompt,
@@ -363,6 +391,9 @@ async def step_oai_chat_completions(context, api_error):
                                             enable_streaming=context.enable_streaming
                                             if hasattr(context, 'enable_streaming') else None,
 
+                                            response_format=context.response_format
+                                            if hasattr(context, 'response_format') else None,
+
                                             seed=await completions_seed(context),
 
                                             user_api_key=context.user_api_key
@@ -422,6 +453,8 @@ async def step_oai_chat_completions(context):
                               if hasattr(context, 'n_predict') else None,
                               enable_streaming=context.enable_streaming
                               if hasattr(context, 'enable_streaming') else None,
+                              response_format=context.response_format
+                              if hasattr(context, 'response_format') else None,
                               seed=await completions_seed(context),
                               user_api_key=context.user_api_key
                               if hasattr(context, 'user_api_key') else None)
@@ -442,6 +475,8 @@ async def step_oai_chat_completions(context):
                               if hasattr(context, 'n_predict') else None,
                               enable_streaming=context.enable_streaming
                               if hasattr(context, 'enable_streaming') else None,
+                              response_format=context.response_format
+                              if hasattr(context, 'response_format') else None,
                               seed=context.seed
                               if hasattr(context, 'seed') else
                               context.server_seed
@@ -493,12 +528,12 @@ async def step_all_embeddings_are_the_same(context):
             embedding1 = np.array(embeddings[i])
             embedding2 = np.array(embeddings[j])
             if context.debug:
-                print(f"embedding1: {embedding1[-8:]}\n")
-                print(f"embedding2: {embedding2[-8:]}\n")
+                print(f"embedding1: {embedding1[-8:]}")
+                print(f"embedding2: {embedding2[-8:]}")
             similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
             msg = f"Similarity between {i} and {j}: {similarity:.10f}"
             if context.debug:
-                print(f"{msg}\n")
+                print(f"{msg}")
             assert np.isclose(similarity, 1.0, rtol=1e-05, atol=1e-08, equal_nan=False), msg
 
 
@@ -615,7 +650,7 @@ async def step_prometheus_metrics_exported(context):
             metrics_raw = await metrics_response.text()
             metric_exported = False
             if context.debug:
-                print(f"/metrics answer:\n{metrics_raw}\n")
+                print(f"/metrics answer:\n{metrics_raw}")
             context.metrics = {}
             for metric in parser.text_string_to_metric_families(metrics_raw):
                 match metric.name:
@@ -724,6 +759,7 @@ async def oai_chat_completions(user_prompt,
                                model=None,
                                n_predict=None,
                                enable_streaming=None,
+                               response_format=None,
                                seed=None,
                                user_api_key=None,
                                expect_api_error=None):
@@ -749,6 +785,8 @@ async def oai_chat_completions(user_prompt,
         "stream": enable_streaming,
         "seed": seed
     }
+    if response_format is not None:
+        payload['response_format'] = response_format
     completion_response = {
         'content': '',
         'timings': {
@@ -809,6 +847,7 @@ async def oai_chat_completions(user_prompt,
                 model=model,
                 max_tokens=n_predict,
                 stream=enable_streaming,
+                response_format=payload.get('response_format'),
                 seed=seed
             )
         except openai.error.AuthenticationError as e:
@@ -917,7 +956,7 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
             last_match = end
         highlighted += content[last_match:]
         if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-          print(f"Checking completion response: {highlighted}\n")
+          print(f"Checking completion response: {highlighted}")
         assert last_match > 0, f'/{re_content}/ must match ```{highlighted}```'
     if expected_predicted_n and expected_predicted_n > 0:
         assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
@@ -927,7 +966,7 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
 async def gather_tasks_results(context):
     n_tasks = len(context.concurrent_tasks)
     if context.debug:
-        print(f"Waiting for all {n_tasks} tasks results...\n")
+        print(f"Waiting for all {n_tasks} tasks results...")
     for task_no in range(n_tasks):
         context.tasks_result.append(await context.concurrent_tasks.pop())
     n_completions = len(context.tasks_result)
@@ -944,7 +983,7 @@ async def wait_for_health_status(context,
                                  slots_processing=None,
                                  expected_slots=None):
     if context.debug:
-        print(f"Starting checking for health for expected_health_status={expected_health_status}\n")
+        print(f"Starting checking for health for expected_health_status={expected_health_status}")
     interval = 0.5
     counter = 0
     if 'GITHUB_ACTIONS' in os.environ:
@@ -1033,13 +1072,18 @@ def start_server_background(context):
     if 'LLAMA_SERVER_BIN_PATH' in os.environ:
         context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
     server_listen_addr = context.server_fqdn
-    if os.name == 'nt':
-        server_listen_addr = '0.0.0.0'
     server_args = [
         '--host', server_listen_addr,
         '--port', context.server_port,
-        '--model', context.model_file
     ]
+    if context.model_file:
+        server_args.extend(['--model', context.model_file])
+    if context.model_url:
+        server_args.extend(['--model-url', context.model_url])
+    if context.model_hf_repo:
+        server_args.extend(['--hf-repo', context.model_hf_repo])
+    if context.model_hf_file:
+        server_args.extend(['--hf-file', context.model_hf_file])
     if context.n_batch:
         server_args.extend(['--batch-size', context.n_batch])
     if context.n_ubatch:
@@ -1070,7 +1114,7 @@ def start_server_background(context):
         server_args.append('--verbose')
     if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
         server_args.extend(['--log-format', "text"])
-    print(f"starting server with: {context.server_path} {server_args}\n")
+    print(f"starting server with: {context.server_path} {server_args}")
     flags = 0
     if 'nt' == os.name:
         flags |= subprocess.DETACHED_PROCESS
@@ -1079,8 +1123,23 @@ def start_server_background(context):
 
     pkwargs = {
         'creationflags': flags,
+        'stdout': subprocess.PIPE,
+        'stderr': subprocess.PIPE
     }
     context.server_process = subprocess.Popen(
         [str(arg) for arg in [context.server_path, *server_args]],
         **pkwargs)
+
+    def log_stdout(process):
+        for line in iter(process.stdout.readline, b''):
+            print(line.decode('utf-8'), end='')
+    thread_stdout = threading.Thread(target=log_stdout, args=(context.server_process,))
+    thread_stdout.start()
+
+    def log_stderr(process):
+        for line in iter(process.stderr.readline, b''):
+            print(line.decode('utf-8'), end='', file=sys.stderr)
+    thread_stderr = threading.Thread(target=log_stderr, args=(context.server_process,))
+    thread_stderr.start()
+
     print(f"server pid={context.server_process.pid}, behave pid={os.getpid()}")

examples/server/utils.hpp

@@ -12,7 +12,7 @@
 
 #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
-using json = nlohmann::json;
+using json = nlohmann::ordered_json;
 
 // https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
 enum error_type {
@@ -95,8 +95,8 @@ static inline void server_log(const char *level, const char *function, int line,
 
         const std::string str = ss.str();
         printf("%.*s\n", (int)str.size(), str.data());
-        fflush(stdout);
     }
+    fflush(stdout);
 }
 
 //
@@ -371,11 +371,23 @@ static json oaicompat_completion_params_parse(
     llama_params["repeat_last_n"]     = json_value(body,   "repeat_last_n",     default_sparams.penalty_last_n);
     llama_params["ignore_eos"]        = json_value(body,   "ignore_eos",        false);
     llama_params["tfs_z"]             = json_value(body,   "tfs_z",             default_sparams.tfs_z);
+    llama_params["n_keep"]            = json_value(body,   "n_keep",            0);
 
-    if (body.count("grammar") != 0) {
+    if (body.contains("grammar")) {
         llama_params["grammar"] = json_value(body, "grammar", json::object());
     }
 
+    if (body.contains("response_format")) {
+        auto response_format = json_value(body, "response_format", json::object());
+        if (response_format.contains("type")) {
+            if (response_format["type"] == "json_object") {
+                llama_params["json_schema"] = json_value(response_format, "schema", json::object());
+            } else {
+                throw std::runtime_error("response_format type not supported: " + response_format["type"].dump());
+            }
+        }
+    }
+
     // Handle 'stop' field
     if (body.contains("stop") && body["stop"].is_string()) {
         llama_params["stop"] = json::array({body["stop"].get<std::string>()});

examples/speculative/speculative.cpp

@@ -219,7 +219,8 @@ int main(int argc, char ** argv) {
                 if (params.sparams.temp > 0) {
                     // stochastic verification
 
-                    llama_token_data_array dist_tgt = llama_sampling_probability_distribution(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+                    llama_token_data_array dist_tgt = llama_sampling_prepare(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft], true, NULL);
+                    llama_sample_softmax(ctx_tgt, &dist_tgt);
                     float p_tgt = 0, p_dft = 0;
 
                     // GGML_ASSERT(dist_tgt.size() == dist_dft.size());

examples/sycl/build.sh

@@ -13,8 +13,11 @@ source /opt/intel/oneapi/setvars.sh
 #for FP32
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
-#build example/main only
+#build example/main
 #cmake --build . --config Release --target main
 
+#build example/llama-bench
+#cmake --build . --config Release --target llama-bench
+
 #build all binary
 cmake --build . --config Release -v

examples/sycl/run-llama2.sh

@@ -9,18 +9,28 @@ source /opt/intel/oneapi/setvars.sh
 
 if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
+    GGML_SYCL_SINGLE_GPU=1
 else
     GGML_SYCL_DEVICE=0
 fi
-echo "use $GGML_SYCL_DEVICE as main GPU"
+
 #export GGML_SYCL_DEBUG=1
 
 
 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
 
-#use all GPUs with same max compute units
-ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
+    echo "use $GGML_SYCL_DEVICE as main GPU"
+    #use signle GPU only
+    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+else
+    #use multiple GPUs with same max compute units
+    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+fi
 
 #use main GPU only
 #ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
 
+#use multiple GPUs with same max compute units
+#ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+

examples/sycl/win-run-llama2.bat

@@ -6,8 +6,6 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
 
-set GGML_SYCL_DEVICE=0
-rem set GGML_SYCL_DEBUG=1
 .\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
 
 

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -711,6 +711,7 @@ static bool load_checkpoint_file(const char * filename, struct my_llama_model *
 
     load_checkpoint_gguf(fctx, f_ggml_ctx, model, train);
 
+    gguf_free(fctx);
     return true;
 }
 

examples/ts-type-to-grammar.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+#
+# ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
+# python examples/json-schema-to-grammar.py https://json.schemastore.org/tsconfig.json
+#
+set -euo pipefail
+
+readonly type="$1"
+
+# Create a temporary directory
+TMPDIR=""
+trap 'rm -fR "$TMPDIR"' EXIT
+TMPDIR=$(mktemp -d)
+
+DTS_FILE="$TMPDIR/type.d.ts"
+SCHEMA_FILE="$TMPDIR/schema.json"
+
+echo "export type MyType = $type" > "$DTS_FILE"
+
+# This is a fork of typescript-json-schema, actively maintained as of March 2024:
+# https://github.com/vega/ts-json-schema-generator
+npx ts-json-schema-generator --unstable --no-top-ref --path "$DTS_FILE" --type MyType -e none > "$SCHEMA_FILE"
+
+# Alternative, not actively maintained as of March 2024:
+# https://github.com/YousefED/typescript-json-schema
+# npx typescript-json-schema --defaultProps --required "$DTS_FILE" MyType | tee "$SCHEMA_FILE" >&2
+
+./examples/json-schema-to-grammar.py "$SCHEMA_FILE"

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1709703039,
-        "narHash": "sha256-6hqgQ8OK6gsMu1VtcGKBxKQInRLHtzulDo9Z5jxHEFY=",
+        "lastModified": 1710451336,
+        "narHash": "sha256-pP86Pcfu3BrAvRO7R64x7hs+GaQrjFes+mEPowCfkxY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9df3e30ce24fd28c7b3e2de0d986769db5d6225d",
+        "rev": "d691274a972b3165335d261cc4671335f5c67de9",
         "type": "github"
       },
       "original": {

ggml-alloc.c

@@ -548,7 +548,11 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
-        if (ggml_is_view(node)) {
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
             struct ggml_tensor * view_src = node->view_src;
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
@@ -565,8 +569,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 
             ggml_gallocr_hash_get(galloc, src)->n_children += 1;
 
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
                 ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
             }
         }

ggml-backend-impl.h

@@ -103,6 +103,11 @@ extern "C" {
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
 
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
         // (optional) event synchronization
         ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
         void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);

ggml-backend.c

@@ -278,7 +278,7 @@ enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_
     return err;
 }
 
-bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return backend->iface.graph_compute(backend, cgraph);
 }
 
@@ -286,6 +286,13 @@ bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor *
     return backend->iface.supports_op(backend, op);
 }
 
+bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
+}
+
 // backend copy
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@@ -761,6 +768,10 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg
 
     if (cpu_plan->cplan.work_size > 0) {
         cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+        if (cpu_plan->cplan.work_data == NULL) {
+            free(cpu_plan);
+            return NULL;
+        }
     }
 
     cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
@@ -834,6 +845,7 @@ static struct ggml_backend_i cpu_backend_i = {
     /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
     /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
     /* .supports_op             = */ ggml_backend_cpu_supports_op,
+    /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
@@ -999,11 +1011,11 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 256
+#define GGML_SCHED_MAX_SPLITS 2048
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
-#define GGML_SCHED_MAX_SPLIT_INPUTS 16
+#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
 #endif
 
 #ifndef GGML_SCHED_MAX_COPIES
@@ -1043,8 +1055,9 @@ struct ggml_backend_sched {
     struct ggml_cgraph * graph;
 
     // graph splits
-    struct ggml_backend_sched_split splits[GGML_SCHED_MAX_SPLITS];
+    struct ggml_backend_sched_split * splits;
     int n_splits;
+    int splits_capacity;
 
     // pipeline parallelism support
     int n_copies;
@@ -1114,40 +1127,48 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     // TODO: use supports_op to check if the backend supports the op
 
     // assign pre-allocated nodes to their backend
-    // dst
-    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor);
-    if (cur_backend != -1) {
+    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor);
+    if (cur_backend_id != -1) {
         SET_CAUSE(tensor, "1.dst");
-        return cur_backend;
+        return cur_backend_id;
     }
 
     // view_src
     if (tensor->view_src != NULL) {
-        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
-        if (cur_backend != -1) {
+        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
+        if (cur_backend_id != -1) {
             SET_CAUSE(tensor, "1.vsrc");
-            return cur_backend;
+            return cur_backend_id;
         }
     }
 
-    // input
+    // graph input
     if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
-        cur_backend = sched->n_backends - 1; // last backend (assumed CPU)
+        cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
         SET_CAUSE(tensor, "1.inp");
-        return cur_backend;
+        return cur_backend_id;
     }
 
     // assign nodes that use weights to the backend of the weights
+    // operations with weights are preferably run on the same backend as the weights
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
         if (src == NULL) {
             continue;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src);
-            // operations with weights are always run on the same backend as the weights
+            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src);
+            // check if a backend with higher prio wants to offload the op
+            if (src_backend_id == sched->n_backends - 1) {
+                for (int b = 0; b < src_backend_id; b++) {
+                    if (ggml_backend_offload_op(sched->backends[b], tensor)) {
+                        SET_CAUSE(tensor, "1.off");
+                        return b;
+                    }
+                }
+            }
             SET_CAUSE(tensor, "1.wgt%d", i);
-            return src_backend;
+            return src_backend_id;
         }
     }
 
@@ -1227,28 +1248,31 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     // pass 1: assign backends to ops with pre-allocated inputs
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
-        if (tensor_backend_id(leaf) != -1) {
+        int * leaf_backend_id = &tensor_backend_id(leaf);
+        if (*leaf_backend_id != -1) {
             // do not overwrite user assignments
             continue;
         }
-        tensor_backend_id(leaf) = ggml_backend_sched_backend_id_from_cur(sched, leaf);
+        *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
     }
 
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        if (tensor_backend_id(node) != -1) {
+        int * node_backend_id = &tensor_backend_id(node);
+        if (*node_backend_id != -1) {
             // do not overwrite user assignments
             continue;
         }
-        tensor_backend_id(node) = ggml_backend_sched_backend_id_from_cur(sched, node);
+        *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
         // src
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
                 continue;
             }
-            if (tensor_backend_id(src) == -1) {
-                tensor_backend_id(src) = ggml_backend_sched_backend_id_from_cur(sched, src);
+            int * src_backend_id = &tensor_backend_id(src);
+            if (*src_backend_id == -1) {
+                *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
             }
         }
     }
@@ -1270,21 +1294,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (ggml_is_view_op(node->op)) {
                 continue;
             }
-            int tensor_backend_id = tensor_backend_id(node);
-            if (tensor_backend_id != -1) {
-                if (tensor_backend_id == sched->n_backends - 1) {
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
                     // skip cpu (lowest prio backend)
                     cur_backend_id = -1;
                 } else {
-                    cur_backend_id = tensor_backend_id;
+                    cur_backend_id = *node_backend_id;
                 }
             } else {
-                tensor_backend_id(node) = cur_backend_id;
+                *node_backend_id = cur_backend_id;
                 SET_CAUSE(node, "2.2");
             }
         }
     }
-
     // pass 2.1 expand gpu up
     {
         int cur_backend_id = -1;
@@ -1293,22 +1316,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (ggml_is_view_op(node->op)) {
                 continue;
             }
-            int tensor_backend_id = tensor_backend_id(node);
-            if (tensor_backend_id != -1) {
-                if (tensor_backend_id == sched->n_backends - 1) {
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
                     // skip cpu (lowest prio backend)
                     cur_backend_id = -1;
                 } else {
-                    cur_backend_id = tensor_backend_id;
+                    cur_backend_id = *node_backend_id;
                 }
             } else {
-                tensor_backend_id(node) = cur_backend_id;
+                *node_backend_id = cur_backend_id;
                 SET_CAUSE(node, "2.1");
             }
         }
     }
-
-
     // pass 2.4 expand rest down
     {
         int cur_backend_id = -1;
@@ -1317,16 +1338,16 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (ggml_is_view_op(node->op)) {
                 continue;
             }
-            int tensor_backend_id = tensor_backend_id(node);
-            if (tensor_backend_id != -1) {
-                cur_backend_id = tensor_backend_id;
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
             } else {
-                tensor_backend_id(node) = cur_backend_id;
+                *node_backend_id = cur_backend_id;
                 SET_CAUSE(node, "2.4");
             }
         }
     }
-        // pass 2.3 expand rest up
+    // pass 2.3 expand rest up
     {
         int cur_backend_id = -1;
         for (int i = graph->n_nodes - 1; i >= 0; i--) {
@@ -1334,11 +1355,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (ggml_is_view_op(node->op)) {
                 continue;
             }
-            int tensor_backend_id = tensor_backend_id(node);
-            if (tensor_backend_id != -1) {
-                cur_backend_id = tensor_backend_id;
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
             } else {
-                tensor_backend_id(node) = cur_backend_id;
+                *node_backend_id = cur_backend_id;
                 SET_CAUSE(node, "2.3");
             }
         }
@@ -1351,9 +1372,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     // pass 3: assign backends to remaining src from dst and view_src
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        int cur_backend_id = tensor_backend_id(node);
-        if (node->view_src != NULL && cur_backend_id == -1) {
-            cur_backend_id = tensor_backend_id(node) = tensor_backend_id(node->view_src);
+        int * cur_backend_id = &tensor_backend_id(node);
+        if (node->view_src != NULL && *cur_backend_id == -1) {
+            *cur_backend_id = tensor_backend_id(node->view_src);
             SET_CAUSE(node, "3.vsrc");
         }
         for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -1361,14 +1382,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (src == NULL) {
                 continue;
             }
-            int src_backend_id = tensor_backend_id(src);
-            if (src_backend_id == -1) {
+            int * src_backend_id = &tensor_backend_id(src);
+            if (*src_backend_id == -1) {
                 if (src->view_src != NULL) {
                     // views are always on the same backend as the source
-                    tensor_backend_id(src) = tensor_backend_id(src->view_src);
+                    *src_backend_id = tensor_backend_id(src->view_src);
                     SET_CAUSE(src, "3.vsrc");
                 } else {
-                    tensor_backend_id(src) = cur_backend_id;
+                    *src_backend_id = *cur_backend_id;
                     SET_CAUSE(src, "3.cur");
                 }
             }
@@ -1380,19 +1401,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
     // pass 4: split graph, find tensors that need to be copied
     {
-        int cur_split = 0;
+        int i_split = 0;
+        struct ggml_backend_sched_split * split = &sched->splits[0];
         // find the backend of the first split, skipping view ops
         for (int i = 0; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
             if (!ggml_is_view_op(node->op)) {
-                sched->splits[0].backend_id = tensor_backend_id(node);
+                split->backend_id = tensor_backend_id(node);
                 break;
             }
         }
-        sched->splits[0].i_start = 0;
-        sched->splits[0].n_inputs = 0;
-        memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
-        int cur_backend_id = sched->splits[0].backend_id;
+        split->i_start = 0;
+        split->n_inputs = 0;
+        memset(split->inputs, 0, sizeof(split->inputs)); //HACK
+        int cur_backend_id = split->backend_id;
         for (int i = 0; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
 
@@ -1400,18 +1422,54 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 continue;
             }
 
-            int tensor_backend_id = tensor_backend_id(node);
+            const int node_backend_id = tensor_backend_id(node);
 
-            GGML_ASSERT(tensor_backend_id != -1); // all nodes should be assigned by now
+            GGML_ASSERT(node_backend_id != -1); // all nodes should be assigned by now
 
-            if (tensor_backend_id != cur_backend_id) {
-                sched->splits[cur_split].i_end = i;
-                cur_split++;
-                GGML_ASSERT(cur_split < GGML_SCHED_MAX_SPLITS);
-                sched->splits[cur_split].backend_id = tensor_backend_id;
-                sched->splits[cur_split].i_start = i;
-                sched->splits[cur_split].n_inputs = 0;
-                cur_backend_id = tensor_backend_id;
+            // check if we should start a new split based on the sources of the current node
+            bool need_new_split = false;
+            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * src = node->src[j];
+                    if (src == NULL) {
+                        continue;
+                    }
+                    // check if a weight is on a different backend
+                    // by starting a new split, the memory of the previously offloaded weights can be reused
+                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                        int src_backend_id = tensor_backend_id(src);
+                        if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                    // check if the split has too many inputs
+                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                        const size_t id = hash_id(src);
+                        int src_backend_id = sched->tensor_backend_id[id];
+                        if (src_backend_id != cur_backend_id && sched->tensor_copies[hash_id(src)][cur_backend_id][0] == NULL) {
+                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (node_backend_id != cur_backend_id || need_new_split) {
+                split->i_end = i;
+                i_split++;
+                if (i_split >= sched->splits_capacity) {
+                    sched->splits_capacity *= 2;
+                    sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
+                    GGML_ASSERT(sched->splits != NULL);
+                }
+                GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS);
+                split = &sched->splits[i_split];
+                split->backend_id = node_backend_id;
+                split->i_start = i;
+                split->n_inputs = 0;
+                cur_backend_id = node_backend_id;
             }
 
             // find inputs that are not on the same backend
@@ -1421,10 +1479,10 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     continue;
                 }
 
-                int src_backend_id = tensor_backend_id(src);
+                const int src_backend_id = tensor_backend_id(src);
                 assert(src_backend_id != -1); // all inputs should be assigned by now
 
-                if (src->flags & GGML_TENSOR_FLAG_INPUT)  {
+                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1)  {
                     size_t id = hash_id(src);
                     if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
                         ggml_backend_t backend = sched->backends[src_backend_id];
@@ -1441,7 +1499,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
                             sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
-                            tensor_backend_id(tensor_copy) = src_backend_id;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
@@ -1450,9 +1507,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                     }
                 }
 
-                if (src_backend_id != tensor_backend_id) {
+                if (src_backend_id != node_backend_id) {
                     // create a copy of the input in the split's backend
-                    size_t id = hash_id(src);
+                    const size_t id = hash_id(src);
                     if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
                         ggml_backend_t backend = sched->backends[cur_backend_id];
                         for (int c = 0; c < sched->n_copies; c++) {
@@ -1463,76 +1520,42 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
                             sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
-                            tensor_backend_id(tensor_copy) = cur_backend_id;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
-                        int n_inputs = sched->splits[cur_split].n_inputs++;
+                        int n_inputs = split->n_inputs++;
                         GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
-                        sched->splits[cur_split].inputs[n_inputs] = src;
+                        split->inputs[n_inputs] = src;
                     }
                     node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
                 }
             }
         }
-        sched->splits[cur_split].i_end = graph->n_nodes;
-        sched->n_splits = cur_split + 1;
+        split->i_end = graph->n_nodes;
+        sched->n_splits = i_split + 1;
     }
 #ifdef DEBUG_PASS4
     fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif
 
-#ifndef NDEBUG
-    // sanity check: all sources should have the same backend as the node
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-        if (tensor_backend == NULL) {
-            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
-        }
-        if (node->view_src != NULL && tensor_backend != ggml_backend_sched_get_tensor_backend(sched, node->view_src)) {
-            fprintf(stderr, "!!!!!!! %s has backend %s, view_src %s has backend %s\n",
-                node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
-                node->view_src->name, ggml_backend_sched_get_tensor_backend(sched, node->view_src) ?
-                    ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, node->view_src)) : "NULL");
-        }
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
-            if (src_backend != tensor_backend /* && src_backend != NULL */) {
-                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
-                    node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
-                    j, src->name, src_backend ? ggml_backend_name(src_backend) : "NULL");
-            }
-            if (src->view_src != NULL && src_backend != ggml_backend_sched_get_tensor_backend(sched, src->view_src)) {
-                fprintf(stderr, "!!!!!!! [src] %s has backend %s, view_src %s has backend %s\n",
-                    src->name, src_backend ? ggml_backend_name(src_backend) : "NULL",
-                    src->view_src->name, ggml_backend_sched_get_tensor_backend(sched, src->view_src) ?
-                        ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, src->view_src)) : "NULL");
-            }
-        }
-    }
-    fflush(stderr);
-#endif
-
     // create copies of the graph for each split
     // TODO: avoid this copy
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS, false);
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2, false);
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
 
         // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
+            assert(graph_copy->size > (graph_copy->n_nodes + 1));
+
             struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id][sched->cur_copy];
+            const size_t input_id = hash_id(input);
+            struct ggml_tensor * input_cpy = sched->tensor_copies[input_id][split->backend_id][sched->cur_copy];
 
             // add a dependency to the input source so that it is not freed before the copy is done
             struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
             input_dep->src[0] = input;
-            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(input);
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->tensor_backend_id[input_id];
             graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
 
             // add a dependency to the input copy so that it is allocated at the start of the split
@@ -1541,6 +1564,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
 
         for (int j = split->i_start; j < split->i_end; j++) {
+            assert(graph_copy->size > graph_copy->n_nodes);
             sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
             graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
         }
@@ -1625,13 +1649,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 }
                 ggml_backend_tensor_copy(input, input_cpy);
             } else {
+                // wait for the split backend to finish using the input before overwriting it
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
                 } else {
                     ggml_backend_synchronize(split_backend);
-                    ggml_backend_synchronize(input_backend);
                 }
-
                 ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
             }
         }
@@ -1701,17 +1724,21 @@ ggml_backend_sched_t ggml_backend_sched_new(
     struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
 
     // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
+    sched->hash_set          = ggml_hash_set_new(graph_size);
     sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
     sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
-    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), graph_size);
-    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), graph_size);
+
+    const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
+    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
 
     sched->n_backends = n_backends;
 
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
-    GGML_ASSERT(sched->n_copies <= GGML_SCHED_MAX_COPIES);
+    const int initial_splits_capacity = 16;
+    sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
+    sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
@@ -1742,6 +1769,7 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
+    free(sched->splits);
     free(sched->hash_set.keys);
     free(sched->tensor_backend_id);
     free(sched->tensor_copies);
@@ -1762,6 +1790,8 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
 }
 
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes);
+
     ggml_backend_sched_split_graph(sched, measure_graph);
 
     // TODO: extract this to a separate function
@@ -1776,7 +1806,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 }
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes);
 
     ggml_backend_sched_split_graph(sched, graph);
 

ggml-backend.h

@@ -70,11 +70,11 @@ extern "C" {
     GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);

ggml-cuda.h

@@ -17,29 +17,17 @@ extern "C" {
 
 #define GGML_CUDA_MAX_DEVICES       16
 
-// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
-GGML_API GGML_CALL void   ggml_init_cublas(void);
-
-// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
-GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
-
-GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
-GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
-
-GGML_API GGML_CALL bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API GGML_CALL bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-GGML_API GGML_CALL int    ggml_cuda_get_device_count(void);
-GGML_API GGML_CALL void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
-
 // backend API
 GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
 
 GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
 
+// device buffer
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
@@ -47,6 +35,9 @@ GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
+GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-kompute.cpp

@@ -1951,6 +1951,7 @@ static struct ggml_backend_i kompute_backend_i = {
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
     /* .supports_op             = */ ggml_backend_kompute_supports_op,
+    /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,

ggml-metal.m

@@ -173,8 +173,9 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,
-  //GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
-  //GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_CPY_F16_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F16_F32,
     GGML_METAL_KERNEL_TYPE_CONCAT,
@@ -598,8 +599,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,            cpy_f32_iq4_nl,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                 true);
@@ -739,6 +741,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
                            case GGML_TYPE_Q8_0:
                            case GGML_TYPE_Q4_0:
                            case GGML_TYPE_Q4_1:
+                           case GGML_TYPE_Q5_0:
+                           case GGML_TYPE_Q5_1:
+                           case GGML_TYPE_IQ4_NL:
                                 return true;
                            default:
                                 return false;
@@ -1387,6 +1392,14 @@ static enum ggml_status ggml_metal_graph_compute(
                             (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;
+                                default: break;
+                            }
+
                             id<MTLComputePipelineState> pipeline = nil;
 
                             switch (src0->type) {
@@ -1701,6 +1714,14 @@ static enum ggml_status ggml_metal_graph_compute(
                             ne20 % 32 == 0 && ne20 >= 64 &&
                             ne11 > ne11_mm_min) {
 
+                            // 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;
+                                default: break;
+                            }
+
                             id<MTLComputePipelineState> pipeline = nil;
 
                             switch (src2->type) {
@@ -2431,13 +2452,14 @@ static enum ggml_status ggml_metal_graph_compute(
                                     GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
 
                                     switch (dstt) {
-                                        case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline;  break;
-                                        case GGML_TYPE_F32:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;  break;
-                                        case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break;
-                                        case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break;
-                                        case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break;
-                                      //case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
-                                      //case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
+                                        case GGML_TYPE_F16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline;  break;
+                                        case GGML_TYPE_F32:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;  break;
+                                        case GGML_TYPE_Q8_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break;
+                                        case GGML_TYPE_Q4_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break;
+                                        case GGML_TYPE_Q4_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break;
+                                        case GGML_TYPE_Q5_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
+                                        case GGML_TYPE_Q5_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
+                                        case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
                                         default: GGML_ASSERT(false && "not implemented");
                                     };
                                 } break;
@@ -2837,6 +2859,7 @@ static struct ggml_backend_i ggml_backend_metal_i = {
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_metal_graph_compute,
     /* .supports_op             = */ ggml_backend_metal_supports_op,
+    /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,

ggml-metal.metal

@@ -2388,6 +2388,242 @@ kernel void kernel_cpy_f32_q4_1(
     }
 }
 
+kernel void kernel_cpy_f32_q5_0(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
+
+    device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK5_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / -16;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK5_0].d = d;
+
+        uint32_t qh = 0;
+        for (int j = 0; j < QK5_0/2; ++j) {
+            const float x0 = src[0       + j]*id;
+            const float x1 = src[QK5_0/2 + j]*id;
+
+            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+        }
+        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+        for (int j = 0; j < 4; ++j) {
+            dst_data[i00/QK5_0].qh[j] = qh8[j];
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q5_1(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
+
+    device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float max = src[0];
+        float min = src[0];
+
+        for (int j = 1; j < QK5_1; j++) {
+            const float v = src[j];
+            min = v < min ? v : min;
+            max = v > max ? v : max;
+        }
+
+        const float d = (max - min) / 31;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK5_1].d = d;
+        dst_data[i00/QK5_1].m = min;
+
+        uint32_t qh = 0;
+        for (int j = 0; j < QK5_1/2; ++j) {
+            const float x0 = (src[0       + j] - min)*id;
+            const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+        }
+        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+        for (int j = 0; j < 4; ++j) {
+            dst_data[i00/QK5_1].qh[j] = qh8[j];
+        }
+    }
+}
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+kernel void kernel_cpy_f32_iq4_nl(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
+
+    device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK4_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / kvalues_iq4nl_f[0];
+        const float id = d ? 1.0f/d : 0.0f;
+
+        float sumqx = 0, sumq2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            const float x0 = src[0        + j]*id;
+            const float x1 = src[QK4_NL/2 + j]*id;
+
+            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
+
+            const float v0 = kvalues_iq4nl_f[xi0];
+            const float v1 = kvalues_iq4nl_f[xi1];
+            const float w0 = src[0        + j]*src[0        + j];
+            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+            sumq2 += w0*v0*v0 + w1*v1*v1;
+
+        }
+
+        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+
+    }
+}
+
 kernel void kernel_concat(
     device  const char * src0,
     device  const char * src1,
@@ -4220,10 +4456,6 @@ void kernel_mul_mv_iq1_s_f32_impl(
     }
 }
 
-constexpr constant static float kvalues_iq4nl_f[16] = {
-    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
-};
-
 void kernel_mul_mv_iq4_nl_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -5528,18 +5760,18 @@ typedef void (mat_mm_t)(
         threadgroup   uchar *,
         uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_f32_f32")]]  kernel mat_mm_t kernel_mul_mm<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_f16_f32")]]  kernel mat_mm_t kernel_mul_mm<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mat_mm_t kernel_mul_mm<float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mat_mm_t kernel_mul_mm<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mat_mm_t kernel_mul_mm<block_q6_K,    QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
@@ -5588,18 +5820,18 @@ typedef void (mat_mm_id_t)(
         threadgroup    uchar *,
         uint3, uint, uint);
 
-template [[host_name("kernel_mul_mm_id_f32_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<float4x4,   1,     dequantize_f32>;
-template [[host_name("kernel_mul_mm_id_f16_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<half4x4,    1,     dequantize_f16>;
-template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2,     dequantize_q4_0>;
-template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2,     dequantize_q4_1>;
-template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2,     dequantize_q5_0>;
-template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2,     dequantize_q5_1>;
-template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2,     dequantize_q8_0>;
-template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
-template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
-template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
-template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
-template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K,    QK_NL, dequantize_q6_K>;
 template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
 template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;

ggml-quants.c

@@ -11705,9 +11705,8 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
         ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
         float * scales, float * weight, uint8_t * L,
         const int8_t * values,
-        const float * quant_weights) {
-
-    const int ntry = 7;
+        const float * quant_weights,
+        const int ntry) {
 
     float sigma2 = 0;
     for (int j = 0; j < super_block_size; ++j) sigma2 += x[j]*x[j];
@@ -11719,6 +11718,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
     float max_scale = 0, amax_scale = 0;
     for (int ib = 0; ib < super_block_size/block_size; ++ib) {
         const float * xb = x + ib*block_size;
+        uint8_t * Lb = L + ib*block_size;
         if (quant_weights) {
             const float * qw = quant_weights + ib*block_size;
             for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
@@ -11736,12 +11736,13 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
             scales[ib] = 0;
             continue;
         }
-        float d = -max/values[0];
+        float d = ntry > 0 ? -max/values[0] : max/values[0];
         float id = 1/d;
         float sumqx = 0, sumq2 = 0;
         for (int j = 0; j < block_size; ++j) {
             float al = id*xb[j];
             int l = best_index_int8(16, values, al);
+            Lb[j] = l;
             float q = values[l];
             float w = weight[j];
             sumqx += w*q*xb[j];
@@ -11796,9 +11797,11 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
         }
     } else {
         dh[0] = GGML_FP32_TO_FP16(scales[0]);
-        float id = scales[0] ? 1/scales[0] : 0;
-        for (int j = 0; j < super_block_size; ++j) {
-            L[j] = best_index_int8(16, values, id*x[j]);
+        if (ntry > 0) {
+            float id = scales[0] ? 1/scales[0] : 0;
+            for (int j = 0; j < super_block_size; ++j) {
+                L[j] = best_index_int8(16, values, id*x[j]);
+            }
         }
     }
 
@@ -11823,7 +11826,7 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int nrow
         for (int ibl = 0; ibl < nblock; ++ibl) {
             const float * qw = quant_weights ? quant_weights + QK4_NL*ibl : NULL;
             quantize_row_iq4_nl_impl(QK4_NL, 32, src + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
-                    &scale, weight, L, kvalues_iq4nl, qw);
+                    &scale, weight, L, kvalues_iq4nl, qw, 7);
         }
         src += n_per_row;
         qrow += nblock*sizeof(block_iq4_nl);
@@ -11832,14 +11835,23 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int nrow
 }
 
 void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK4_NL == 0);
-    block_iq4_nl * restrict y = vy;
-    quantize_row_iq4_nl_reference(x, y, k);
+    GGML_ASSERT(k%QK4_NL == 0);
+    int nblock = k/QK4_NL;
+    uint8_t L[QK4_NL];
+    float weight[QK4_NL];
+    uint16_t unused_h;
+    uint8_t * unused_l = NULL;
+    float scale;
+    block_iq4_nl * iq4 = (block_iq4_nl *)vy;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
+                &scale, weight, L, kvalues_iq4nl, NULL, -1);
+    }
 }
 
 void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int k) {
     assert(k % QK4_NL == 0);
-    quantize_iq4_nl(x, y, 1, k, NULL);
+    quantize_row_iq4_nl(x, y, k);
 }
 
 size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) {
@@ -11857,7 +11869,7 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int nrow
         for (int ibl = 0; ibl < nblock; ++ibl) {
             const float * qw = quant_weights ? quant_weights + QK_K*ibl : NULL;
             quantize_row_iq4_nl_impl(QK_K, 32, src + QK_K*ibl, &iq4[ibl].d, iq4[ibl].qs, &iq4[ibl].scales_h, iq4[ibl].scales_l,
-                    scales, weight, L, kvalues_iq4nl, qw);
+                    scales, weight, L, kvalues_iq4nl, qw, 7);
         }
         src += n_per_row;
         qrow += nblock*sizeof(block_iq4_xs);

ggml-sycl.cpp

@@ -16,6 +16,7 @@
 #include <cinttypes>
 #include <cstddef>
 #include <cstdint>
+#include <cstdlib>
 #include <float.h>
 #include <limits>
 #include <stdint.h>
@@ -24,10 +25,9 @@
 #include <cmath>
 #include <iostream>
 #include <fstream>
-
 #include <stdio.h>
 #include <stdlib.h>
-
+#include <regex>
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
@@ -82,6 +82,30 @@ Following definition copied from DPCT head files, which are used by ggml-sycl.cp
 #define __dpct_noinline__ __attribute__((noinline))
 #endif
 
+
+std::string get_device_type_name(const sycl::device &Device) {
+    auto DeviceType = Device.get_info<sycl::info::device::device_type>();
+    switch (DeviceType) {
+    case sycl::info::device_type::cpu:
+        return "cpu";
+    case sycl::info::device_type::gpu:
+        return "gpu";
+    case sycl::info::device_type::host:
+        return "host";
+    case sycl::info::device_type::accelerator:
+        return "acc";
+    default:
+        return "unknown";
+    }
+}
+
+std::string get_device_backend_and_type(const sycl::device &device) {
+    std::stringstream device_type;
+    sycl::backend backend = device.get_backend();
+    device_type <<  backend << ":" << get_device_type_name(device);
+    return device_type.str();
+}
+
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;
@@ -716,11 +740,7 @@ namespace dpct
 
         sycl::queue &default_queue()
         {
-#ifdef DPCT_USM_LEVEL_NONE
-            return out_of_order_queue();
-#else
             return in_order_queue();
-#endif // DPCT_USM_LEVEL_NONE
         }
 
         void queues_wait_and_throw()
@@ -739,11 +759,7 @@ namespace dpct
 
         sycl::queue *create_queue(bool enable_exception_handler = false)
         {
-#ifdef DPCT_USM_LEVEL_NONE
-            return create_out_of_order_queue(enable_exception_handler);
-#else
             return create_in_order_queue(enable_exception_handler);
-#endif // DPCT_USM_LEVEL_NONE
         }
 
         sycl::queue *create_queue(sycl::context context, sycl::device device,
@@ -942,17 +958,67 @@ namespace dpct
 
     private:
         mutable std::recursive_mutex m_mutex;
+        static bool compare_dev(sycl::device &device1, sycl::device &device2)
+        {
+            dpct::device_info prop1;
+            dpct::get_device_info(prop1, device1);
+            dpct::device_info prop2;
+            dpct::get_device_info(prop2, device2);
+            return prop1.get_max_compute_units() > prop2.get_max_compute_units();
+        }
+        static int convert_backend_index(std::string & backend) {
+            if (backend == "ext_oneapi_level_zero:gpu") return 0;
+            if (backend == "opencl:gpu") return 1;
+            if (backend == "ext_oneapi_cuda:gpu") return 2;
+            if (backend == "ext_oneapi_hip:gpu") return 3;
+            if (backend == "opencl:cpu") return 4;
+            if (backend == "opencl:acc") return 5;
+            printf("convert_backend_index: can't handle backend=%s\n", backend.c_str());
+            GGML_ASSERT(false);
+        }
+        static bool compare_backend(std::string &backend1, std::string &backend2) {
+            return convert_backend_index(backend1) < convert_backend_index(backend2);
+        }
         dev_mgr()
         {
             sycl::device default_device =
                 sycl::device(sycl::default_selector_v);
             _devs.push_back(std::make_shared<device_ext>(default_device));
 
-            std::vector<sycl::device> sycl_all_devs =
-                sycl::device::get_devices(sycl::info::device_type::all);
+            std::vector<sycl::device> sycl_all_devs;
             // Collect other devices except for the default device.
             if (default_device.is_cpu())
                 _cpu_device = 0;
+
+            auto Platforms = sycl::platform::get_platforms();
+            // Keep track of the number of devices per backend
+            std::map<sycl::backend, size_t> DeviceNums;
+            std::map<std::string, std::vector<sycl::device>> backend_devices;
+
+            while (!Platforms.empty()) {
+                auto Platform = Platforms.back();
+                Platforms.pop_back();
+                auto devices = Platform.get_devices();
+                std::string backend_type = get_device_backend_and_type(devices[0]);
+                for (const auto &device : devices) {
+                    backend_devices[backend_type].push_back(device);
+                }
+            }
+
+            std::vector<std::string> keys;
+            for(auto it = backend_devices.begin(); it != backend_devices.end(); ++it) {
+                keys.push_back(it->first);
+            }
+            std::sort(keys.begin(), keys.end(), compare_backend);
+
+            for (auto &key : keys) {
+                std::vector<sycl::device> devs = backend_devices[key];
+                std::sort(devs.begin(), devs.end(), compare_dev);
+                for (const auto &dev : devs) {
+                    sycl_all_devs.push_back(dev);
+                }
+            }
+
             for (auto &dev : sycl_all_devs)
             {
                 if (dev == default_device)
@@ -1001,11 +1067,6 @@ namespace dpct
         static pointer_access_attribute get_pointer_attribute(sycl::queue &q,
                                                               const void *ptr)
         {
-#ifdef DPCT_USM_LEVEL_NONE
-            return mem_mgr::instance().is_device_ptr(ptr)
-                       ? pointer_access_attribute::device_only
-                       : pointer_access_attribute::host_only;
-#else
             switch (sycl::get_pointer_type(ptr, q.get_context()))
             {
             case sycl::usm::alloc::unknown:
@@ -1016,7 +1077,6 @@ namespace dpct
             case sycl::usm::alloc::host:
                 return pointer_access_attribute::host_device;
             }
-#endif
         }
 
         template <typename ArgT>
@@ -1199,11 +1259,7 @@ namespace dpct
 
         static inline void *dpct_malloc(size_t size, sycl::queue &q)
         {
-#ifdef DPCT_USM_LEVEL_NONE
-            return mem_mgr::instance().mem_alloc(size * sizeof(byte_t));
-#else
             return sycl::malloc_device(size, q.get_device(), q.get_context());
-#endif // DPCT_USM_LEVEL_NONE
         }
 
 #define PITCH_DEFAULT_ALIGN(x) (((x) + 31) & ~(0x1F))
@@ -1227,25 +1283,7 @@ namespace dpct
         static inline sycl::event dpct_memset(sycl::queue &q, void *dev_ptr,
                                               valueT value, size_t size)
         {
-#ifdef DPCT_USM_LEVEL_NONE
-            auto &mm = mem_mgr::instance();
-            assert(mm.is_device_ptr(dev_ptr));
-            auto alloc = mm.translate_ptr(dev_ptr);
-            size_t offset = (valueT *)dev_ptr - (valueT *)alloc.alloc_ptr;
-
-            return q.submit([&](sycl::handler &cgh)
-                            {
-    auto r = sycl::range<1>(size);
-    auto o = sycl::id<1>(offset);
-    auto new_buffer = alloc.buffer.reinterpret<valueT>(
-        sycl::range<1>(alloc.size / sizeof(valueT)));
-    sycl::accessor<valueT, 1, sycl::access_mode::write,
-                sycl::access::target::device>
-        acc(new_buffer, cgh, r, o);
-    cgh.fill(acc, value); });
-#else
             return q.fill(dev_ptr, value, size);
-#endif // DPCT_USM_LEVEL_NONE
         }
 
         /**
@@ -1339,72 +1377,8 @@ namespace dpct
         {
             if (!size)
                 return sycl::event{};
-#ifdef DPCT_USM_LEVEL_NONE
-            auto &mm = mem_mgr::instance();
-            auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction);
-
-            switch (real_direction)
-            {
-            case host_to_host:
-                return q.submit([&](sycl::handler &cgh)
-                                {
-    cgh.depends_on(dep_events);
-    cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); });
-            case host_to_device:
-            {
-                auto alloc = mm.translate_ptr(to_ptr);
-                size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr;
-                return q.submit([&](sycl::handler &cgh)
-                                {
-    cgh.depends_on(dep_events);
-    auto r = sycl::range<1>(size);
-    auto o = sycl::id<1>(offset);
-    sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                        sycl::access::target::device>
-        acc(alloc.buffer, cgh, r, o);
-    cgh.copy(from_ptr, acc); });
-            }
-            case device_to_host:
-            {
-                auto alloc = mm.translate_ptr(from_ptr);
-                size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr;
-                return q.submit([&](sycl::handler &cgh)
-                                {
-    cgh.depends_on(dep_events);
-    auto r = sycl::range<1>(size);
-    auto o = sycl::id<1>(offset);
-    sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                        sycl::access::target::device>
-        acc(alloc.buffer, cgh, r, o);
-    cgh.copy(acc, to_ptr); });
-            }
-            case device_to_device:
-            {
-                auto to_alloc = mm.translate_ptr(to_ptr);
-                auto from_alloc = mm.translate_ptr(from_ptr);
-                size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr;
-                size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr;
-                return q.submit([&](sycl::handler &cgh)
-                                {
-    cgh.depends_on(dep_events);
-    auto r = sycl::range<1>(size);
-    auto to_o = sycl::id<1>(to_offset);
-    auto from_o = sycl::id<1>(from_offset);
-    sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                        sycl::access::target::device>
-        to_acc(to_alloc.buffer, cgh, r, to_o);
-    sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                        sycl::access::target::device>
-        from_acc(from_alloc.buffer, cgh, r, from_o);
-    cgh.copy(from_acc, to_acc); });
-            }
-            default:
-                throw std::runtime_error("dpct_memcpy: invalid direction value");
-            }
-#else
             return q.memcpy(to_ptr, from_ptr, size, dep_events);
             GGML_UNUSED(direction);
-#endif // DPCT_USM_LEVEL_NONE
         }
 
         // Get actual copy range and make sure it will not exceed range.
@@ -1544,45 +1518,15 @@ namespace dpct
                 break;
             }
             case device_to_device:
-#ifdef DPCT_USM_LEVEL_NONE
-            {
-                auto &mm = mem_mgr::instance();
-                auto to_alloc = mm.translate_ptr(to_surface);
-                auto from_alloc = mm.translate_ptr(from_surface);
-                size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr;
-                size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr;
-                event_list.push_back(q.submit([&](sycl::handler &cgh)
-                                              {
-    cgh.depends_on(dep_events);
-    auto to_o = sycl::id<1>(to_offset);
-    auto from_o = sycl::id<1>(from_offset);
-    sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                        sycl::access::target::device>
-        to_acc(to_alloc.buffer, cgh,
-                get_copy_range(size, to_slice, to_range.get(0)), to_o);
-    sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                        sycl::access::target::device>
-        from_acc(from_alloc.buffer, cgh,
-                get_copy_range(size, from_slice, from_range.get(0)), from_o);
-    cgh.parallel_for<class dpct_memcpy_3d_detail_usmnone>(
-        size,
-        [=](sycl::id<3> id) {
-            to_acc[get_offset(id, to_slice, to_range.get(0))] =
-                from_acc[get_offset(id, from_slice, from_range.get(0))];
-        }); }));
-            }
-#else
-                event_list.push_back(q.submit([&](sycl::handler &cgh)
-                                              {
-    cgh.depends_on(dep_events);
-    cgh.parallel_for<class dpct_memcpy_3d_detail>(
-        size,
-        [=](sycl::id<3> id) {
-            to_surface[get_offset(id, to_slice, to_range.get(0))] =
-                from_surface[get_offset(id, from_slice, from_range.get(0))];
-        }); }));
-#endif
-            break;
+                event_list.push_back(q.submit([&](sycl::handler &cgh){
+                cgh.depends_on(dep_events);
+                cgh.parallel_for<class dpct_memcpy_3d_detail>(
+                    size,
+                    [=](sycl::id<3> id) {
+                        to_surface[get_offset(id, to_slice, to_range.get(0))] =
+                            from_surface[get_offset(id, from_slice, from_range.get(0))];
+                    }); }));
+                break;
             default:
                 throw std::runtime_error("dpct_memcpy: invalid direction value");
             }
@@ -1680,11 +1624,7 @@ namespace dpct
         {
             if (ptr)
             {
-#ifdef DPCT_USM_LEVEL_NONE
-                detail::mem_mgr::instance().mem_free(ptr);
-#else
                 sycl::free(ptr, q.get_context());
-#endif // DPCT_USM_LEVEL_NONE
             }
         }
 
@@ -1692,11 +1632,7 @@ namespace dpct
         inline auto get_memory(const void *x)
         {
             T *new_x = reinterpret_cast<T *>(const_cast<void *>(x));
-#ifdef DPCT_USM_LEVEL_NONE
-            return dpct::get_buffer<std::remove_cv_t<T>>(new_x);
-#else
             return new_x;
-#endif
         }
 
         template <typename T>
@@ -2148,72 +2084,8 @@ namespace dpct
     {
         if (!size)
             return sycl::event{};
-#ifdef DPCT_USM_LEVEL_NONE
-        auto &mm = mem_mgr::instance();
-        auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction);
-
-        switch (real_direction)
-        {
-        case host_to_host:
-            return q.submit([&](sycl::handler &cgh)
-                            {
-        cgh.depends_on(dep_events);
-        cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); });
-        case host_to_device:
-        {
-            auto alloc = mm.translate_ptr(to_ptr);
-            size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr;
-            return q.submit([&](sycl::handler &cgh)
-                            {
-        cgh.depends_on(dep_events);
-        auto r = sycl::range<1>(size);
-        auto o = sycl::id<1>(offset);
-        sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                            sycl::access::target::device>
-            acc(alloc.buffer, cgh, r, o);
-        cgh.copy(from_ptr, acc); });
-        }
-        case device_to_host:
-        {
-            auto alloc = mm.translate_ptr(from_ptr);
-            size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr;
-            return q.submit([&](sycl::handler &cgh)
-                            {
-        cgh.depends_on(dep_events);
-        auto r = sycl::range<1>(size);
-        auto o = sycl::id<1>(offset);
-        sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                            sycl::access::target::device>
-            acc(alloc.buffer, cgh, r, o);
-        cgh.copy(acc, to_ptr); });
-        }
-        case device_to_device:
-        {
-            auto to_alloc = mm.translate_ptr(to_ptr);
-            auto from_alloc = mm.translate_ptr(from_ptr);
-            size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr;
-            size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr;
-            return q.submit([&](sycl::handler &cgh)
-                            {
-        cgh.depends_on(dep_events);
-        auto r = sycl::range<1>(size);
-        auto to_o = sycl::id<1>(to_offset);
-        auto from_o = sycl::id<1>(from_offset);
-        sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                            sycl::access::target::device>
-            to_acc(to_alloc.buffer, cgh, r, to_o);
-        sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                            sycl::access::target::device>
-            from_acc(from_alloc.buffer, cgh, r, from_o);
-        cgh.copy(from_acc, to_acc); });
-        }
-        default:
-            throw std::runtime_error("dpct_memcpy: invalid direction value");
-        }
-#else
         return q.memcpy(to_ptr, from_ptr, size, dep_events);
         GGML_UNUSED(direction);
-#endif // DPCT_USM_LEVEL_NONE
     }
 
     // Get actual copy range and make sure it will not exceed range.
@@ -2353,34 +2225,6 @@ namespace dpct
             break;
         }
         case device_to_device:
-#ifdef DPCT_USM_LEVEL_NONE
-        {
-            auto &mm = mem_mgr::instance();
-            auto to_alloc = mm.translate_ptr(to_surface);
-            auto from_alloc = mm.translate_ptr(from_surface);
-            size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr;
-            size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr;
-            event_list.push_back(q.submit([&](sycl::handler &cgh)
-                                          {
-        cgh.depends_on(dep_events);
-        auto to_o = sycl::id<1>(to_offset);
-        auto from_o = sycl::id<1>(from_offset);
-        sycl::accessor<byte_t, 1, sycl::access_mode::write,
-                            sycl::access::target::device>
-            to_acc(to_alloc.buffer, cgh,
-                    get_copy_range(size, to_slice, to_range.get(0)), to_o);
-        sycl::accessor<byte_t, 1, sycl::access_mode::read,
-                            sycl::access::target::device>
-            from_acc(from_alloc.buffer, cgh,
-                    get_copy_range(size, from_slice, from_range.get(0)), from_o);
-        cgh.parallel_for<class dpct_memcpy_3d_detail_usmnone>(
-            size,
-            [=](sycl::id<3> id) {
-                to_acc[get_offset(id, to_slice, to_range.get(0))] =
-                    from_acc[get_offset(id, from_slice, from_range.get(0))];
-            }); }));
-        }
-#else
             event_list.push_back(q.submit([&](sycl::handler &cgh)
                                           {
         cgh.depends_on(dep_events);
@@ -2390,7 +2234,6 @@ namespace dpct
                 to_surface[get_offset(id, to_slice, to_range.get(0))] =
                     from_surface[get_offset(id, from_slice, from_range.get(0))];
             }); }));
-#endif
         break;
         default:
             throw std::runtime_error("dpct_memcpy: invalid direction value");
@@ -2581,9 +2424,6 @@ namespace dpct
                            void *c[], library_data_t c_type, int ldc,
                            int batch_size, library_data_t scaling_type)
     {
-#ifdef DPCT_USM_LEVEL_NONE
-        throw std::runtime_error("this API is unsupported when USM level is none");
-#else
         if (scaling_type == library_data_t::real_float &&
             c_type == library_data_t::complex_float)
         {
@@ -2718,7 +2558,6 @@ namespace dpct
         default:
             throw std::runtime_error("the combination of data type is unsupported");
         }
-#endif
     }
 
     /// Computes a batch of matrix-matrix product with general matrices.
@@ -3057,24 +2896,9 @@ namespace dpct
             template <size_t D = Dimension>
             typename std::enable_if<D == 1, T>::type &operator[](size_t index) {
                 init();
-        #ifdef DPCT_USM_LEVEL_NONE
-                return dpct::get_buffer<typename std::enable_if<D == 1, T>::type>(
-                        _device_ptr)
-                    .template get_access<sycl::access_mode::read_write>()[index];
-        #else
                 return _device_ptr[index];
-        #endif // DPCT_USM_LEVEL_NONE
             }
 
-        #ifdef DPCT_USM_LEVEL_NONE
-            /// Get sycl::accessor for the device memory object when usm is not used.
-            accessor_t get_access(sycl::handler &cgh) {
-                return get_buffer(_device_ptr)
-                    .template reinterpret<T, Dimension>(_range)
-                    .template get_access<detail::memory_traits<Memory, T>::mode,
-                                        detail::memory_traits<Memory, T>::target>(cgh);
-            }
-        #else
             /// Get dpct::accessor with dimension info for the device memory object
             /// when usm is used and dimension is greater than 1.
             template <size_t D = Dimension>
@@ -3082,7 +2906,6 @@ namespace dpct
             get_access(sycl::handler &cgh) {
                 return dpct_accessor_t((T *)_device_ptr, _range);
             }
-        #endif // DPCT_USM_LEVEL_NONE
 
         private:
             device_memory(value_t *memory_ptr, size_t size)
@@ -3127,15 +2950,6 @@ namespace dpct
 
             /// Default constructor
             device_memory() : base(1) {}
-
-        #ifdef DPCT_USM_LEVEL_NONE
-            /// Get sycl::accessor for the device memory object when usm is not used.
-            accessor_t get_access(sycl::handler &cgh) {
-                auto buf = get_buffer(base::get_ptr())
-                            .template reinterpret<T, 1>(sycl::range<1>(1));
-                return accessor_t(buf, cgh);
-            }
-        #endif // DPCT_USM_LEVEL_NONE
         };
         } // namespace detail
 
@@ -3202,6 +3016,11 @@ static int g_work_group_size = 0;
 #define GGML_SYCL_MMV_Y 1
 #endif
 
+enum ggml_sycl_backend_gpu_mode {
+    SYCL_UNSET_GPU_MODE = -1,
+    SYCL_SINGLE_GPU_MODE = 0,
+    SYCL_MUL_GPU_MODE
+};
 
 static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
@@ -3401,12 +3220,31 @@ class sycl_gpu_mgr {
         int work_group_size = 0;
         std::string gpus_list = "";
 
+        /*
+        Use all GPUs with same top max compute units
+        */
         sycl_gpu_mgr() {
             detect_sycl_gpu_list_with_max_cu();
             get_allow_gpus();
             create_context_with_gpus();
         }
 
+        /*
+        Only use the assigned GPU
+        */
+        sycl_gpu_mgr(int main_gpu_id) {
+            sycl::device device = dpct::dev_mgr::instance().get_device(main_gpu_id);
+            dpct::device_info prop;
+            dpct::get_device_info(prop, device);
+            gpus.push_back(main_gpu_id);
+            devices.push_back(device);
+            work_group_size = prop.get_max_work_group_size();
+            max_compute_units = prop.get_max_compute_units();
+
+            get_allow_gpus();
+            create_context_with_gpus();
+        }
+
         void create_context_with_gpus() {
             sycl::context ctx = sycl::context(devices);
             assert(gpus.size() > 0);
@@ -3422,7 +3260,7 @@ class sycl_gpu_mgr {
                 gpus_list += std::to_string(gpus[i]);
                 gpus_list += ",";
             }
-            if (gpus_list.length() > 2) {
+            if (gpus_list.length() > 1) {
                 gpus_list.pop_back();
             }
         }
@@ -3451,7 +3289,7 @@ class sycl_gpu_mgr {
                 dpct::device_info prop;
                 dpct::get_device_info(prop, device);
                 if (max_compute_units == prop.get_max_compute_units() &&
-                    prop.get_major_version() == 1) {
+                    is_ext_oneapi_device(device)) {
                     gpus.push_back(id);
                     devices.push_back(device);
                     work_group_size = prop.get_max_work_group_size();
@@ -3471,8 +3309,8 @@ class sycl_gpu_mgr {
                 if (gpus[i] == id)
                     return i;
             }
-            assert(false);
-            return -1;
+            printf("miss to get device index by id=%d\n", id);
+            GGML_ASSERT(false);
         }
 
         int get_next_index(int id) {
@@ -3481,8 +3319,16 @@ class sycl_gpu_mgr {
                 if (gpus[i] == id)
                     return i;
             }
-            assert(false);
-            return -1;
+            GGML_ASSERT(false);
+        }
+
+        bool is_ext_oneapi_device(const sycl::device &dev) {
+            sycl::backend dev_backend = dev.get_backend();
+            if (dev_backend == sycl::backend::ext_oneapi_level_zero ||
+                dev_backend == sycl::backend::ext_oneapi_cuda ||
+                dev_backend == sycl::backend::ext_oneapi_hip)
+                return true;
+            return false;
         }
 };
 
@@ -3491,11 +3337,14 @@ static int g_device_count = -1;
 static int g_all_sycl_device_count = -1;
 static int g_main_device = -1;
 static int g_main_device_id = -1;
+static bool g_ggml_backend_sycl_buffer_type_initialized = false;
 
 static std::array<float, GGML_SYCL_MAX_DEVICES> g_default_tensor_split = {};
 
 static float g_tensor_split[GGML_SYCL_MAX_DEVICES] = {0};
 
+static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode = SYCL_UNSET_GPU_MODE;
+
 struct sycl_device_capabilities {
     int     cc;                 // compute capability
     bool    vmm;                // virtual memory support
@@ -12999,17 +12848,20 @@ bool ggml_sycl_loaded(void) {
     return g_sycl_loaded;
 }
 
-void print_device_detail(int id) {
+void print_device_detail(int id, sycl::device &device, std::string device_type) {
+
     dpct::device_info prop;
     SYCL_CHECK(CHECK_TRY_ERROR(
-        dpct::get_device_info(prop, dpct::dev_mgr::instance().get_device(id))));
-    sycl::device cur_device = dpct::dev_mgr::instance().get_device(id);
+        dpct::get_device_info(prop, device)));
+
     std::string version;
     version += std::to_string(prop.get_major_version());
     version += ".";
     version += std::to_string(prop.get_minor_version());
 
-    fprintf(stderr, "|%2d|%45s|%18s|%17d|%14d|%13d|%15lu|\n", id,
+    device_type = std::regex_replace(device_type, std::regex("ext_oneapi_"), "");
+
+    fprintf(stderr, "|%2d|%18s|%45s|%10s|%11d|%8d|%7d|%15lu|\n", id, device_type.c_str(),
             prop.get_name(), version.c_str(), prop.get_max_compute_units(),
             prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
             prop.get_global_mem_size());
@@ -13017,19 +12869,35 @@ void print_device_detail(int id) {
 
 void ggml_backend_sycl_print_sycl_devices() {
     int device_count = dpct::dev_mgr::instance().device_count();
+    std::map<std::string, size_t> DeviceNums;
     fprintf(stderr, "found %d SYCL devices:\n", device_count);
-    fprintf(stderr, "|ID| Name                                        |compute capability|Max compute units|Max work group|Max sub group|Global mem size|\n");
-    fprintf(stderr, "|--|---------------------------------------------|------------------|-----------------|--------------|-------------|---------------|\n");
+    fprintf(stderr, "|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |\n");
+    fprintf(stderr, "|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|\n");
+    fprintf(stderr, "|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|\n");
     for (int id = 0; id < device_count; ++id) {
-        print_device_detail(id);
+        sycl::device device = dpct::dev_mgr::instance().get_device(id);
+        sycl::backend backend = device.get_backend();
+        std::string backend_type = get_device_backend_and_type(device);
+        int type_id=DeviceNums[backend_type]++;
+        std::stringstream device_type;
+        device_type << "[" <<  backend_type << ":" << std::to_string(type_id) << "]";
+        print_device_detail(id, device, device_type.str());
     }
 }
 
 void print_gpu_device_list() {
-    fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n",
-            g_sycl_gpu_mgr->get_gpu_count(),
-            g_sycl_gpu_mgr->gpus_list.c_str(),
-            g_sycl_gpu_mgr->max_compute_units);
+    GGML_ASSERT(g_sycl_gpu_mgr);
+
+    char* hint=NULL;
+    if (g_ggml_sycl_backend_gpu_mode == SYCL_SINGLE_GPU_MODE) {
+        hint = "use %d SYCL GPUs: [%s] with Max compute units:%d\n";
+    } else {
+        hint = "detect %d SYCL GPUs: [%s] with top Max compute units:%d\n";
+    }
+    fprintf(stderr, hint,
+        g_sycl_gpu_mgr->get_gpu_count(),
+        g_sycl_gpu_mgr->gpus_list.c_str(),
+        g_sycl_gpu_mgr->max_compute_units);
 }
 
 int get_sycl_env(const char *env_name, int default_val) {
@@ -13053,7 +12921,7 @@ int get_work_group_size(int user_device_id) {
     return prop.get_max_work_group_size();
 }
 
-void ggml_init_sycl() try {
+static void ggml_init_sycl() try {
     static bool initialized = false;
 
     if (!initialized) {
@@ -13065,23 +12933,6 @@ void ggml_init_sycl() try {
 #else
         fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
 #endif
-        if (CHECK_TRY_ERROR(g_all_sycl_device_count =
-                            dpct::dev_mgr::instance().device_count()) != 0) {
-            initialized = true;
-            g_sycl_loaded = false;
-            return;
-        }
-        GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
-        ggml_backend_sycl_print_sycl_devices();
-
-        if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
-
-        g_device_count = g_sycl_gpu_mgr->get_gpu_count();
-        g_work_group_size = g_sycl_gpu_mgr->work_group_size;
-
-        print_gpu_device_list();
-
-        int64_t total_vram = 0;
 
 /* NOT REMOVE, keep it for next optimize for XMX.
 #if defined(SYCL_USE_XMX)
@@ -13090,49 +12941,15 @@ void ggml_init_sycl() try {
         fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
 #endif
 */
-        for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) {
-            g_device_caps[id].vmm = 0;
-            g_device_caps[id].device_id = -1;
-            g_device_caps[id].cc = 0;
-            g_tensor_split[id] = 0;
-            g_default_tensor_split[id] = 0;
+
+        if (CHECK_TRY_ERROR(g_all_sycl_device_count =
+                            dpct::dev_mgr::instance().device_count()) != 0) {
+            initialized = true;
+            g_sycl_loaded = false;
+            return;
         }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            int device_id = g_sycl_gpu_mgr->gpus[i];
-            g_device_caps[i].vmm = 0;
-
-            dpct::device_info prop;
-            SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
-                prop, dpct::dev_mgr::instance().get_device(device_id))));
-
-            g_default_tensor_split[i] = total_vram;
-            total_vram += prop.get_global_mem_size();
-
-            g_device_caps[i].cc =
-                100 * prop.get_major_version() + 10 * prop.get_minor_version();
-        }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            g_default_tensor_split[i] /= total_vram;
-        }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            SYCL_CHECK(ggml_sycl_set_device(i));
-
-            // create sycl streams
-            for (int is = 0; is < MAX_STREAMS; ++is) {
-                SYCL_CHECK(CHECK_TRY_ERROR(
-                    g_syclStreams[i][is] =
-                        dpct::get_current_device().create_queue(
-                            g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device())));
-            }
-
-            const dpct::queue_ptr stream = g_syclStreams[i][0];
-            // create sycl handle
-            SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream));
-        }
-
+        GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
+        ggml_backend_sycl_print_sycl_devices();
         initialized = true;
         g_sycl_loaded = true;
     }
@@ -13143,6 +12960,63 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
+void ggml_init_by_gpus(int device_count) try {
+    g_device_count = device_count;
+    g_work_group_size = g_sycl_gpu_mgr->work_group_size;
+
+    int64_t total_vram = 0;
+
+    print_gpu_device_list();
+
+    for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) {
+        g_device_caps[id].vmm = 0;
+        g_device_caps[id].device_id = -1;
+        g_device_caps[id].cc = 0;
+        g_tensor_split[id] = 0;
+        g_default_tensor_split[id] = 0;
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        int device_id = g_sycl_gpu_mgr->gpus[i];
+        g_device_caps[i].vmm = 0;
+
+        dpct::device_info prop;
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
+            prop, dpct::dev_mgr::instance().get_device(device_id))));
+
+        g_default_tensor_split[i] = total_vram;
+        total_vram += prop.get_global_mem_size();
+
+        g_device_caps[i].cc =
+            100 * prop.get_major_version() + 10 * prop.get_minor_version();
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        g_default_tensor_split[i] /= total_vram;
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        SYCL_CHECK(ggml_sycl_set_device(i));
+
+        // create sycl streams
+        for (int is = 0; is < MAX_STREAMS; ++is) {
+            SYCL_CHECK(CHECK_TRY_ERROR(
+                g_syclStreams[i][is] =
+                    dpct::get_current_device().create_queue(
+                        g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device())));
+        }
+
+        const dpct::queue_ptr stream = g_syclStreams[i][0];
+        // create sycl handle
+        SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream));
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
 void *ggml_sycl_host_malloc(size_t size) try {
     if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
         return nullptr;
@@ -16542,22 +16416,25 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
     /* .is_host          = */ nullptr,
 };
 
-ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
+ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) {
+    ggml_init_sycl();
+    if (device_index>=g_device_count or device_index<0) {
+        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+            device_index, g_device_count-1);
+        GGML_ASSERT(device_index<g_device_count);
+    }
     static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
 
-    static bool ggml_backend_sycl_buffer_type_initialized = false;
-
-    if (!ggml_backend_sycl_buffer_type_initialized) {
+    if (!g_ggml_backend_sycl_buffer_type_initialized) {
         for (int i = 0; i < g_device_count; i++) {
             ggml_backend_sycl_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
                 /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(g_sycl_gpu_mgr->gpus[i])},
             };
         }
-        ggml_backend_sycl_buffer_type_initialized = true;
+        g_ggml_backend_sycl_buffer_type_initialized = true;
     }
-
-    return &ggml_backend_sycl_buffer_types[device];
+    return &ggml_backend_sycl_buffer_types[device_index];
 }
 
 // sycl split buffer type
@@ -16910,6 +16787,7 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface
 };
 
 GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
+    ggml_init_sycl();
     // FIXME: this is not thread safe
     static std::map<std::array<float, GGML_SYCL_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
 
@@ -17243,6 +17121,13 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
     UNUSED(backend);
 }
 
+GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+    const int min_batch_size = 32;
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
+    GGML_UNUSED(backend);
+}
+
+
 static ggml_backend_i ggml_backend_sycl_interface = {
     /* .get_name                = */ ggml_backend_sycl_name,
     /* .free                    = */ ggml_backend_sycl_free,
@@ -17256,6 +17141,7 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
     /* .supports_op             = */ ggml_backend_sycl_supports_op,
+    /* .offload_op              = */ ggml_backend_sycl_offload_op,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
@@ -17269,7 +17155,7 @@ static ggml_guid_t ggml_backend_sycl_guid() {
 }
 
 GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) {
-    ggml_init_sycl(); // TODO: remove from ggml.c
+    ggml_init_sycl();
 
     check_allow_gpu_index(device);
 
@@ -17310,11 +17196,42 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) {
     return g_sycl_gpu_mgr->get_index(device_id);
 }
 
+GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index) {
+    return g_sycl_gpu_mgr->gpus[device_index];
+}
+
+GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id) {
+    GGML_ASSERT(main_gpu_id<g_all_sycl_device_count);
+    fprintf(stderr, "ggml_backend_sycl_set_single_device: use single device: [%d]\n", main_gpu_id);
+    if (g_sycl_gpu_mgr) {
+        delete g_sycl_gpu_mgr;
+    }
+    g_sycl_gpu_mgr = new sycl_gpu_mgr(main_gpu_id);
+    g_ggml_sycl_backend_gpu_mode = SYCL_SINGLE_GPU_MODE;
+    ggml_init_by_gpus(g_sycl_gpu_mgr->get_gpu_count());
+    g_ggml_backend_sycl_buffer_type_initialized = false;
+}
+
+GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode() {
+    if (g_ggml_sycl_backend_gpu_mode == SYCL_MUL_GPU_MODE) {
+        return;
+    }
+
+    fprintf(stderr, "ggml_backend_sycl_set_mul_device_mode: true\n");
+
+    if (g_sycl_gpu_mgr) {
+        delete g_sycl_gpu_mgr;
+    }
+    g_sycl_gpu_mgr = new sycl_gpu_mgr();
+    g_ggml_sycl_backend_gpu_mode = SYCL_MUL_GPU_MODE;
+    ggml_init_by_gpus(g_sycl_gpu_mgr->get_gpu_count());
+    g_ggml_backend_sycl_buffer_type_initialized = false;
+}
+
 extern "C" int ggml_backend_sycl_reg_devices();
 
 int ggml_backend_sycl_reg_devices() {
-    if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
-    g_device_count = g_sycl_gpu_mgr->get_gpu_count();
+    ggml_backend_sycl_set_mul_device_mode();
     assert(g_device_count>0);
     for (int i = 0; i < g_device_count; i++) {
         int id = g_sycl_gpu_mgr->gpus[i];

ggml-sycl.h

@@ -13,22 +13,37 @@
 extern "C" {
 #endif
 
-#define GGML_SYCL_MAX_DEVICES       16
+#define GGML_SYCL_MAX_DEVICES       48
 #define GGML_SYCL_NAME "SYCL"
 
-GGML_API void   ggml_init_sycl(void);
-GGML_API bool   ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+// backend API
 GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+// devide buffer
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+
 GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
 GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
 GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
 GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
 
+// TODO: these are temporary
+//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
+GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
+GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
+GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
+
+// SYCL doesn't support registering host memory, keep here for reference
+// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
 #ifdef  __cplusplus
 }
 #endif

ggml-vulkan.cpp

@@ -710,6 +710,12 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope
         }
     }
 
+    // All commands that are allowed on a queue that supports transfer operations are also allowed on a queue that supports either graphics or compute operations.
+    // Thus, if the capabilities of a queue family include VK_QUEUE_GRAPHICS_BIT or VK_QUEUE_COMPUTE_BIT, then reporting the VK_QUEUE_TRANSFER_BIT capability separately for that queue family is optional.
+    if (compute_index >= 0) {
+        return compute_index;
+    }
+
     std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;
 
     for(auto &q_family : queue_family_props) {
@@ -5693,6 +5699,7 @@ static ggml_backend_i ggml_backend_vk_interface = {
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_vk_graph_compute,
     /* .supports_op             = */ ggml_backend_vk_supports_op,
+    /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,

ggml.c

@@ -3,6 +3,7 @@
 
 #include "ggml-impl.h"
 #include "ggml-quants.h"
+#include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -43,6 +44,10 @@
 
 #if defined(_WIN32)
 
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
 #include <windows.h>
 
 typedef volatile LONG atomic_int;
@@ -282,14 +287,10 @@ inline static void * ggml_calloc(size_t num, size_t size) {
 #else
 #include <cblas.h>
 #endif
-#elif defined(GGML_USE_CUBLAS)
-#include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
 #elif defined(GGML_USE_VULKAN)
 #include "ggml-vulkan.h"
-#elif defined(GGML_USE_SYCL)
-#include "ggml-sycl.h"
 #endif
 
 // floating point type used to accumulate sums
@@ -432,6 +433,57 @@ int64_t ggml_cycles_per_ms(void) {
 #define ggml_perf_cycles_per_ms() 0
 #endif
 
+//
+// cross-platform UTF-8 file paths
+//
+
+#ifdef _WIN32
+static wchar_t * ggml_mbstowcs(const char * mbs) {
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, NULL, 0);
+    if (!wlen) {
+        errno = EINVAL;
+        return NULL;
+    }
+
+    wchar_t * wbuf = GGML_MALLOC(wlen * sizeof(wchar_t));
+    wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, wbuf, wlen);
+    if (!wlen) {
+        GGML_FREE(wbuf);
+        errno = EINVAL;
+        return NULL;
+    }
+
+    return wbuf;
+}
+#endif
+
+FILE * ggml_fopen(const char * fname, const char * mode) {
+#ifdef _WIN32
+    FILE * file = NULL;
+
+    // convert fname (UTF-8)
+    wchar_t * wfname = ggml_mbstowcs(fname);
+    if (wfname) {
+        // convert mode (ANSI)
+        wchar_t * wmode = GGML_MALLOC(strlen(mode) + 1);
+        wchar_t * wmode_p = wmode;
+        do {
+            *wmode_p++ = (wchar_t)*mode;
+        } while (*mode++);
+
+        // open file
+        file = _wfopen(wfname, wmode);
+
+        GGML_FREE(wfname);
+        GGML_FREE(wmode);
+    }
+
+    return file;
+#else
+    return fopen(fname, mode);
+#endif
+}
+
 //
 // cache line
 //
@@ -470,6 +522,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(int32_t),
         .is_quantized             = false,
     },
+    [GGML_TYPE_I64] = {
+        .type_name                = "i64",
+        .blck_size                = 1,
+        .type_size                = sizeof(int64_t),
+        .is_quantized             = false,
+    },
+    [GGML_TYPE_F64] = {
+        .type_name                = "f64",
+        .blck_size                = 1,
+        .type_size                = sizeof(double),
+        .is_quantized             = false,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_F32] = {
         .type_name                = "f32",
         .blck_size                = 1,
@@ -918,6 +983,101 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif
 
+#elif defined(__AVX512F__)
+
+#define GGML_SIMD
+
+// F32 AVX512
+
+#define GGML_F32_STEP 64
+#define GGML_F32_EPR  16
+
+#define GGML_F32x16         __m512
+#define GGML_F32x16_ZERO    _mm512_setzero_ps()
+#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
+#define GGML_F32x16_LOAD    _mm512_loadu_ps
+#define GGML_F32x16_STORE   _mm512_storeu_ps
+// _mm512_fmadd_ps is defined in AVX512F so no guard is required
+#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32x16_ADD     _mm512_add_ps
+#define GGML_F32x16_MUL     _mm512_mul_ps
+#define GGML_F32x16_REDUCE(res, x)                                    \
+do {                                                                  \
+    int offset = GGML_F32_ARR >> 1;                                   \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    res = _mm512_reduce_add_ps(x[0]);                                 \
+} while (0)
+
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC        GGML_F32x16
+#define GGML_F32_VEC_ZERO   GGML_F32x16_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x16_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x16_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x16_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x16_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x16_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x16_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
+
+// F16 AVX512
+
+// F16 AVX
+
+#define GGML_F16_STEP 64
+#define GGML_F16_EPR  16
+
+// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead
+
+#define GGML_F32Cx16             __m512
+#define GGML_F32Cx16_ZERO        _mm512_setzero_ps()
+#define GGML_F32Cx16_SET1(x)     _mm512_set1_ps(x)
+
+// unlike  _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
+// so F16C guard isn't required
+#define GGML_F32Cx16_LOAD(x)     _mm512_cvtph_ps(_mm256_loadu_si256((__m256i *)(x)))
+#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
+
+#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32Cx16_ADD         _mm512_add_ps
+#define GGML_F32Cx16_MUL         _mm512_mul_ps
+#define GGML_F32Cx16_REDUCE(res, x)                               \
+do {                                                              \
+    int offset = GGML_F32_ARR >> 1;                               \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    res = _mm512_reduce_add_ps(x[0]);                             \
+} while (0)
+
+#define GGML_F16_VEC                GGML_F32Cx16
+#define GGML_F16_VEC_ZERO           GGML_F32Cx16_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32Cx16_SET1
+#define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx16_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32Cx16_FMA
+#define GGML_F16_VEC_ADD            GGML_F32Cx16_ADD
+#define GGML_F16_VEC_MUL            GGML_F32Cx16_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
+
 #elif defined(__AVX__)
 
 #define GGML_SIMD
@@ -2532,14 +2692,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
-#if defined(GGML_USE_CUBLAS)
-        ggml_init_cublas();
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
         ggml_cl_init();
 #elif defined(GGML_USE_VULKAN)
         ggml_vk_init_cpu_assist();
-#elif defined(GGML_USE_SYCL)
-        ggml_init_sycl();
 #endif
 
         ggml_setup_op_has_task_pass();
@@ -10997,7 +11153,6 @@ static void ggml_compute_forward_out_prod_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
     // TODO: #if defined(GGML_USE_CLBLAST)
 
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
@@ -11197,7 +11352,6 @@ static void ggml_compute_forward_out_prod_q_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
     // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
 
     if (params->type == GGML_TASK_TYPE_INIT) {
@@ -12418,6 +12572,8 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+        case GGML_TYPE_F64:
         case GGML_TYPE_COUNT:
             {
                 GGML_ASSERT(false);
@@ -12504,6 +12660,8 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+        case GGML_TYPE_F64:
         case GGML_TYPE_COUNT:
             {
                 GGML_ASSERT(false);
@@ -15939,14 +16097,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
         return;
     }
 
-#ifdef GGML_USE_CUBLAS
-    bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
-    if (skip_cpu) {
-        return;
-    }
-    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU);
-    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
-#elif defined(GGML_USE_VULKAN)
+#if defined(GGML_USE_VULKAN)
     const bool skip_cpu = ggml_vk_compute_forward_cpu_assist(params, tensor);
 #ifdef GGML_VULKAN_CHECK_RESULTS
     if (skip_cpu) {
@@ -15958,14 +16109,8 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
     GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU);
     GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
-#endif // GGML_USE_CUBLAS
+#endif // GGML_USE_VULKAN
 
-#ifdef GGML_USE_SYCL
-    bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
-    if (skip_cpu) {
-        return;
-    }
-#endif // GGML_USE_SYCL
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -18640,7 +18785,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
     // write binary data
     {
-        FILE * fout = fopen(fname, "wb");
+        FILE * fout = ggml_fopen(fname, "wb");
 
         if (!fout) {
             fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
@@ -18778,7 +18923,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
     // read file into data
     {
-        FILE * fin = fopen(fname, "rb");
+        FILE * fin = ggml_fopen(fname, "rb");
         if (!fin) {
             fprintf(stderr, "%s: failed to open %s\n", __func__, fname);
             return result;
@@ -19114,7 +19259,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
 void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
     char color[16];
 
-    FILE * fp = fopen(filename, "w");
+    FILE * fp = ggml_fopen(filename, "w");
     GGML_ASSERT(fp);
 
     fprintf(fp, "digraph G {\n");
@@ -20432,7 +20577,7 @@ struct gguf_context * gguf_init_empty(void) {
 }
 
 struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) {
-    FILE * file = fopen(fname, "rb");
+    FILE * file = ggml_fopen(fname, "rb");
     if (!file) {
         return NULL;
     }
@@ -21387,7 +21532,7 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
 }
 
 void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
-    FILE * file = fopen(fname, "wb");
+    FILE * file = ggml_fopen(fname, "wb");
     if (!file) {
         GGML_ASSERT(false && "failed to open file for writing");
     }

ggml.h

@@ -214,9 +214,10 @@
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#include <stdint.h>
-#include <stddef.h>
 #include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
 
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
 #define GGML_FILE_VERSION 1
@@ -366,6 +367,8 @@ extern "C" {
         GGML_TYPE_I8      = 24,
         GGML_TYPE_I16     = 25,
         GGML_TYPE_I32     = 26,
+        GGML_TYPE_I64     = 27,
+        GGML_TYPE_F64     = 28,
         GGML_TYPE_COUNT,
     };
 
@@ -706,6 +709,9 @@ extern "C" {
 
     GGML_API void    ggml_print_backtrace(void);
 
+    // accepts a UTF-8 path, even on Windows
+    GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
+
     GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node