server : enable continuous batching by default (#6231)

* metal : proper assert for mat-mat memory alignment

ggml-ci

* readme : add notice about the bug fix

* metal : fix the fix

ggml-ci

.github/workflows/build.yml

@@ -21,6 +21,118 @@ env:
   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
 
@@ -48,6 +160,28 @@ 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 +210,40 @@ 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 $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v3
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+#
+#      - name: Test
+#        id: cmake_test
+#        run: |
+#          cd build
+#          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -333,6 +467,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -361,6 +496,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -724,6 +860,8 @@ 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: 1000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/server.yml

@@ -24,18 +24,15 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
+        build_type: [Debug]
         include:
           - build_type: Release
             sanitizer: ""
-        exclude:
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
-            sanitizer: THREAD
-          - build_type: Release
-            sanitizer: UNDEFINED
+            disabled_on_pr: true
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     container:
       image: ubuntu:latest
@@ -60,7 +57,8 @@ jobs:
             cmake \
             python3-pip \
             wget \
-            language-pack-en
+            language-pack-en \
+            libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
@@ -70,6 +68,7 @@ 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
@@ -81,13 +80,14 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' && matrix.build_type == 'Release' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
@@ -103,12 +103,21 @@ 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
@@ -122,15 +131,21 @@ 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 }}
         run: |
           cd examples/server/tests
           behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests
         id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
           behave.exe --stop --no-skipped --no-capture --tags slow

.github/workflows/tidy-post.yml

@@ -1,20 +0,0 @@
-name: clang-tidy review post comments
-
-on:
-  workflow_dispatch:
-    workflows: ["clang-tidy-review"]
-    types:
-      - completed
-
-jobs:
-  build:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: ZedThree/clang-tidy-review/post@v0.13.0
-        # lgtm_comment_body, max_comments, and annotations need to be set on the posting workflow in a split setup
-        with:
-          # adjust options as necessary
-          lgtm_comment_body: ''
-          annotations: false
-          max_comments: 25

.github/workflows/tidy-review.yml

@@ -1,23 +0,0 @@
-name: clang-tidy-review
-
-on:
-  pull_request:
-    branches:
-      - master
-
-jobs:
-  clang-tidy-review:
-    runs-on: ubuntu-latest
-
-    steps:
-    - uses: actions/checkout@v3
-
-    - uses: ZedThree/clang-tidy-review@v0.13.0
-      id: review
-      with:
-        lgtm_comment_body: ''
-        build_dir: build
-        cmake_command: cmake . -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=on
-        split_workflow: true
-
-    - uses: ZedThree/clang-tidy-review/upload@v0.13.0

.gitignore

@@ -11,7 +11,10 @@
 *.gcda
 *.dot
 *.bat
+*.tmp
 *.metallib
+*.etag
+*.lastModified
 .DS_Store
 .build/
 .cache/
@@ -25,6 +28,8 @@
 .vscode/
 .idea/
 
+ggml-metal-embed.metal
+
 lcov-report/
 gcovr-report/
 

CMakeLists.txt

@@ -99,6 +99,7 @@ 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_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)
@@ -118,6 +119,7 @@ option(LLAMA_SYCL                            "llama: use SYCL"
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
 option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
+set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
@@ -147,6 +149,8 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -197,9 +201,6 @@ if (LLAMA_METAL)
         add_compile_definitions(GGML_METAL_NDEBUG)
     endif()
 
-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
     # copy ggml-common.h and ggml-metal.metal to bin directory
     configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
@@ -208,53 +209,62 @@ if (LLAMA_METAL)
         enable_language(ASM)
         add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
+        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
         set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+
         file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+        # merge ggml-common.h and ggml-metal.metal into a single file
+        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
 
         add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
+            OUTPUT ${METALLIB_EMBED_ASM}
+            COMMAND echo "Embedding Metal library"
+            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
+            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+            DEPENDS ggml-metal.metal ggml-common.h
             COMMENT "Generate assembly for embedded Metal library"
         )
 
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
-    if (LLAMA_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
-        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        if (LLAMA_QKK_64)
-            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
+        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
+    else()
+        if (LLAMA_METAL_SHADER_DEBUG)
+            # custom command to do the following:
+            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+            #
+            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+            #       disabling fast math is needed in order to pass tests/test-backend-ops
+            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+            set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        else()
+            set(XC_FLAGS -O3)
         endif()
 
         add_custom_command(
             OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
             COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
             COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DEPENDS ggml-metal.metal
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+            DEPENDS ggml-metal.metal ggml-common.h
             COMMENT "Compiling Metal kernels"
-        )
+            )
 
         add_custom_target(
             ggml-metal ALL
             DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
-    endif()
+            )
+    endif() # LLAMA_METAL_EMBED_LIBRARY
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}

Makefile

@@ -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
@@ -167,6 +168,10 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
+ifdef LLAMA_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+endif
+
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
@@ -549,19 +554,20 @@ 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
-ggml-metal-embed.o: ggml-metal.metal
+ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 	@echo "Embedding Metal library"
+	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
 	@$(AS) $(TEMP_ASSEMBLY) -o $@
 	@rm -f ${TEMP_ASSEMBLY}
 endif
@@ -590,6 +596,11 @@ 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
 #
@@ -656,6 +667,9 @@ 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 $@
 
@@ -735,7 +749,7 @@ 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)
 
@@ -743,6 +757,10 @@ 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)
@@ -851,6 +869,10 @@ 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

@@ -10,12 +10,15 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
+- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
 - [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
 ### 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
@@ -110,6 +113,7 @@ 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:**
 
@@ -131,6 +135,7 @@ 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)
@@ -902,6 +907,9 @@ First, install the essential packages for termux:
 pkg install clang wget git cmake
 ```
 Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
+
+You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
+
 ```
 $ mkdir build-android
 $ cd build-android
@@ -910,7 +918,28 @@ $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROI
 $ make
 ```
 Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
-Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+```
+
+Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 

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)
 
@@ -60,6 +62,7 @@ add_library(${TARGET} STATIC
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    json.hpp
     train.h
     train.cpp
     )
@@ -68,6 +71,17 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${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
@@ -51,7 +54,8 @@ struct gpt_params {
     int32_t n_threads_batch_draft = -1;
     int32_t n_predict             = -1;    // new tokens to predict
     int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
     int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
@@ -85,6 +89,7 @@ struct gpt_params {
     struct llama_sampling_params sparams;
 
     std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_url         = ""; // model url to download
     std::string model_draft       = "";                              // draft model for speculative decoding
     std::string model_alias       = "unknown"; // model alias
     std::string prompt            = "";
@@ -102,6 +107,11 @@ 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)
@@ -129,7 +139,7 @@ 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
@@ -182,6 +192,9 @@ 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,
+                                                         struct llama_model_params     params);
+
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -267,3 +280,25 @@ 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);

common/json-schema-to-grammar.cpp

@@ -0,0 +1,725 @@
+#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::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(json(literal).dump(), 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, nlohmann::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(prop_name) + " 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(nlohmann::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) {
+        if (!value.is_string()) {
+            _errors.push_back("Only std::string constants are supported, got " + value.dump());
+            return "";
+        }
+        return format_literal(value.get<std::string>());
+    }
+
+    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::json& schema);

common/sampling.cpp

@@ -17,6 +17,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
             return nullptr;
         }
 
+        // Ensure that there is a "root" node.
+        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
+            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
+            delete result;
+            return nullptr;
+        }
+
         std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
 
         result->grammar = llama_grammar_init(

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,

convert-hf-to-gguf.py

@@ -1634,7 +1634,7 @@ 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 +1965,23 @@ 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-bench/batched-bench.cpp

@@ -138,6 +138,8 @@ int main(int argc, char ** argv) {
                 LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
+
+            llama_synchronize(ctx);
         }
 
         return true;

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/embedding/embedding.cpp

@@ -107,18 +107,25 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch == params.n_ctx);
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
     // 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++) {
@@ -167,15 +174,26 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print first 3 embeddings
-    for (int j = 0; j < std::min(3, n_prompts); j++) {
-        fprintf(stderr, "embedding %d: ", j);
-        for (int i = 0; i < n_embd; i++) {
-            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+    // print the first part of the embeddings
+    fprintf(stdout, "\n");
+    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, "%9.6f ", emb[j * n_embd + i]);
         }
-        fprintf(stderr, "\n\n");
+        fprintf(stdout, "\n");
+    }
+
+    // print cosine similarity matrix
+    fprintf(stdout, "\n");
+    printf("cosine similarity matrix:\n\n");
+    for (int i = 0; i < n_prompts; i++) {
+        for (int j = 0; j < n_prompts; j++) {
+            float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+            fprintf(stdout, "%6.2f ", sim);
+        }
+        fprintf(stdout, "\n");
     }
-    fprintf(stderr, "\n");
 
     // clean up
     llama_print_timings(ctx);

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 +1,489 @@
+#include "llama.h"
+#include "ggml.h"
+#include "common.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <fstream>
+#include <ios>
+#include <string>
+#include <vector>
+
+#include <stdio.h>
+#include <fcntl.h>
+#include <string.h>
+
+enum split_operation : uint8_t {
+    SPLIT_OP_SPLIT,
+    SPLIT_OP_MERGE,
+};
+
+static const char * const LLM_KV_GENERAL_SPLIT_I_SPLIT = "general.split";
+static const char * const LLM_KV_GENERAL_SPLIT_N_SPLIT = "general.split_count";
+
+static const int SPLIT_FILENAME_MAX = 256;
+
+static const char * const SPLIT_FILENAME_FORMAT = "%s-%05d-of-%05d.gguf";
+
+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(1);
+        }
+    }
+    catch (const std::invalid_argument & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        split_print_usage(argv[0]);
+        exit(1);
+    }
+    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);
+    }
+}
+
+static std::string split_file_name(const std::string & path, int i_split, int n_split) {
+    char f_split[SPLIT_FILENAME_MAX] = {0};
+    snprintf(f_split, sizeof(f_split), SPLIT_FILENAME_FORMAT, path.c_str(), i_split + 1, n_split);
+    return std::string(f_split);
+}
+
+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_u8(ctx_out, LLM_KV_GENERAL_SPLIT_I_SPLIT, i_split);
+        gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, n_split);
+
+        // 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);
+        }
+
+        auto split_name = split_file_name(params.output, i_split, n_split);
+
+        fprintf(stderr, "%s: %s ...", __func__, split_name.c_str());
+        fout = std::ofstream(split_name, 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(1);
+    }
+
+    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(1);
+    }
+
+    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
+            __func__, split_params.input.c_str(),
+            split_file_name(split_params.output, strategy.i_split, strategy.n_split).c_str(),
+            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;
+
+    std::string split_prefix;
+
+    // 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,
+        };
+
+        auto split_name = split_params.input;
+        if (i_split > 0) {
+            split_name = split_file_name(split_prefix, i_split, n_split);
+        }
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_name.c_str());
+
+        auto * ctx_gguf = gguf_init_from_file(split_name.c_str(), params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
+            exit(1);
+        }
+        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_GENERAL_SPLIT_N_SPLIT);
+            if (key_n_split < 0) {
+                fprintf(stderr,
+                        "\n%s: input file does not contain %s metadata\n",
+                        __func__,
+                        LLM_KV_GENERAL_SPLIT_N_SPLIT);
+                gguf_free(ctx_gguf);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            n_split = gguf_get_val_u8(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);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            // Do not trigger merge if we try to merge again the output
+            gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, 0);
+
+            // Set metadata from the first split
+            gguf_set_kv(ctx_out, ctx_gguf);
+        }
+
+        // Verify the file naming
+        {
+            int i_split_file = 0;
+            int n_split_file = 0;
+            const char * i_split_format = "-00000-of-00000.gguf";
+
+            if (split_name.size() < strlen(i_split_format)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s\n", __func__, split_params.input.c_str());
+                for (auto * _ctx_gguf : ctx_ggufs) {
+                    gguf_free(_ctx_gguf);
+                }
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+
+            split_prefix = split_name.substr(0, split_name.size() - strlen(i_split_format));
+
+            const char * split_name_c_str = split_name.c_str();
+            int n_part = sscanf(&split_name_c_str[0] + split_prefix.size(), "-%d-of-%d", &i_split_file, &n_split_file);
+
+            if (n_part != 2 || i_split_file - 1 != i_split || n_split_file != n_split) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d i_split_file=%d"
+                                " n_split=%d n_split_file=%d\n", __func__,
+                        split_params.input.c_str(),
+                        i_split, i_split_file,
+                        n_split, n_split_file);
+                for (auto * _ctx_gguf : ctx_ggufs) {
+                    gguf_free(_ctx_gguf);
+                }
+                gguf_free(ctx_out);
+                fout.close();
+                exit(1);
+            }
+        }
+
+        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++) {
+        auto split_name = split_file_name(split_prefix, i_split, n_split);
+        std::ifstream f_input(split_name.c_str(), std::ios::binary);
+        if (!f_input.is_open()) {
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_name.c_str());
+            for (auto * _ctx_gguf : ctx_ggufs) {
+                gguf_free(_ctx_gguf);
+            }
+            gguf_free(ctx_out);
+            fout.close();
+            exit(1);
+        }
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_name.c_str());
+
+        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(1);
+    }
+
+    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/gritlm/gritlm.cpp

@@ -6,22 +6,6 @@
 
 // #define GRIT_DEBUG
 
-static float dot_product(const std::vector<float> & v1, const std::vector<float> & v2) {
-    float dot = 0.0f;
-    for (uint64_t i = 0; i < v1.size(); ++i) {
-        dot += v1[i] * v2[i];
-    }
-    return dot;
-}
-
-static float norm(const std::vector<float> & v) {
-    return std::sqrt(dot_product(v, v));
-}
-
-static float cosine_similarity(const std::vector<float> & v1, const std::vector<float> & v2) {
-    return dot_product(v1, v2) / (norm(v1) * norm(v2));
-}
-
 static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
     std::vector<std::vector<float>> result;
 
@@ -203,10 +187,12 @@ int main(int argc, char * argv[]) {
         const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
         const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
 
-        const float cosine_sim_q0_d0 = cosine_similarity(q_rep[0], d_rep[0]);
-        const float cosine_sim_q0_d1 = cosine_similarity(q_rep[0], d_rep[1]);
-        const float cosine_sim_q1_d0 = cosine_similarity(q_rep[1], d_rep[0]);
-        const float cosine_sim_q1_d1 = cosine_similarity(q_rep[1], d_rep[1]);
+        const int n_embd = llama_n_embd(mdl);
+
+        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
 
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);

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?
@@ -12,22 +14,50 @@ 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(
@@ -41,78 +71,421 @@ 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):
+        assert isinstance(value, str), f'Only string constants are supported, got {value}'
+        return self._format_literal(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(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 +502,47 @@ 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 += "/";
@@ -164,6 +166,7 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
+    std::vector<int> n_ubatch;
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
@@ -183,7 +186,8 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_batch       */ {512},
+    /* n_batch       */ {2048},
+    /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
@@ -208,6 +212,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
     printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
@@ -217,7 +222,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
     printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
-    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
@@ -244,6 +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;
 }
@@ -297,6 +305,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
         } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -455,6 +470,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
+    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
     if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
     if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
@@ -474,6 +490,7 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
+    int n_ubatch;
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
@@ -511,6 +528,7 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
+        cparams.n_ubatch = n_ubatch;
         cparams.type_k = type_k;
         cparams.type_v = type_v;
         cparams.offload_kqv = !no_kv_offload;
@@ -532,6 +550,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
+    for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
     for (const auto & nkvo : params.no_kv_offload)
@@ -545,6 +564,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -568,6 +588,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -604,6 +625,7 @@ struct test {
     uint64_t model_size;
     uint64_t model_n_params;
     int n_batch;
+    int n_ubatch;
     int n_threads;
     ggml_type type_k;
     ggml_type type_v;
@@ -627,6 +649,7 @@ struct test {
         model_size = llama_model_size(lmodel);
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
+        n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -705,7 +728,8 @@ struct test {
             "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "type_k", "type_v",
+            "n_batch", "n_ubatch",
+            "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
             "tensor_split", "use_mmap", "embeddings",
@@ -719,7 +743,8 @@ struct test {
     enum field_type {STRING, BOOL, INT, FLOAT};
 
     static field_type get_field_type(const std::string & field) {
-        if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
+            field == "n_threads" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -759,7 +784,8 @@ struct test {
             std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_batch), std::to_string(n_ubatch),
+            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload),
             tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
@@ -957,6 +983,9 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
+        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
+            fields.emplace_back("n_ubatch");
+        }
         if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
             fields.emplace_back("type_k");
         }
@@ -1096,25 +1125,40 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
-    int n_processed = 0;
-
     llama_set_n_threads(ctx, n_threads, n_threads);
 
+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token> tokens(n_batch);
+
+    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;
     }
+
+    llama_synchronize(ctx);
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(llama_get_model(ctx));
-
     llama_set_n_threads(ctx, n_threads, n_threads);
 
+    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;
     }
 }
 
@@ -1203,7 +1247,8 @@ int main(int argc, char ** argv) {
 
         // warmup run
         if (t.n_prompt > 0) {
-            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
+            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             test_gen(ctx, 1, 0, t.n_threads);
@@ -1219,6 +1264,7 @@ int main(int argc, char ** argv) {
             if (t.n_gen > 0) {
                 test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
+
             uint64_t t_ns = get_time_ns() - t_start;
             t.samples_ns.push_back(t_ns);
         }

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

@@ -221,6 +221,7 @@ actor LlamaContext {
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
+            llama_synchronize(context)
 
             let t_pp_end = ggml_time_us()
 
@@ -240,6 +241,7 @@ actor LlamaContext {
                 if llama_decode(context, batch) != 0 {
                     print("llama_decode() failed during text generation")
                 }
+                llama_synchronize(context)
             }
 
             let t_tg_end = ggml_time_us()

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/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/perplexity/perplexity.cpp

@@ -589,9 +589,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             }
         }
 
-        const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
+            llama_synchronize(ctx);
+            const auto t_end = std::chrono::high_resolution_clock::now();
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
             fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
             int total_seconds = (int)(t_total*n_chunk/n_seq);

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

@@ -20,6 +20,7 @@ The project is under active development, and we are [looking for feedback and co
 - `-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-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 (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
 - `-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.
@@ -259,7 +260,7 @@ 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

@@ -1,25 +1,50 @@
+// 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) {
@@ -30,83 +55,490 @@ export class SchemaConverter {
     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) {
+    if (typeof value !== 'string') {
+      throw new Error('Only string constants are supported, got ' + JSON.stringify(value));
+    }
+    return this._formatLiteral(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));
+      if (typeof schema.const !== 'string') {
+        throw new Error('Only string constants are supported, got ' + JSON.stringify(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(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"
 
@@ -147,7 +148,7 @@ struct server_slot {
     int32_t n_decoded   = 0;
     int32_t n_remaining = -1;
     int32_t i_batch     = -1;
-    int32_t n_predict   = -1;
+    int32_t n_predict   = -1; // TODO: disambiguate from params.n_predict
 
     int32_t n_prompt_tokens           = 0;
     int32_t n_prompt_tokens_processed = 0;
@@ -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
@@ -739,7 +741,13 @@ struct server_context {
         default_generation_settings_for_props = get_formated_generation(slots.front());
         default_generation_settings_for_props["seed"] = -1;
 
-        batch = llama_batch_init(n_ctx, 0, params.n_parallel);
+        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
+        {
+            const int32_t n_batch = llama_n_batch(ctx);
+
+            batch = llama_batch_init(n_batch, 0, params.n_parallel);
+        }
 
         metrics.init();
     }
@@ -839,7 +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);
 
@@ -1036,8 +1054,10 @@ struct server_context {
                 llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
             }
 
-            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
+            const int32_t n_batch = llama_n_batch(ctx);
+
+            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
+                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
                 llama_batch batch_view = {
                     n_tokens,
                     batch.token    + i,
@@ -1226,7 +1246,7 @@ struct server_context {
             {"mirostat_eta",              slot.sparams.mirostat_eta},
             {"penalize_nl",               slot.sparams.penalize_nl},
             {"stop",                      slot.params.antiprompt},
-            {"n_predict",                 slot.params.n_predict},
+            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
             {"n_keep",                    params.n_keep},
             {"ignore_eos",                ignore_eos},
             {"stream",                    slot.params.stream},
@@ -1738,7 +1758,8 @@ struct server_context {
         }
 
         // process in chunks of params.n_batch
-        int32_t n_batch = params.n_batch;
+        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
         if (params.cont_batching || batch.n_tokens == 0) {
@@ -1811,7 +1832,7 @@ struct server_context {
 
                         if (slot.embedding) {
                             // this prompt is too large to process - discard it
-                            if (slot.n_prompt_tokens > n_batch) {
+                            if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.state = SLOT_STATE_PROCESSING;
                                 slot.command = SLOT_COMMAND_NONE;
                                 slot.release();
@@ -2157,7 +2178,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
     printf("  --pooling {none,mean,cls} pooling type for embeddings, use model default if unspecified\n");
     printf("  -dt N, --defrag-thold N\n");
     printf("                            KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
-    printf("  -b N, --batch-size N      batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  -b N, --batch-size N      logical maximum batch size (default: %d)\n", params.n_batch);
+    printf("  -ub N, --ubatch-size N    physical maximum batch size (default: %d)\n", params.n_ubatch);
     printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
     printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_supports_mlock()) {
@@ -2185,6 +2207,8 @@ 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: %s)\n", params.model_url.c_str());
     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");
@@ -2201,7 +2225,7 @@ 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");
@@ -2307,6 +2331,12 @@ 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 == "-a" || arg == "--alias") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -2424,6 +2454,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 break;
             }
             params.n_batch = std::stoi(argv[i]);
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ubatch = std::stoi(argv[i]);
         } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -2763,6 +2799,7 @@ int main(int argc, char ** argv) {
         res.set_header("Access-Control-Allow-Credentials", "true");
         res.set_header("Access-Control-Allow-Methods",     "POST");
         res.set_header("Access-Control-Allow-Headers",     "*");
+        return res.set_content("", "application/json; charset=utf-8");
     });
 
     svr->set_logger(log_server_request);
@@ -3371,44 +3408,37 @@ int main(int argc, char ** argv) {
         const json body = json::parse(req.body);
         bool is_openai = false;
 
-        // an input prompt can string or a list of tokens (integer)
-        std::vector<json> prompts;
+        // an input prompt can be a string or a list of tokens (integer)
+        json prompt;
         if (body.count("input") != 0) {
             is_openai = true;
-            if (body["input"].is_array()) {
-                // support multiple prompts
-                for (const json & elem : body["input"]) {
-                    prompts.push_back(elem);
-                }
-            } else {
-                // single input prompt
-                prompts.push_back(body["input"]);
-            }
+            prompt = body["input"];
         } else if (body.count("content") != 0) {
-            // only support single prompt here
-            std::string content = body["content"];
-            prompts.push_back(content);
+            // with "content", we only support single prompt
+            prompt = std::vector<std::string>{body["content"]};
         } else {
             res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
-        // process all prompts
-        json responses = json::array();
-        for (auto & prompt : prompts) {
-            // TODO @ngxson : maybe support multitask for this endpoint?
-            // create and queue the task
+        // create and queue the task
+        json responses;
+        {
             const int id_task = ctx_server.queue_tasks.get_new_id();
-
             ctx_server.queue_results.add_waiting_task_id(id_task);
-            ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0}}, false, true);
+            ctx_server.request_completion(id_task, -1, {{"prompt", prompt}}, false, true);
 
             // get the result
             server_task_result result = ctx_server.queue_results.recv(id_task);
             ctx_server.queue_results.remove_waiting_task_id(id_task);
             if (!result.error) {
-                // append to the responses
-                responses.push_back(result.data);
+                if (result.data.count("results")) {
+                    // result for multi-task
+                    responses = result.data["results"];
+                } else {
+                    // result for single task
+                    responses = std::vector<json>{result.data};
+                }
             } else {
                 // error received, ignore everything else
                 res_error(res, result.data);
@@ -3417,24 +3447,19 @@ int main(int argc, char ** argv) {
         }
 
         // write JSON response
-        json root;
-        if (is_openai) {
-            json res_oai = json::array();
-            int i = 0;
-            for (auto & elem : responses) {
-                res_oai.push_back(json{
-                    {"embedding", json_value(elem, "embedding", json::array())},
-                    {"index",     i++},
-                    {"object",    "embedding"}
-                });
-            }
-            root = format_embeddings_response_oaicompat(body, res_oai);
-        } else {
-            root = responses[0];
-        }
+        json root = is_openai
+            ? format_embeddings_response_oaicompat(body, responses)
+            : responses[0];
         return res.set_content(root.dump(), "application/json; charset=utf-8");
     };
 
+    auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
+        return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
+            res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
+            return false;
+        };
+    };
+
     //
     // Router
     //
@@ -3446,17 +3471,6 @@ int main(int argc, char ** argv) {
     }
 
     // using embedded static files
-    auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
-        return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
-            res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
-            return false;
-        };
-    };
-
-    svr->Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) {
-        // TODO @ngxson : I have no idea what it is... maybe this is redundant?
-        return res.set_content("", "application/json; charset=utf-8");
-    });
     svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
     svr->Get("/index.js", handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
     svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));

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

@@ -4,11 +4,13 @@ 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
     And   1024 as batch size
+    And   1024 as ubatch size
     And   2048 KV cache size
     And   embeddings extraction
     Then  the server is starting

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/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

@@ -4,7 +4,8 @@ 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 url https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf
+    And   a model file stories260K.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
@@ -22,21 +24,27 @@ 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_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
@@ -51,6 +59,7 @@ 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 = []
@@ -61,7 +70,17 @@ def step_server_config(context, server_fqdn, server_port):
 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")
+        print(f"model file: {context.model_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}')
@@ -118,14 +137,21 @@ def step_server_metrics(context):
 def step_start_server(context):
     start_server_background(context)
     attempts = 0
+    max_attempts = 20
+    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
             attempts += 1
-            if attempts > 20:
+            if attempts > max_attempts:
                 assert False, "server not started"
             print(f"waiting for server to start, connect error code = {result}...")
             time.sleep(0.1)
@@ -136,7 +162,8 @@ 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',
@@ -189,7 +216,7 @@ 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}"
 
@@ -243,6 +270,11 @@ 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'
@@ -278,6 +310,11 @@ def step_n_batch(context, n_batch):
     context.n_batch = n_batch
 
 
+@step('{n_ubatch:d} as ubatch size')
+def step_n_ubatch(context, n_ubatch):
+    context.n_ubatch = n_ubatch
+
+
 @step('{seed:d} as seed')
 def step_seed(context, seed):
     context.seed = seed
@@ -329,7 +366,7 @@ 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)
 
@@ -338,7 +375,7 @@ 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,
@@ -353,6 +390,9 @@ 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
@@ -412,6 +452,8 @@ 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)
@@ -432,6 +474,8 @@ 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
@@ -483,12 +527,12 @@ 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
 
 
@@ -605,7 +649,7 @@ 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:
@@ -714,6 +758,7 @@ 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):
@@ -739,6 +784,8 @@ 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': {
@@ -799,6 +846,7 @@ 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:
@@ -907,7 +955,7 @@ 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:'
@@ -917,7 +965,7 @@ 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)
@@ -934,9 +982,12 @@ 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:
+        timeout *= 2
+
     async with aiohttp.ClientSession() as session:
         while True:
             async with await session.get(f'{base_url}/health', params=params) as health_response:
@@ -1020,15 +1071,18 @@ 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.n_batch:
         server_args.extend(['--batch-size', context.n_batch])
+    if context.n_ubatch:
+        server_args.extend(['--ubatch-size', context.n_ubatch])
     if context.n_gpu_layer:
         server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
     if context.server_continuous_batching:
@@ -1055,7 +1109,7 @@ 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
@@ -1064,8 +1118,23 @@ 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

@@ -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>()});
@@ -529,6 +541,16 @@ static std::vector<json> format_partial_response_oaicompat(json result, const st
 }
 
 static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
+    json data = json::array();
+    int i = 0;
+    for (auto & elem : embeddings) {
+        data.push_back(json{
+            {"embedding", json_value(elem, "embedding", json::array())},
+            {"index",     i++},
+            {"object",    "embedding"}
+        });
+    }
+
     json res = json {
         {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
         {"object", "list"},
@@ -536,7 +558,7 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
             {"prompt_tokens", 0},
             {"total_tokens", 0}
         }},
-        {"data", embeddings}
+        {"data", data}
     };
 
     return res;

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

@@ -61,7 +61,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
-// TODO: GGML_PAD ?
 static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
     assert(alignment && !(alignment & (alignment - 1))); // power of 2
     size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
@@ -69,25 +68,14 @@ static size_t aligned_offset(const void * buffer, size_t offset, size_t alignmen
 }
 
 // tallocr
-struct ggml_tallocr {
-    ggml_backend_buffer_t buffer;
-    void * base;
-    size_t alignment;
-    size_t offset;
-};
-
-ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
-    ggml_tallocr_t talloc = malloc(sizeof(struct ggml_tallocr));
-    if (talloc == NULL) {
-        return NULL;
-    }
 
+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
     void * base = ggml_backend_buffer_get_base(buffer);
     size_t align = ggml_backend_buffer_get_alignment(buffer);
 
     assert(align && !(align & (align - 1))); // power of 2
 
-    *talloc = (struct ggml_tallocr) {
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
         /*.buffer    = */ buffer,
         /*.base      = */ base,
         /*.alignment = */ align,
@@ -96,11 +84,7 @@ ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
     return talloc;
 }
 
-void ggml_tallocr_free(ggml_tallocr_t talloc) {
-    free(talloc);
-}
-
-void ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor) {
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
     size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
     size = GGML_PAD(size, talloc->alignment);
 
@@ -354,12 +338,16 @@ struct hash_node {
     bool allocated;
 };
 
-//
 struct tensor_alloc {
     size_t offset;
     size_t size_max; // 0 = pre-allocated, unused, or view
 };
 
+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
 struct node_alloc {
     int buffer_id;
     struct tensor_alloc dst;
@@ -378,7 +366,7 @@ struct ggml_gallocr {
     struct node_alloc * node_allocs; // [n_nodes]
     int n_nodes;
 
-    struct tensor_alloc * leaf_allocs; // [n_leafs]
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
     int n_leafs;
 };
 
@@ -543,17 +531,28 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
     return node_buffer_ids ? node_buffer_ids[i] : 0;
 }
 
-static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     // clear hash tables
     memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
     memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
 
+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
     // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
+    // allocate other graph inputs and leafs first to avoid overwriting them
     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;
         }
@@ -570,26 +569,13 @@ 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));
             }
         }
     }
 
-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
-
     // allocate tensors
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
@@ -652,7 +638,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     }
 }
 
-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     size_t hash_size = graph->visited_hash_table.size;
 
     // initialize hash table
@@ -676,7 +662,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
 
     // allocate in hash table
-    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids);
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
 
     // set the node_allocs from the hash table
     if (galloc->n_nodes < graph->n_nodes) {
@@ -711,15 +697,16 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
     if (galloc->n_leafs < graph->n_leafs) {
         free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
+        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
         GGML_ASSERT(galloc->leaf_allocs != NULL);
     }
     galloc->n_leafs = graph->n_leafs;
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        galloc->leaf_allocs[i].leaf.offset = hn->offset;
+        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
     }
 
     // reallocate buffers if needed
@@ -727,7 +714,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
         size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
 
-        if (new_size > cur_size) {
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
             fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
@@ -744,30 +732,30 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 }
 
 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-    return ggml_gallocr_reserve_n(galloc, graph, NULL);
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
 
-    if (node->view_src != NULL) {
-        if (node->buffer == NULL) {
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
             assert(tensor_alloc->offset == SIZE_MAX);
-            if (node->view_src->buffer == NULL) {
+            if (tensor->view_src->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
         }
     } else {
-        if (node->data == NULL) {
+        if (tensor->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
             void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
         } else {
-            if (node->buffer == NULL) {
+            if (tensor->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
@@ -843,13 +831,18 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
 
     // reset buffers
     for (int i = 0; i < galloc->n_buffers; i++) {
-        // zero size buffers are not allocated
         if (galloc->buffers[i] != NULL) {
             ggml_backend_buffer_reset(galloc->buffers[i]);
         }
     }
 
     // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
+    }
     // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
@@ -863,12 +856,6 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
         }
         ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
     }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
-    }
 
     return true;
 }
@@ -900,12 +887,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
         return false;
     }
 
-    struct ggml_tallocr * tallocr = ggml_tallocr_new(buffer);
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
 
     for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
         if (t->data == NULL) {
             if (t->view_src == NULL) {
-                ggml_tallocr_alloc(tallocr, t);
+                ggml_tallocr_alloc(&tallocr, t);
             } else if (t->buffer == NULL) {
                 ggml_backend_view_init(buffer, t);
             }
@@ -917,8 +904,6 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
         }
     }
 
-    ggml_tallocr_free(tallocr);
-
     *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
     (*buffers)[(*n_buffers)++] = buffer;
 

ggml-alloc.h

@@ -11,11 +11,15 @@ typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
 typedef struct ggml_backend * ggml_backend_t;
 
 // Tensor allocator
-typedef struct ggml_tallocr * ggml_tallocr_t;
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
 
-GGML_API ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer);
-GGML_API void           ggml_tallocr_free(ggml_tallocr_t talloc);
-GGML_API void           ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor);
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
 
 // Graph allocator
 /*
@@ -50,7 +54,11 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-GGML_API bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
 
 // automatic reallocation if the topology changes when using a single buffer
 // returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)

ggml-backend-impl.h

@@ -86,12 +86,12 @@ extern "C" {
         // (optional) asynchronous tensor data access
         void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
         // (optional) complete all pending operations
         void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        // create a plan for ggml_cgraph and free it
+        // compute graph with a plan (not used currently)
         ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
@@ -102,16 +102,32 @@ 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);
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
     };
 
     struct ggml_backend {
         ggml_guid_t guid;
 
         struct ggml_backend_i iface;
-
         ggml_backend_context_t context;
     };
 
+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
+
     //
     // Backend registry
     //

ggml-backend.h

@@ -9,6 +9,7 @@ extern "C" {
 
     typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
     typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
     typedef struct ggml_backend * ggml_backend_t;
     typedef void * ggml_backend_graph_plan_t;
 
@@ -69,14 +70,27 @@ 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 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_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
 
     //
     // CPU backend
@@ -123,27 +137,31 @@ extern "C" {
     /*
       Example usage:
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
-        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be asigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
 
-        // in build_graph:
-        build_graph(...) {
-            // manually assign nodes to a backend (optional, should not be needed in most cases)
-            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
-        }
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);
 
-        // allocate backend buffers from measure graph
-        ggml_backend_sched_init_measure(sched, measure_graph);
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
 
-        // the scheduler is now ready to compute graphs
+        ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
         graph = build_graph(sched);
         ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
     */
 
     struct ggml_backend_sched;
@@ -158,20 +176,26 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+
     // Initialize backend buffers from a measure graph
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
     // Get the number of splits of the last graph
     GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
 
     GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
-    GGML_API void                 ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
     // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
     // Reset all assignments and allocators - must be called before changing the node backends
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);

ggml-common.h

@@ -1,4 +1,408 @@
-#pragma once
+#ifndef GGML_COMMON_DECL
+
+#if defined(GGML_COMMON_DECL_C)
+#include <stdint.h>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_METAL)
+#include <metal_stdlib>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CUDA)
+#include <cuda_fp16.h>
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_HIP)
+#include <hip/hip_fp16.h>
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_SYCL)
+#include <sycl/half_type.hpp>
+#include <cstdint>
+
+typedef sycl::half  ggml_half;
+typedef sycl::half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#endif
+
+#if defined(GGML_COMMON_DECL)
+
+#ifndef __cplusplus
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+#endif // __cplusplus
+
+// QK = number of values after dequantization
+// QK_K = super-block size
+
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif // GGML_QKK_64
+
+#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
+// QR = QK / number of values before dequantization
+// QI = number of 32 bit integers before dequantization
+
+#define QI4_0 (QK4_0 / (4 * QR4_0))
+#define QR4_0 2
+
+#define QI4_1 (QK4_1 / (4 * QR4_1))
+#define QR4_1 2
+
+#define QI5_0 (QK5_0 / (4 * QR5_0))
+#define QR5_0 2
+
+#define QI5_1 (QK5_1 / (4 * QR5_1))
+#define QR5_1 2
+
+#define QI8_0 (QK8_0 / (4 * QR8_0))
+#define QR8_0 1
+
+#define QI8_1 (QK8_1 / (4 * QR8_1))
+#define QR8_1 1
+
+#define QI2_K (QK_K / (4*QR2_K))
+#define QR2_K 4
+
+#define QI3_K (QK_K / (4*QR3_K))
+#define QR3_K 4
+
+#define QI4_K (QK_K / (4*QR4_K))
+#define QR4_K 2
+
+#define QI5_K (QK_K / (4*QR5_K))
+#define QR5_K 2
+
+#define QI6_K (QK_K / (4*QR6_K))
+#define QR6_K 2
+
+#define QI2_XXS (QK_K / (4*QR2_XXS))
+#define QR2_XXS 8
+
+#define QI2_XS (QK_K / (4*QR2_XS))
+#define QR2_XS 8
+
+#define QI2_S (QK_K / (4*QR2_S))
+#define QR2_S 8
+
+#define QI3_XXS (QK_K / (4*QR3_XXS))
+#define QR3_XXS 8
+
+#define QI3_XS (QK_K / (4*QR3_XS))
+#define QR3_XS 8
+
+#define QI1_S (QK_K / (4*QR1_S))
+#define QR1_S 8
+
+#define QI4_NL (QK4_NL / (4*QR4_NL))
+#define QR4_NL 2
+
+#if QK_K == 64
+#define QI4_XS QI4_NL
+#define QR4_XS QR4_NL
+#else
+#define QI4_XS (QK_K / (4*QR4_XS))
+#define QR4_XS 8
+#endif
+
+#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
+
+#define QK4_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qs[QK4_0 / 2]; // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qs[QK4_1 / 2]; // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_half d;       // delta
+    int8_t  qs[QK8_0]; // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half s; // d * sum(qs[i])
+        } GGML_COMMON_AGGR;
+        ggml_half2 ds;
+    };
+    int8_t qs[QK8_1]; // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8]; // quants - high bit
+    uint8_t qs[QK_K/4];    // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_half d;           // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8]; // quants - high bit
+    uint8_t qs[QK_K/4];    // quants - low 2 bits
+    uint8_t scales[12];    // scales, quantized with 6 bits
+    ggml_half d;           // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_half d[2];     // super-block scales/mins
+    uint8_t scales[2];  // 4-bit block scales/mins
+    uint8_t qs[QK_K/2]; // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];           // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_half d;             // super-block scale
+    int8_t  scales[QK_K/16]; // 8-bit block scales
+    uint8_t qh[QK_K/8];      // quants, high bit
+    uint8_t qs[QK_K/2];      // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];           // quants, high bit
+    uint8_t qs[QK_K/2];           // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_half d;             // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+// (Almost) "true" 2-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 2.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+// 2.3125 bpw quants
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
+
+// 2.5625 bpw quants
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t scales[QK_K/32];
+} block_iq2_s;
+static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
+
+// (Almost) "true" 3-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 3.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint8_t qs[3*QK_K/8];
+} block_iq3_xxs;
+static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
+
+// 3.4375 bpw
+#if QK_K == 64
+#define IQ3S_N_SCALE 2
+#else
+#define IQ3S_N_SCALE QK_K/64
+#endif
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t signs[QK_K/8];
+    uint8_t scales[IQ3S_N_SCALE];
+} block_iq3_s;
+static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
+
+typedef struct {
+    ggml_half d;
+    uint8_t  qs[QK_K/8];
+    uint16_t qh[QK_K/32];
+} block_iq1_s;
+static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
+
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
+static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
+
+#if QK_K == 64
+#define block_iq4_xs block_iq4_nl
+#else
+typedef struct {
+    ggml_half d;
+    uint16_t scales_h;
+    uint8_t  scales_l[QK_K/64];
+    uint8_t  qs[QK_K/2];
+} block_iq4_xs;
+static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
+#endif
+
+#endif // GGML_COMMON_DECL
+#endif // GGML_COMMON_DECL
+
+////////////////////////////////////////////////////////////////////////////////
+
+#ifndef GGML_COMMON_IMPL
 
 #if defined(GGML_COMMON_IMPL_C)
 #include <stdint.h>
@@ -14,7 +418,7 @@
 #define GGML_TABLE_END() };
 
 #define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA)
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
 #include <cstdint>
 
 #define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
@@ -1423,3 +1827,4 @@ GGML_TABLE_END()
 #endif
 
 #endif // GGML_COMMON_IMPL
+#endif // GGML_COMMON_IMPL

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 +1951,12 @@ 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,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_kompute_guid() {

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,
@@ -280,6 +281,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
     id<MTLLibrary> metal_library;
 
     // load library
+    //
+    // - first check if the library is embedded
+    // - then check if the library is in the bundle
+    // - if not found, load the source and compile it
+    // - if that fails, return NULL
     {
         NSBundle * bundle = nil;
 #ifdef SWIFT_PACKAGE
@@ -287,12 +293,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #else
         bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
+
         NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (libPath != nil) {
+
+#if GGML_METAL_EMBED_LIBRARY
+        const bool try_metallib = false;
+#else
+        const bool try_metallib = true;
+#endif
+
+        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (try_metallib && path_lib != nil) {
             // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:libPath];
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
+            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
             metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
@@ -305,38 +320,41 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             extern const char ggml_metallib_start[];
             extern const char ggml_metallib_end[];
 
-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
+            NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
 #else
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
-            NSString * sourcePath;
-            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+            NSString * path_source;
+            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
 
-            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil");
+            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
 
-            if (ggmlMetalPathResources) {
-                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
+            if (path_resource) {
+                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
             } else {
-                sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
             }
-            if (sourcePath == nil) {
+
+            if (path_source == nil) {
                 GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                sourcePath = @"ggml-metal.metal";
+                path_source = @"ggml-metal.metal";
             }
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
-            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
+
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+            NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
             if (error) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
             }
-#endif
+#endif // GGML_METAL_EMBED_LIBRARY
 
             @autoreleasepool {
                 // dictionary of preprocessor macros
                 NSMutableDictionary * prep = [NSMutableDictionary dictionary];
 
 #ifdef GGML_QKK_64
-                prep[@"QK_K"] = @(64);
+                prep[@"GGML_QKK_64"] = @(1);
 #endif
 
                 MTLCompileOptions* options = [MTLCompileOptions new];
@@ -581,8 +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);
@@ -722,6 +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;
@@ -1370,6 +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) {
@@ -1684,6 +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) {
@@ -2414,13 +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;
@@ -2820,6 +2859,12 @@ 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,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };
 
 void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {

ggml-metal.metal

@@ -1,49 +1,15 @@
-#include <metal_stdlib>
-
+#define GGML_COMMON_DECL_METAL
 #define GGML_COMMON_IMPL_METAL
 #include "ggml-common.h"
 
+#include <metal_stdlib>
+
 using namespace metal;
 
 #define MAX(x, y) ((x) > (y) ? (x) : (y))
 #define MIN(x, y) ((x) < (y) ? (x) : (y))
 #define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
 
-#define QK4_0 32
-#define QR4_0 2
-typedef struct {
-    half    d;             // delta
-    uint8_t qs[QK4_0 / 2]; // nibbles / quants
-} block_q4_0;
-
-#define QK4_1 32
-typedef struct {
-    half d;                 // delta
-    half m;                 // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-
-#define QK5_0 32
-typedef struct {
-    half d;                // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-
-#define QK5_1 32
-typedef struct {
-    half d;                 // delta
-    half m;                 // min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-
-#define QK8_0 32
-typedef struct {
-    half    d;         // delta
-    int8_t  qs[QK8_0]; // quants
-} block_q8_0;
-
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32
 
 enum ggml_sort_order {
@@ -2422,6 +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,
@@ -2478,147 +2680,6 @@ kernel void kernel_concat(
     }
 }
 
-//============================================ k-quants ======================================================
-
-#ifndef QK_K
-#define QK_K 256
-#else
-static_assert(QK_K == 256 || QK_K == 64, "QK_K must be 256 or 64");
-#endif
-
-#if QK_K == 256
-#define K_SCALE_SIZE 12
-#else
-#define K_SCALE_SIZE 4
-#endif
-
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    half d;           // super-block scale for quantized scales
-    half dmin;        // super-block scale for quantized mins
-} block_q2_K;
-// 84 bytes / block
-
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#if QK_K == 64
-    uint8_t scales[2];
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    half d;             // super-block scale
-} block_q3_K;
-
-#if QK_K == 64
-typedef struct {
-    half    d[2];          // super-block scales/mins
-    uint8_t scales[2];
-    uint8_t qs[QK_K/2];    // 4-bit quants
-} block_q4_K;
-#else
-typedef struct {
-    half d;             // super-block scale for quantized scales
-    half dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-#endif
-
-#if QK_K == 64
-typedef struct {
-    half  d;                     // super-block scales/mins
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-#else
-typedef struct {
-    half d;                      // super-block scale for quantized scales
-    half dmin;                   // super-block scale for quantized mins
-    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-// 176 bytes / block
-#endif
-
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    half d;                  // super-block scale
-} block_q6_K;
-// 210 bytes / block
-
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-// 66 bytes / block for QK_K = 256, so 2.0625 bpw
-
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-// 74 bytes / block for QK_K = 256, so 2.3125 bpw
-
-// 2.5625 bpw quants
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-
-typedef struct {
-    half d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-// 98 bytes / block for QK_K = 256, so 3.0625 bpw
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-
-typedef struct {
-    half d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    half    d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
-typedef struct {
-    half d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-#endif
-
-//====================================== dot products =========================
-
 void kernel_mul_mv_q2_K_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -4395,10 +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,
@@ -5703,18 +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>;
@@ -5763,18 +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

@@ -1,3 +1,6 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 
@@ -951,7 +954,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         int sum = 0;
 
@@ -966,7 +969,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
             sum += y[i].qs[QK8_1/2 + j];
         }
 
-        y[i].s = sum*d;
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
     }
 }
 
@@ -994,7 +997,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         int32x4_t accv = vdupq_n_s32(0);
 
@@ -1010,7 +1013,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
             accv = vaddq_s32(accv, vi);
         }
 
-        y[i].s = d * vaddvq_s32(accv);
+        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
     }
 #elif defined(__wasm_simd128__)
     for (int i = 0; i < nb; i++) {
@@ -1033,7 +1036,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         v128_t accv = wasm_i32x4_splat(0);
 
@@ -1049,10 +1052,11 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
             accv = wasm_i32x4_add(accv, vi);
         }
 
-        y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
-                      wasm_i32x4_extract_lane(accv, 1) +
-                      wasm_i32x4_extract_lane(accv, 2) +
-                      wasm_i32x4_extract_lane(accv, 3));
+        y[i].s = GGML_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
     }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
@@ -1077,7 +1081,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
 
         // Quantize these floats
         const float d = maxScalar / 127.f;
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
         const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -1101,7 +1105,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
 
 #if defined(__AVX2__)
         // Compute the sum of the quants and set y[i].s
-        y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
 
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -1131,7 +1135,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
         const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
 
         // Convert int32 to int16
         ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -1162,7 +1166,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
         const float d  = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
 
         vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
 
@@ -1179,7 +1183,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
 
         // set y[i].s
         int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = sum*d;
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
     }
 #else
     GGML_UNUSED(nb);
@@ -4019,10 +4023,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             const block_q8_1 * restrict b_y0 = &vy0[i];
             const block_q8_1 * restrict b_y1 = &vy1[i];
 
-            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
-                                   GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
-                                   GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
-                                   GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
+            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                   GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
             summs0 += summs_t;
 
             const uint8x16_t m4b = vdupq_n_u8(0x0F);
@@ -4087,7 +4091,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         const block_q8_1 * restrict y0 = &y[i + 0];
         const block_q8_1 * restrict y1 = &y[i + 1];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -4110,8 +4114,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
         const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
     *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@@ -4124,9 +4128,9 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     // Main loop
     for (int i = 0; i < nb; ++i) {
         const float d0 = GGML_FP16_TO_FP32(x[i].d);
-        const float d1 = y[i].d;
+        const float d1 = GGML_FP16_TO_FP32(y[i].d);
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
 
         const __m256 d0v = _mm256_set1_ps( d0 );
         const __m256 d1v = _mm256_set1_ps( d1 );
@@ -4178,7 +4182,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
     }
 
     *s = sumf;
@@ -4196,7 +4200,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
     }
 
     *s = sumf;
@@ -4532,8 +4536,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
-        summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s;
-        summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s;
+        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
 
         // extract the 5th bit via lookup table ((b) << 4)
         memcpy(&qh0, x0->qh, sizeof(qh0));
@@ -4577,10 +4581,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
     }
 
     *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@@ -4597,7 +4601,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         const block_q5_1 * restrict x0 = &x[i];
         const block_q8_1 * restrict y0 = &y[i];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s;
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
 
         const v128_t m4b = wasm_i8x16_splat(0x0F);
 
@@ -4644,7 +4648,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
                                            wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                             wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                            wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
     }
 
     *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -4659,14 +4663,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     for (int i = 0; i < nb; i++) {
         const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
 
         __m256i qx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
         qx = _mm256_or_si256(qx, bxhi);
 
-        const __m256 dy = _mm256_set1_ps(y[i].d);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
         const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -4686,7 +4690,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
     for (int i = 0; i < nb; i++) {
         const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
 
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
 
         __m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
         const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@@ -4700,7 +4704,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
         bxh = _mm_or_si128(bxh, bxhih);
         bx_0 = MM256_SET_M128I(bxh, bxl);
 
-        const __m256 dy = _mm256_set1_ps(y[i].d);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
         const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@@ -4767,7 +4771,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
     }
 
     *s = sumf;
@@ -4791,7 +4795,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
             sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
         }
 
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
     }
 
     *s = sumf;
@@ -9025,7 +9029,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
                                      vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
             qs += 8;
 
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
             vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
             vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vceqq_u8(vs.val[0], mask2);
@@ -9034,7 +9038,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
             q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
             q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
 
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
             vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
             vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vceqq_u8(vs.val[0], mask2);
@@ -9105,12 +9109,12 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
                                                    iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
             qs += 8;
 
-            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
             aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
             const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
             const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
 
-            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
             aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
             const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
             const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
@@ -9386,7 +9390,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
                                                         iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
 
 
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
             vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
             vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
@@ -9395,7 +9399,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
             q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
             q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
 
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
             vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
             vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
@@ -11701,9 +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];
@@ -11715,6 +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]);
@@ -11732,12 +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];
@@ -11792,9 +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]);
+            }
         }
     }
 
@@ -11819,7 +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);
@@ -11828,14 +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) {
@@ -11853,7 +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-quants.h

@@ -1,248 +1,12 @@
 #pragma once
 
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
 // GGML internal header
 
-#include "ggml-impl.h"
-
-#include <stdint.h>
-#include <stddef.h>
-
-#define QK4_0 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    ggml_fp16_t m;          // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // min
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    int8_t  qs[QK8_0];     // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    float d;               // delta
-    float s;               // d * sum(qs[i])
-    int8_t  qs[QK8_1];     // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    ggml_fp16_t d;           // super-block scale for quantized scales
-    ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;             // super-block scale for quantized scales
-    ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;               // super-block scale for quantized scales
-    ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-// (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 2.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-// 2.3125 bpw quants
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-// 2.5625 bpw quants
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
-// (Almost) "true" 3-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 3.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-//typedef struct block_iq4_nl block_iq4_xs;
-#else
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif

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;
@@ -202,24 +226,29 @@ namespace dpct
             // Version string has the following format:
             // a. OpenCL<space><major.minor><space><vendor-specific-information>
             // b. <major.minor>
+            // c. <AmdGcnArchName> e.g gfx1030
             std::string ver;
             ver = dev.get_info<sycl::info::device::version>();
             std::string::size_type i = 0;
-            while (i < ver.size())
-            {
-                if (isdigit(ver[i]))
-                    break;
-                i++;
+            while (i < ver.size()) {
+              if (isdigit(ver[i]))
+                break;
+              i++;
             }
             major = std::stoi(&(ver[i]));
-            while (i < ver.size())
-            {
-                if (ver[i] == '.')
-                    break;
-                i++;
+            while (i < ver.size()) {
+              if (ver[i] == '.')
+                break;
+              i++;
+            }
+            if (i < ver.size()) {
+              // a. and b.
+              i++;
+              minor = std::stoi(&(ver[i]));
+            } else {
+              // c.
+              minor = 0;
             }
-            i++;
-            minor = std::stoi(&(ver[i]));
         }
 
         template <typename tag, typename T>
@@ -937,17 +966,67 @@ 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)
@@ -3144,6 +3223,7 @@ namespace dpct
 
 } // COPY from DPCT head files
 
+#define GGML_COMMON_DECL_SYCL
 #define GGML_COMMON_IMPL_SYCL
 #include "ggml-common.h"
 
@@ -3196,6 +3276,11 @@ 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");
 
@@ -3312,66 +3397,6 @@ typedef void (*ggml_sycl_op_flatten_t)(const ggml_tensor *src0,
                                        const float *src1_dd, float *dst_dd,
                                        const dpct::queue_ptr &main_stream);
 
-// QK = number of values after dequantization
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QK4_0 32
-#define QR4_0 2
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-typedef struct dpct_type_block_q4_0 {
-    sycl::half d;           // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-#define QR4_1 2
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-typedef struct dpct_type_block_q4_1 {
-    sycl::half2 dm;         // dm.x = delta, dm.y = min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-#define QR5_0 2
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-typedef struct dpct_type_block_q5_0 {
-    sycl::half d;           // delta
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2];  // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-#define QR5_1 2
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-typedef struct dpct_type_block_q5_1 {
-    sycl::half2 dm;         // dm.x = delta, dm.y = min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-#define QR8_0 1
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-typedef struct dpct_type_block_q8_0 {
-    sycl::half d;           // delta
-    int8_t  qs[QK8_0];      // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-#define QR8_1 1
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-typedef struct dpct_type_block_q8_1 {
-    sycl::half2 ds;         // ds.x = delta, ds.y = sum
-    int8_t  qs[QK8_0];      // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
-
 typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_sycl_t)(int **x_ql, sycl::half2 **x_dm,
                                       int **x_qh, int **x_sc);
@@ -3388,137 +3413,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
     const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ms,
     const int &i, const int &j, const int &k);
 
-//================================= k-quants
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-#define QR2_K 4
-#define QI2_K (QK_K / (4*QR2_K))
-typedef struct dpct_type_block_q2_K {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    sycl::half2 dm;          // super-block scale for quantized scales/mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-#define QR3_K 4
-#define QI3_K (QK_K / (4*QR3_K))
-typedef struct dpct_type_block_q3_K {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#ifdef GGML_QKK_64
-    uint8_t scales[2]; // scales, quantized with 8 bits
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    sycl::half d; // super-block scale
-} block_q3_K;
-//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
-
-#define QR4_K 2
-#define QI4_K (QK_K / (4*QR4_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    sycl::half dm[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == sizeof(sycl::half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct dpct_type_block_q4_K {
-    sycl::half2 dm;            // super-block scale for quantized scales/mins
-    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-#define QR5_K 2
-#define QI5_K (QK_K / (4*QR5_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    sycl::half d;                  // super-block scale
-    int8_t scales[QK_K/16];  // block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct dpct_type_block_q5_K {
-    sycl::half2 dm;               // super-block scale for quantized scales/mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-#define QR6_K 2
-#define QI6_K (QK_K / (4*QR6_K))
-typedef struct dpct_type_block_q6_K {
-    uint8_t ql[QK_K/2];   // quants, lower 4 bits
-    uint8_t qh[QK_K/4];   // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales
-    sycl::half d;            // delta
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
-
-#define QR2_XXS 8
-#define QI2_XXS (QK_K / (4*QR2_XXS))
-typedef struct dpct_type_block_iq2_xxs {
-    sycl::half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-#define QR2_XS 8
-#define QI2_XS (QK_K / (4*QR2_XS))
-typedef struct dpct_type_block_iq2_xs {
-    sycl::half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-#define QR3_XXS 8
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-typedef struct dpct_type_block_iq3_xxs {
-    sycl::half d;
-    uint8_t qs[3*(QK_K/8)];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-#define QR3_XS 8
-#define QI3_XS (QK_K / (4*QR3_XS))
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    sycl::half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-#define QR1_S 8
-#define QI1_S (QK_K / (4*QR1_S))
-typedef struct {
-    sycl::half d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -3586,12 +3480,31 @@ 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);
@@ -3607,7 +3520,7 @@ 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();
             }
         }
@@ -3636,7 +3549,7 @@ 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();
@@ -3656,8 +3569,8 @@ 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) {
@@ -3666,8 +3579,16 @@ 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;
         }
 };
 
@@ -3676,11 +3597,14 @@ 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
@@ -4891,10 +4815,9 @@ static void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restr
 template<typename dst_t>
 static void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                      const sycl::nd_item<3> &item_ct1,
-                                     const uint64_t *iq1s_grid,
+                                     const uint32_t *iq1s_grid,
                                      const uint8_t *ksigns_iq2xs,
                                      const uint8_t *kmask_iq2xs) {
-
     const int i = item_ct1.get_group(2);
     const block_iq1_s * x = (const block_iq1_s  *) vx;
 
@@ -4903,11 +4826,15 @@ static void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restr
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const int i8 = 4*ib+il;
-    uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
-    const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
-    const float d = (float)x[i].d * (2*(h & 7) + 1);
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
+    const uint8_t  * qs = x[i].qs + 8*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq1s_grid + qs[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq1s_grid + qs[2*il+1]);
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 0xf) + 1);
+    const uint8_t signs = ksigns_iq2xs[(x[i].qh[ib] >> 3*il) & 7];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
 #else
     assert(false);
 #endif
@@ -7803,28 +7730,27 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
 static __dpct_inline__ float
 vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
                      const block_q8_1 *__restrict__ bq8_1, const int &iqs,
-                     const uint64_t *iq1s_grid, const uint64_t *ksigns64) {
+                     const uint32_t *iq1s_grid, const uint64_t *ksigns64) {
 #if QK_K == 256
     const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
 
     const int ib32 = iqs;
-    int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
-    const uint8_t h1 = bq1->scales[2*ib32+0];
-    const uint8_t h2 = bq1->scales[2*ib32+1];
-    const int * q8 = (const int *)bq8_1[ib32].qs;
-    const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 2; ++j) {
-        sumi1 = dpct::dp4a(q8[j+0], grid1[j], sumi1);
-        sumi2 = dpct::dp4a(q8[j+2], grid2[j], sumi2);
-        sumi3 = dpct::dp4a(q8[j+4], grid3[j], sumi3);
-        sumi4 = dpct::dp4a(q8[j+6], grid4[j], sumi4);
+    const uint8_t  * qs = bq1->qs + 4*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq1s_grid + qs[l]);
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (qs[l] >> 8));
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs[0], signs[0], std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs[1], signs[1], std::minus<>());
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        q8 += 8;
     }
-    const float d = (float)bq1->d  * bq8_1[ib32].ds[0];
-    return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
-                sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
+    const float d = (float)bq1->d * bq8_1[ib32].ds[0] * 0.25f;
+    return d * sumi;
 #else
     assert(false);
     return 0.f;
@@ -8644,7 +8570,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void * __restrict__ vx, const void *
 template <int qk, int qi, typename block_q_t, int vdr>
 static void mul_mat_vec_q_iq1_s_q8_1(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
                           const sycl::nd_item<3> &item_ct1,
-                          const uint64_t *iq1s_grid_ptr, const uint64_t *ksigns64_ptr ) {
+                          const uint32_t *iq1s_grid_ptr, const uint64_t *ksigns64_ptr ) {
     const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
                     item_ct1.get_local_id(1);
 
@@ -10406,7 +10332,7 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
                                         dpct::queue_ptr stream) {
     const int nb = k / QK_K;
     {
-        iq1s_grid.init(*stream);
+        iq1s_grid_gpu.init(*stream);
         ksigns_iq2xs.init(*stream);
         kmask_iq2xs.init(*stream);
 
@@ -10414,7 +10340,7 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
                                      {sycl::aspect::fp16});
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid.get_ptr();
+            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
             auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
             auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();
 
@@ -11154,11 +11080,11 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        iq1s_grid.init(*stream);
+        iq1s_grid_gpu.init(*stream);
         ksigns64.init(*stream);
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid.get_ptr();
+            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
             auto ksigns64_ptr_ct1 = ksigns64.get_ptr();
 
             cgh.parallel_for(
@@ -13182,17 +13108,20 @@ 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());
@@ -13200,19 +13129,35 @@ 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) {
@@ -13248,23 +13193,6 @@ 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)
@@ -13273,49 +13201,15 @@ 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;
     }
@@ -13326,6 +13220,63 @@ 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;
@@ -16725,22 +16676,24 @@ 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) {
+    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
@@ -17432,13 +17385,19 @@ static ggml_backend_i ggml_backend_sycl_interface = {
     /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type,
     /* .set_tensor_async        = */ ggml_backend_sycl_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_sycl_get_tensor_async,
-    /* .cpy_tensor_async        = */ ggml_backend_sycl_cpy_tensor_async,
+    /* .cpy_tensor_async        = */ NULL, //ggml_backend_sycl_cpy_tensor_async, // TODO: update for the new interface
     /* .synchronize             = */ ggml_backend_sycl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_compute      = */ NULL,
     /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
     /* .supports_op             = */ ggml_backend_sycl_supports_op,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -17488,11 +17447,42 @@ 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

@@ -13,7 +13,7 @@
 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);
@@ -29,6 +29,11 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_typ
 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();
 #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 +5699,12 @@ 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,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_vk_guid() {

ggml.c

@@ -282,8 +282,6 @@ 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)
@@ -470,6 +468,19 @@ 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 +929,101 @@ 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
@@ -2532,9 +2638,7 @@ 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();
@@ -10997,7 +11101,6 @@ 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 +11300,6 @@ 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) {
@@ -11560,8 +11662,6 @@ static void ggml_compute_forward_get_rows_q(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    assert(params->ith == 0);
-
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
     }
@@ -11569,7 +11669,7 @@ static void ggml_compute_forward_get_rows_q(
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
@@ -11579,17 +11679,25 @@ static void ggml_compute_forward_get_rows_q(
     assert(nb00 == ggml_type_size(type));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                dequantize_row_q(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        dequantize_row_q(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
 }
 
@@ -11600,8 +11708,6 @@ static void ggml_compute_forward_get_rows_f16(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    assert(params->ith == 0);
-
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
     }
@@ -11609,24 +11715,32 @@ static void ggml_compute_forward_get_rows_f16(
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     assert(ne0  == nc);
     assert(ne02 == ne11);
     assert(nb00 == sizeof(ggml_fp16_t));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                ggml_fp16_to_fp32_row(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        ggml_fp16_to_fp32_row(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
 }
 
@@ -11637,8 +11751,6 @@ static void ggml_compute_forward_get_rows_f32(
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
 
-    assert(params->ith == 0);
-
     if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
         return;
     }
@@ -11646,24 +11758,32 @@ static void ggml_compute_forward_get_rows_f32(
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     assert(ne0  == nc);
     assert(ne02 == ne11);
     assert(nb00 == sizeof(float));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                ggml_vec_cpy_f32(nc,
-                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
-                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        ggml_vec_cpy_f32(nc,
+                (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
     }
 }
 
@@ -12400,6 +12520,8 @@ 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);
@@ -12486,6 +12608,8 @@ 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);
@@ -15921,14 +16045,7 @@ 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) {
@@ -15940,7 +16057,7 @@ 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);
@@ -17796,7 +17913,7 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const
     node->perf_time_us += time_us_cur;
 }
 
-static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_threads) {
     int n_tasks = 0;
 
     switch (node->op) {
@@ -17877,6 +17994,12 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = n_threads;
             } break;
+        case GGML_OP_GET_ROWS:
+            {
+                // FIXME: the cost of launching additional threads decreases performance with GPU offloading
+                //n_tasks = MIN(n_threads, ggml_nelements(node->src[1]));
+                n_tasks = MIN(n_cur_threads, ggml_nelements(node->src[1]));
+            } break;
         case GGML_OP_SCALE:
         case GGML_OP_SET:
         case GGML_OP_CONT:
@@ -17884,7 +18007,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
-        case GGML_OP_GET_ROWS:
         case GGML_OP_GET_ROWS_BACK:
         case GGML_OP_DIAG:
             {
@@ -18102,7 +18224,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
                 /* FINALIZE */
                 struct ggml_tensor * node = cgraph->nodes[node_n];
                 if (GGML_OP_HAS_FINALIZE[node->op]) {
-                    params.nth = ggml_get_n_tasks(node, n_threads);
+                    params.nth = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);
                     ggml_compute_forward(&params, node);
                 }
                 ggml_graph_compute_perf_stats_node(node, state->shared);
@@ -18112,7 +18234,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             while (++node_n < cgraph->n_nodes) {
                 GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
                 struct ggml_tensor * node = cgraph->nodes[node_n];
-                const int n_tasks = ggml_get_n_tasks(node, n_threads);
+                const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);
 
                 state->shared->perf_node_start_cycles  = ggml_perf_cycles();
                 state->shared->perf_node_start_time_us = ggml_perf_time_us();
@@ -18160,7 +18282,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
         /* INIT & COMPUTE */
         struct ggml_tensor * node = cgraph->nodes[node_n];
-        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+        const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);
 
         struct ggml_compute_params params = {
             /*.type  =*/ GGML_TASK_TYPE_INIT,
@@ -18225,7 +18347,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
-        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+        const int n_tasks = ggml_get_n_tasks(node, n_threads, 1);
 
         max_tasks = MAX(max_tasks, n_tasks);
 

ggml.h

@@ -337,24 +337,24 @@ extern "C" {
     struct ggml_object;
     struct ggml_context;
 
+    // NOTE: always add types at the end of the enum to keep backward compatibility
     enum ggml_type {
-        GGML_TYPE_F32  = 0,
-        GGML_TYPE_F16  = 1,
-        GGML_TYPE_Q4_0 = 2,
-        GGML_TYPE_Q4_1 = 3,
+        GGML_TYPE_F32     = 0,
+        GGML_TYPE_F16     = 1,
+        GGML_TYPE_Q4_0    = 2,
+        GGML_TYPE_Q4_1    = 3,
         // GGML_TYPE_Q4_2 = 4, support has been removed
-        // GGML_TYPE_Q4_3 (5) support has been removed
-        GGML_TYPE_Q5_0 = 6,
-        GGML_TYPE_Q5_1 = 7,
-        GGML_TYPE_Q8_0 = 8,
-        GGML_TYPE_Q8_1 = 9,
-        // k-quantizations
-        GGML_TYPE_Q2_K = 10,
-        GGML_TYPE_Q3_K = 11,
-        GGML_TYPE_Q4_K = 12,
-        GGML_TYPE_Q5_K = 13,
-        GGML_TYPE_Q6_K = 14,
-        GGML_TYPE_Q8_K = 15,
+        // GGML_TYPE_Q4_3 = 5, support has been removed
+        GGML_TYPE_Q5_0    = 6,
+        GGML_TYPE_Q5_1    = 7,
+        GGML_TYPE_Q8_0    = 8,
+        GGML_TYPE_Q8_1    = 9,
+        GGML_TYPE_Q2_K    = 10,
+        GGML_TYPE_Q3_K    = 11,
+        GGML_TYPE_Q4_K    = 12,
+        GGML_TYPE_Q5_K    = 13,
+        GGML_TYPE_Q6_K    = 14,
+        GGML_TYPE_Q8_K    = 15,
         GGML_TYPE_IQ2_XXS = 16,
         GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_IQ3_XXS = 18,
@@ -363,9 +363,11 @@ extern "C" {
         GGML_TYPE_IQ3_S   = 21,
         GGML_TYPE_IQ2_S   = 22,
         GGML_TYPE_IQ4_XS  = 23,
-        GGML_TYPE_I8,
-        GGML_TYPE_I16,
-        GGML_TYPE_I32,
+        GGML_TYPE_I8      = 24,
+        GGML_TYPE_I16     = 25,
+        GGML_TYPE_I32     = 26,
+        GGML_TYPE_I64     = 27,
+        GGML_TYPE_F64     = 28,
         GGML_TYPE_COUNT,
     };
 
@@ -383,20 +385,20 @@ extern "C" {
 
     // model file types
     enum ggml_ftype {
-        GGML_FTYPE_UNKNOWN     = -1,
-        GGML_FTYPE_ALL_F32     = 0,
-        GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        GGML_FTYPE_UNKNOWN        = -1,
+        GGML_FTYPE_ALL_F32        = 0,
+        GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors

gguf-py/gguf/constants.py

@@ -32,6 +32,7 @@ class Keys:
         FILE_TYPE            = "general.file_type"
 
     class LLM:
+        VOCAB_SIZE            = "{arch}.vocab_size"
         CONTEXT_LENGTH        = "{arch}.context_length"
         EMBEDDING_LENGTH      = "{arch}.embedding_length"
         BLOCK_COUNT           = "{arch}.block_count"
@@ -41,6 +42,7 @@ class Keys:
         EXPERT_COUNT          = "{arch}.expert_count"
         EXPERT_USED_COUNT     = "{arch}.expert_used_count"
         POOLING_TYPE          = "{arch}.pooling_type"
+        LOGIT_SCALE           = "{arch}.logit_scale"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -120,6 +122,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA      = auto()
     STARCODER2 = auto()
     MAMBA      = auto()
+    COMMAND_R  = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -186,6 +189,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA:          "gemma",
     MODEL_ARCH.STARCODER2:     "starcoder2",
     MODEL_ARCH.MAMBA:          "mamba",
+    MODEL_ARCH.COMMAND_R:      "command-r",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -578,6 +582,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_D,
         MODEL_TENSOR.SSM_OUT,
     ],
+    MODEL_ARCH.COMMAND_R: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -661,6 +677,11 @@ class GGMLQuantizationType(IntEnum):
     IQ3_S   = 21
     IQ2_S   = 22
     IQ4_XS  = 23
+    I8      = 24
+    I16     = 25
+    I32     = 26
+    I64     = 27
+    F64     = 28
 
 
 class GGUFEndian(IntEnum):
@@ -727,6 +748,11 @@ GGML_QUANT_SIZES = {
     GGMLQuantizationType.IQ3_S:   (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4),
     GGMLQuantizationType.IQ2_S:   (256, 2 + QK_K // 4 + QK_K // 16),
     GGMLQuantizationType.IQ4_XS:  (256, 2 + 2 + QK_K // 2 + QK_K // 64),
+    GGMLQuantizationType.I8:      (1, 1),
+    GGMLQuantizationType.I16:     (1, 2),
+    GGMLQuantizationType.I32:     (1, 4),
+    GGMLQuantizationType.I64:     (1, 8),
+    GGMLQuantizationType.F64:     (1, 8),
 }
 
 
@@ -746,6 +772,7 @@ KEY_GENERAL_SOURCE_HF_REPO       = Keys.General.SOURCE_HF_REPO
 KEY_GENERAL_FILE_TYPE            = Keys.General.FILE_TYPE
 
 # LLM
+KEY_VOCAB_SIZE            = Keys.LLM.VOCAB_SIZE
 KEY_CONTEXT_LENGTH        = Keys.LLM.CONTEXT_LENGTH
 KEY_EMBEDDING_LENGTH      = Keys.LLM.EMBEDDING_LENGTH
 KEY_BLOCK_COUNT           = Keys.LLM.BLOCK_COUNT

gguf-py/gguf/gguf_reader.py

@@ -242,12 +242,27 @@ class GGUFReader:
             n_bytes = n_elems * type_size // block_size
             data_offs = int(start_offs + offset_tensor[0])
             item_type: npt.DTypeLike
-            if ggml_type == GGMLQuantizationType.F32:
-                item_count = n_elems
-                item_type = np.float32
-            elif ggml_type == GGMLQuantizationType.F16:
+            if ggml_type == GGMLQuantizationType.F16:
                 item_count = n_elems
                 item_type = np.float16
+            elif ggml_type == GGMLQuantizationType.F32:
+                item_count = n_elems
+                item_type = np.float32
+            elif ggml_type == GGMLQuantizationType.F64:
+                item_count = n_elems
+                item_type = np.float64
+            elif ggml_type == GGMLQuantizationType.I8:
+                item_count = n_elems
+                item_type = np.int8
+            elif ggml_type == GGMLQuantizationType.I16:
+                item_count = n_elems
+                item_type = np.int16
+            elif ggml_type == GGMLQuantizationType.I32:
+                item_count = n_elems
+                item_type = np.int32
+            elif ggml_type == GGMLQuantizationType.I64:
+                item_count = n_elems
+                item_type = np.int64
             else:
                 item_count = n_bytes
                 item_type = np.uint8

gguf-py/gguf/gguf_writer.py

@@ -196,9 +196,6 @@ class GGUFWriter:
         if self.state is not WriterState.EMPTY:
             raise ValueError(f'Expected output file to be empty, got {self.state}')
 
-        if raw_dtype is None and tensor_dtype not in (np.float32, np.float16):
-            raise ValueError("Only F32 and F16 tensors are supported for now")
-
         encoded_name = name.encode("utf8")
         self.ti_data += self._pack("Q", len(encoded_name))
         self.ti_data += encoded_name
@@ -207,7 +204,22 @@ class GGUFWriter:
         for i in range(n_dims):
             self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
         if raw_dtype is None:
-            dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
+            if tensor_dtype == np.float16:
+                dtype = GGMLQuantizationType.F16
+            elif tensor_dtype == np.float32:
+                dtype = GGMLQuantizationType.F32
+            elif tensor_dtype == np.float64:
+                dtype = GGMLQuantizationType.F64
+            elif tensor_dtype == np.int8:
+                dtype = GGMLQuantizationType.I8
+            elif tensor_dtype == np.int16:
+                dtype = GGMLQuantizationType.I16
+            elif tensor_dtype == np.int32:
+                dtype = GGMLQuantizationType.I32
+            elif tensor_dtype == np.int64:
+                dtype = GGMLQuantizationType.I64
+            else:
+                raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
         else:
             dtype = raw_dtype
         self.ti_data += self._pack("I", dtype)
@@ -313,6 +325,9 @@ class GGUFWriter:
         self.data_alignment = alignment
         self.add_uint32(Keys.General.ALIGNMENT, alignment)
 
+    def add_vocab_size(self, size: int) -> None:
+        self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
+
     def add_context_length(self, length: int) -> None:
         self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length)
 
@@ -346,6 +361,9 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
+    def add_logit_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
+
     def add_expert_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)
 

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.7.0"
+version = "0.8.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

llama.h

@@ -59,9 +59,10 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
-        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
-        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
+        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -234,7 +235,8 @@ extern "C" {
     struct llama_context_params {
         uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
-        uint32_t n_batch;           // prompt processing maximum batch size
+        uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
+        uint32_t n_ubatch;          // physical maximum batch size
         uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
         uint32_t n_threads;         // number of threads to use for generation
         uint32_t n_threads_batch;   // number of threads to use for batch processing
@@ -377,6 +379,7 @@ extern "C" {
 
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
 
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
@@ -385,6 +388,7 @@ extern "C" {
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
 
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
@@ -432,10 +436,24 @@ extern "C" {
     // Returns 0 on success
     LLAMA_API int32_t llama_model_apply_lora_from_file(
             const struct llama_model * model,
-                      const char * path_lora,
-                           float   scale,
-                      const char * path_base_model,
-                         int32_t   n_threads);
+                          const char * path_lora,
+                               float   scale,
+                          const char * path_base_model,
+                             int32_t   n_threads);
+
+    // Apply a loaded control vector to a llama_context, or if data is NULL, clear
+    // the currently loaded vector.
+    // n_embd should be the size of a single layer's control, and data should point
+    // to an n_embd x n_layers buffer starting from layer 1.
+    // il_start and il_end are the layer range the vector should apply to (both inclusive)
+    // See llama_control_vector_load in common to load a control vector.
+    LLAMA_API int32_t llama_control_vector_apply(
+            struct llama_context * lctx,
+                     const float * data,
+                          size_t   len,
+                         int32_t   n_embd,
+                         int32_t   il_start,
+                         int32_t   il_end);
 
     //
     // KV cache
@@ -650,6 +668,11 @@ extern "C" {
     // Set abort callback
     LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
 
+    // Wait until all computations are finished
+    // This is automatically done when using one of the functions below to obtain the computation results
+    // and is not necessary to call it explicitly in most cases
+    LLAMA_API void llama_synchronize(struct llama_context * ctx);
+
     // Token logits obtained from the last call to llama_decode()
     // The logits for the last token are stored in the last row
     // Logits for which llama_batch.logits[i] == 0 are undefined

tests/CMakeLists.txt

@@ -1,66 +1,75 @@
-function(llama_build_executable source)
-    get_filename_component(TEST_TARGET ${source} NAME_WE)
+# Builds and runs a test source file.
+# Optional args:
+# - NAME: name of the executable & test target (defaults to the source file name without extension)
+# - LABEL: label for the test (defaults to main)
+# - ARGS: arguments to pass to the test executable
+# - WORKING_DIRECTORY
+function(llama_test source)
+    include(CMakeParseArguments)
+    set(options)
+    set(oneValueArgs NAME LABEL WORKING_DIRECTORY)
+    set(multiValueArgs ARGS)
+    cmake_parse_arguments(LLAMA_TEST "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    if (NOT DEFINED LLAMA_TEST_LABEL)
+        set(LLAMA_TEST_LABEL "main")
+    endif()
+    if (NOT DEFINED LLAMA_TEST_WORKING_DIRECTORY)
+        set(LLAMA_TEST_WORKING_DIRECTORY .)
+    endif()
+    if (DEFINED LLAMA_TEST_NAME)
+        set(TEST_TARGET ${LLAMA_TEST_NAME})
+    else()
+        get_filename_component(TEST_TARGET ${source} NAME_WE)
+    endif()
+
     add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
-    target_link_libraries(${TEST_TARGET} PRIVATE common)
+    target_link_libraries(${TEST_TARGET} PRIVATE common json-schema-to-grammar)
+    add_test(
+        NAME ${TEST_TARGET}
+        WORKING_DIRECTORY ${LLAMA_TEST_WORKING_DIRECTORY}
+        COMMAND $<TARGET_FILE:${TEST_TARGET}>
+        ${LLAMA_TEST_ARGS})
+
+    set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${LLAMA_TEST_LABEL})
 endfunction()
 
-function(llama_test_executable name source)
-    get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_test(NAME ${name} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
-    set_property(TEST ${name} PROPERTY LABELS "main")
-endfunction()
+# llama_test(test-double-float.cpp) # SLOW
+llama_test(test-quantize-fns.cpp)
+llama_test(test-quantize-perf.cpp)
+llama_test(test-sampling.cpp)
+llama_test(test-chat-template.cpp)
 
-function(llama_build_and_test_executable source)
-    llama_build_and_test_executable_with_label(${source} "main")
-endfunction()
+llama_test(test-tokenizer-0-llama.cpp  NAME test-tokenizer-0-llama            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test(test-tokenizer-0-falcon.cpp NAME test-tokenizer-0-falcon           ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 
-function(llama_build_and_test_executable_with_label source label)
-    get_filename_component(TEST_TARGET ${source} NAME_WE)
-    add_executable(${TEST_TARGET} ${source} get-model.cpp)
-    install(TARGETS ${TEST_TARGET} RUNTIME)
-    target_link_libraries(${TEST_TARGET} PRIVATE common)
-    add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
-    set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label})
-endfunction()
+llama_test(test-tokenizer-1-llama.cpp  NAME test-tokenizer-1-llama            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test(test-tokenizer-1-llama.cpp  NAME test-tokenizer-1-baichuan         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
 
-# llama_build_and_test_executable(test-double-float.cpp) # SLOW
-llama_build_and_test_executable(test-quantize-fns.cpp)
-llama_build_and_test_executable(test-quantize-perf.cpp)
-llama_build_and_test_executable(test-sampling.cpp)
-llama_build_and_test_executable(test-chat-template.cpp)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-falcon           ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-aquila           ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-mpt              ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-stablelm-3b-4e1t ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-gpt-neox         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-refact           ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-starcoder        ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-gpt2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt2.gguf)
+#llama_test(test-tokenizer-1-bpe.cpp    NAME test-tokenizer-1-bloom            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
 
-llama_build_executable(test-tokenizer-0-llama.cpp)
-llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
+llama_test(test-grammar-parser.cpp)
+llama_test(test-llama-grammar.cpp)
+llama_test(test-grad0.cpp)
+# llama_test(test-opt.cpp) # SLOW
+llama_test(test-backend-ops.cpp)
 
-llama_build_executable(test-tokenizer-0-falcon.cpp)
-llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
+llama_test(test-rope.cpp)
 
-llama_build_executable(test-tokenizer-1-llama.cpp)
-llama_test_executable (test-tokenizer-1-llama    test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
-llama_test_executable (test-tokenizer-1-baichuan test-tokenizer-1-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+llama_test(test-model-load-cancel.cpp  LABEL "model")
+llama_test(test-autorelease.cpp        LABEL "model")
 
-llama_build_executable(test-tokenizer-1-bpe.cpp)
-llama_test_executable (test-tokenizer-1-falcon           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_test_executable (test-tokenizer-1-aquila           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-llama_test_executable (test-tokenizer-1-mpt              test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-llama_test_executable (test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf)
-llama_test_executable (test-tokenizer-1-gpt-neox         test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-llama_test_executable (test-tokenizer-1-refact           test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-llama_test_executable (test-tokenizer-1-starcoder        test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
-llama_test_executable (test-tokenizer-1-gpt2             test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt2.gguf)
-# llama_test_executable (test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG
-
-llama_build_and_test_executable(test-grammar-parser.cpp)
-llama_build_and_test_executable(test-llama-grammar.cpp)
-llama_build_and_test_executable(test-grad0.cpp)
-# llama_build_and_test_executable(test-opt.cpp) # SLOW
-llama_build_and_test_executable(test-backend-ops.cpp)
-
-llama_build_and_test_executable(test-rope.cpp)
-
-llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model")
-llama_build_and_test_executable_with_label(test-autorelease.cpp "model")
+llama_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
+target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server)
 
 # dummy executable - not installed
 get_filename_component(TEST_TARGET test-c.c NAME_WE)

tests/run-json-schema-to-grammar.mjs

@@ -0,0 +1,10 @@
+import { readFileSync } from "fs"
+import { SchemaConverter } from "../examples/server/public/json-schema-to-grammar.mjs"
+
+const [, , file] = process.argv
+const url = `file://${file}`
+let schema = JSON.parse(readFileSync(file, "utf8"));
+const converter = new SchemaConverter({})
+schema = await converter.resolveRefs(schema, url)
+converter.visit(schema, '')
+console.log(converter.formatGrammar())

tests/test-backend-ops.cpp

@@ -2091,6 +2091,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         }
     }
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,  128, { 8,  1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,  128, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 2,   64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,   64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+
     for (ggml_type type_a : all_types) {
         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
             for (int n_mats : {2, 4, 8}) {
@@ -2222,8 +2229,8 @@ static void usage(char ** argv) {
 
 int main(int argc, char ** argv) {
     test_mode mode = MODE_TEST;
-    const char * op_name = NULL;
-    const char * backend = NULL;
+    const char * op_name_filter = NULL;
+    const char * backend_filter = NULL;
 
     for (int i = 1; i < argc; i++) {
         if (strcmp(argv[i], "test") == 0) {
@@ -2232,14 +2239,14 @@ int main(int argc, char ** argv) {
             mode = MODE_PERF;
         } else if (strcmp(argv[i], "-o") == 0) {
             if (i + 1 < argc) {
-                op_name = argv[++i];
+                op_name_filter = argv[++i];
             } else {
                 usage(argv);
                 return 1;
             }
         } else if (strcmp(argv[i], "-b") == 0) {
             if (i + 1 < argc) {
-                backend = argv[++i];
+                backend_filter = argv[++i];
             } else {
                 usage(argv);
                 return 1;
@@ -2258,7 +2265,7 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < ggml_backend_reg_get_count(); i++) {
         printf("Backend %zu/%zu (%s)\n", i + 1, ggml_backend_reg_get_count(), ggml_backend_reg_get_name(i));
 
-        if (backend != NULL && strcmp(backend, ggml_backend_reg_get_name(i)) != 0) {
+        if (backend_filter != NULL && strcmp(backend_filter, ggml_backend_reg_get_name(i)) != 0) {
             printf("  Skipping\n");
             n_ok++;
             continue;
@@ -2266,9 +2273,17 @@ int main(int argc, char ** argv) {
 
         ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL);
         GGML_ASSERT(backend != NULL);
+
+        if (backend_filter == NULL && ggml_backend_is_cpu(backend)) {
+            printf("  Skipping CPU backend\n");
+            ggml_backend_free(backend);
+            n_ok++;
+            continue;
+        }
+
         printf("  Backend name: %s\n", ggml_backend_name(backend));
 
-        bool ok = test_backend(backend, mode, op_name);
+        bool ok = test_backend(backend, mode, op_name_filter);
 
         printf("  Backend %s: ", ggml_backend_name(backend));
         if (ok) {

tests/test-chat-template.cpp

@@ -31,6 +31,8 @@ int main(void) {
         "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}",
         // google/gemma-7b-it
         "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\\n' + message['content'] | trim + '<end_of_turn>\\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\\n'}}{% endif %}",
+        // OrionStarAI/Orion-14B-Chat
+        "{% for message in messages %}{% if loop.first %}{{ bos_token }}{% endif %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'] + '\\n\\nAssistant: ' + eos_token }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token }}{% endif %}{% endfor %}",
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -45,6 +47,8 @@ int main(void) {
         "system\nYou are a helpful assistant</s>\n<s>user\nHello</s>\n<s>assistant\nHi there</s>\n<s>user\nWho are you</s>\n<s>assistant\n   I am an assistant   </s>\n<s>user\nAnother question</s>\n<s>assistant\n",
         // google/gemma-7b-it
         "<start_of_turn>user\nYou are a helpful assistant\n\nHello<end_of_turn>\n<start_of_turn>model\nHi there<end_of_turn>\n<start_of_turn>user\nWho are you<end_of_turn>\n<start_of_turn>model\nI am an assistant<end_of_turn>\n<start_of_turn>user\nAnother question<end_of_turn>\n<start_of_turn>model\n",
+        // OrionStarAI/Orion-14B-Chat
+        "Human: You are a helpful assistant\n\nHello\n\nAssistant: </s>Hi there</s>Human: Who are you\n\nAssistant: </s>   I am an assistant   </s>Human: Another question\n\nAssistant: </s>",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;

tests/test-json-schema-to-grammar.cpp

@@ -0,0 +1,824 @@
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
+#include <fstream>
+#include <sstream>
+#include <regex>
+
+#include "json-schema-to-grammar.h"
+#include "grammar-parser.h"
+
+static std::string trim(const std::string & source) {
+    std::string s(source);
+    s.erase(0,s.find_first_not_of(" \n\r\t"));
+    s.erase(s.find_last_not_of(" \n\r\t")+1);
+    return std::regex_replace(s, std::regex("(^|\n)[ \t]+"), "$1");
+}
+
+enum TestCaseStatus {
+    SUCCESS,
+    FAILURE
+};
+
+struct TestCase {
+    TestCaseStatus expected_status;
+    std::string name;
+    std::string schema;
+    std::string expected_grammar;
+
+    void _print_failure_header() const {
+        fprintf(stderr, "#\n# Test '%s' failed.\n#\n%s\n", name.c_str(), schema.c_str());
+    }
+    void verify(const std::string & actual_grammar) const {
+        if (trim(actual_grammar) != trim(expected_grammar)) {
+        _print_failure_header();
+        fprintf(stderr, "# EXPECTED:\n%s\n# ACTUAL:\n%s\n", expected_grammar.c_str(), actual_grammar.c_str());
+        assert(false);
+        }
+    }
+    void verify_expectation_parseable() const {
+        try {
+            auto state = grammar_parser::parse(expected_grammar.c_str());
+            if (state.symbol_ids.find("root") == state.symbol_ids.end()) {
+                throw std::runtime_error("Grammar failed to parse:\n" + expected_grammar);
+            }
+        } catch (const std::runtime_error & ex) {
+            _print_failure_header();
+            fprintf(stderr, "# GRAMMAR ERROR: %s\n", ex.what());
+            assert(false);
+        }
+    }
+    void verify_status(TestCaseStatus status) const {
+        if (status != expected_status) {
+            _print_failure_header();
+            fprintf(stderr, "# EXPECTED STATUS: %s\n", expected_status == SUCCESS ? "SUCCESS" : "FAILURE");
+            fprintf(stderr, "# ACTUAL STATUS: %s\n", status == SUCCESS ? "SUCCESS" : "FAILURE");
+            assert(false);
+        }
+    }
+};
+
+static void write(const std::string & file, const std::string & content) {
+    std::ofstream f;
+    f.open(file.c_str());
+    f << content.c_str();
+    f.close();
+}
+
+static std::string read(const std::string & file) {
+    std::ostringstream actuals;
+    actuals << std::ifstream(file.c_str()).rdbuf();
+    return actuals.str();
+}
+
+static void test_all(const std::string & lang, std::function<void(const TestCase &)> runner) {
+    fprintf(stderr, "#\n# Testing JSON schema conversion (%s)\n#\n", lang.c_str());
+    auto test = [&](const TestCase & tc) {
+        fprintf(stderr, "- %s%s\n", tc.name.c_str(), tc.expected_status == FAILURE ? " (failure expected)" : "");
+        runner(tc);
+    };
+
+    test({
+        FAILURE,
+        "unknown type",
+        R"""({
+            "type": "kaboom"
+        })""",
+        ""
+    });
+
+    test({
+        FAILURE,
+        "invalid type type",
+        R"""({
+            "type": 123
+        })""",
+        ""
+    });
+
+    test({
+        SUCCESS,
+        "empty schema (object)",
+        "{}",
+        R"""(
+            array ::= "[" space ( value ("," space value)* )? "]" space
+            boolean ::= ("true" | "false") space
+            null ::= "null" space
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
+            root ::= object
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+            value ::= object | array | string | number | boolean
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "exotic formats",
+        R"""({
+            "items": [
+                { "format": "date" },
+                { "format": "uuid" },
+                { "format": "time" },
+                { "format": "date-time" }
+            ]
+        })""",
+        R"""(
+            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] )
+            date-string ::= "\"" date "\"" space
+            date-time ::= date "T" time
+            date-time-string ::= "\"" date-time "\"" space
+            root ::= "[" space date-string "," space uuid "," space time-string "," space date-time-string "]" space
+            space ::= " "?
+            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] )
+            time-string ::= "\"" time "\"" 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
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string",
+        R"""({
+            "type": "string"
+        })""",
+        R"""(
+            root ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "boolean",
+        R"""({
+            "type": "boolean"
+        })""",
+        R"""(
+            root ::= ("true" | "false") space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "integer",
+        R"""({
+            "type": "integer"
+        })""",
+        R"""(
+            root ::= ("-"? ([0-9] | [1-9] [0-9]*)) space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "string const",
+        R"""({
+            "const": "foo"
+        })""",
+        R"""(
+            root ::= "\"foo\""
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        FAILURE,
+        "non-string const",
+        R"""({
+            "const": 123
+        })""",
+        ""
+    });
+
+    test({
+        FAILURE,
+        "non-string enum",
+        R"""({
+            "enum": [123]
+        })""",
+        ""
+    });
+
+    test({
+        SUCCESS,
+        "tuple1",
+        R"""({
+            "prefixItems": [{ "type": "string" }]
+        })""",
+        R"""(
+            root ::= "[" space string "]" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "tuple2",
+        R"""({
+            "prefixItems": [{ "type": "string" }, { "type": "number" }]
+        })""",
+        R"""(
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "[" space string "," space number "]" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "number",
+        R"""({
+            "type": "number"
+        })""",
+        R"""(
+            root ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "minItems",
+        R"""({
+            "items": {
+                "type": "boolean"
+            },
+            "minItems": 2
+        })""",
+        R"""(
+            boolean ::= ("true" | "false") space
+            root ::= "[" space boolean ( "," space boolean )( "," space boolean )* "]" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "maxItems 1",
+        R"""({
+            "items": {
+                "type": "boolean"
+            },
+            "maxItems": 1
+        })""",
+        R"""(
+            boolean ::= ("true" | "false") space
+            root ::= "[" space ( boolean  )? "]" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "maxItems 2",
+        R"""({
+            "items": {
+                "type": "boolean"
+            },
+            "maxItems": 2
+        })""",
+        R"""(
+            boolean ::= ("true" | "false") space
+            root ::= "[" space ( boolean ( "," space boolean )? )? "]" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "min + maxItems",
+        R"""({
+            "items": {
+                "type": ["number", "integer"]
+            },
+            "minItems": 3,
+            "maxItems": 5
+        })""",
+        R"""(
+            integer ::= ("-"? ([0-9] | [1-9] [0-9]*)) space
+            item ::= number | integer
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "[" space item ( "," space item )( "," space item )( "," space item )?( "," space item )? "]" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "simple regexp",
+        R"""({
+            "type": "string",
+            "pattern": "^abc?d*efg+(hij)?kl$"
+        })""",
+        R"""(
+            root ::= "\"" "ab" "c"? "d"* "ef" "g"+ ("hij")? "kl" "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "regexp escapes",
+        R"""({
+            "type": "string",
+            "pattern": "^\\[\\]\\{\\}\\(\\)\\|\\+\\*\\?$"
+        })""",
+        R"""(
+            root ::= "\"" "[]{}()|+*?" "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "regexp quote",
+        R"""({
+            "type": "string",
+            "pattern": "^\"$"
+        })""",
+        R"""(
+            root ::= "\"" "\"" "\"" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "regexp",
+        R"""({
+            "type": "string",
+            "pattern": "^(\\([0-9]{1,3}\\))?[0-9]{3}-[0-9]{4} and...$"
+        })""",
+        R"""(
+            dot ::= [\U00000000-\x09\x0B\x0C\x0E-\U0010FFFF]
+            root ::= "\"" ("(" root-1 root-1? root-1? ")")? root-1 root-1 root-1 "-" root-1 root-1 root-1 root-1 " and" dot dot dot "\"" space
+            root-1 ::= [0-9]
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "required props",
+        R"""({
+            "type": "object",
+            "properties": {
+                "a": {
+                "type": "string"
+                },
+                "b": {
+                "type": "string"
+                }
+            },
+            "required": [
+                "a",
+                "b"
+            ],
+            "additionalProperties": false,
+            "definitions": {}
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space string
+            b-kv ::= "\"b\"" space ":" space string
+            root ::= "{" space a-kv "," space b-kv "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "1 optional prop",
+        R"""({
+            "properties": {
+                "a": {
+                "type": "string"
+                }
+            },
+            "additionalProperties": false
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space string
+            root ::= "{" space  (a-kv )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "N optional props",
+        R"""({
+            "properties": {
+                "a": {"type": "string"},
+                "b": {"type": "string"},
+                "c": {"type": "string"}
+            },
+            "additionalProperties": false
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space string
+            a-rest ::= ( "," space b-kv )? b-rest
+            b-kv ::= "\"b\"" space ":" space string
+            b-rest ::= ( "," space c-kv )?
+            c-kv ::= "\"c\"" space ":" space string
+            root ::= "{" space  (a-kv a-rest | b-kv b-rest | c-kv )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "required + optional props",
+        R"""({
+            "properties": {
+                "a": {"type": "string"},
+                "b": {"type": "string"},
+                "c": {"type": "string"},
+                "d": {"type": "string"}
+            },
+            "required": ["a", "b"],
+            "additionalProperties": false
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space string
+            b-kv ::= "\"b\"" space ":" space string
+            c-kv ::= "\"c\"" space ":" space string
+            c-rest ::= ( "," space d-kv )?
+            d-kv ::= "\"d\"" space ":" space string
+            root ::= "{" space a-kv "," space b-kv ( "," space ( c-kv c-rest | d-kv ) )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "additional props",
+        R"""({
+            "type": "object",
+            "additionalProperties": {"type": "array", "items": {"type": "number"}}
+        })""",
+        R"""(
+            additional-kv ::= string ":" space additional-value
+            additional-kvs ::= additional-kv ( "," space additional-kv )*
+            additional-value ::= "[" space ( number ( "," space number )* )? "]" space
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "{" space  (additional-kvs )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "additional props (true)",
+        R"""({
+            "type": "object",
+            "additionalProperties": true
+        })""",
+        R"""(
+            array ::= "[" space ( value ("," space value)* )? "]" space
+            boolean ::= ("true" | "false") space
+            null ::= "null" space
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
+            root ::= object
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+            value ::= object | array | string | number | boolean
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "additional props (implicit)",
+        R"""({
+            "type": "object"
+        })""",
+        R"""(
+            array ::= "[" space ( value ("," space value)* )? "]" space
+            boolean ::= ("true" | "false") space
+            null ::= "null" space
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            object ::= "{" space ( string ":" space value ("," space string ":" space value)* )? "}" space
+            root ::= object
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+            value ::= object | array | string | number | boolean
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "empty w/o additional props",
+        R"""({
+            "type": "object",
+            "additionalProperties": false
+        })""",
+        R"""(
+            root ::= "{" space  "}" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "required + additional props",
+        R"""({
+            "type": "object",
+            "properties": {
+                "a": {"type": "number"}
+            },
+            "required": ["a"],
+            "additionalProperties": {"type": "string"}
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space number
+            additional-kv ::= string ":" space string
+            additional-kvs ::= additional-kv ( "," space additional-kv )*
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "{" space a-kv ( "," space ( additional-kvs ) )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "optional + additional props",
+        R"""({
+            "type": "object",
+            "properties": {
+                "a": {"type": "number"}
+            },
+            "additionalProperties": {"type": "number"}
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space number
+            a-rest ::= additional-kvs
+            additional-kv ::= string ":" space number
+            additional-kvs ::= additional-kv ( "," space additional-kv )*
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "{" space  (a-kv a-rest | additional-kvs )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "required + optional + additional props",
+        R"""({
+            "type": "object",
+            "properties": {
+                "a": {"type": "number"},
+                "b": {"type": "number"}
+            },
+            "required": ["a"],
+            "additionalProperties": {"type": "number"}
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space number
+            additional-kv ::= string ":" space number
+            additional-kvs ::= additional-kv ( "," space additional-kv )*
+            b-kv ::= "\"b\"" space ":" space number
+            b-rest ::= additional-kvs
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "{" space a-kv ( "," space ( b-kv b-rest | additional-kvs ) )? "}" space
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "top-level $ref",
+        R"""({
+            "$ref": "#/definitions/MyType",
+            "definitions": {
+                "MyType": {
+                "type": "object",
+                "properties": {
+                    "a": {
+                    "type": "string"
+                    }
+                },
+                "required": [
+                    "a"
+                ],
+                "additionalProperties": false
+                }
+            }
+        })""",
+        R"""(
+            MyType ::= "{" space MyType-a-kv "}" space
+            MyType-a-kv ::= "\"a\"" space ":" space string
+            root ::= MyType
+            space ::= " "?
+            string ::=  "\"" (
+                    [^"\\] |
+                    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
+                    )* "\"" space
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "anyOf",
+        R"""({
+            "anyOf": [
+                {"$ref": "#/definitions/foo"},
+                {"$ref": "#/definitions/bar"}
+            ],
+            "definitions": {
+                "foo": {
+                "properties": {"a": {"type": "number"}}
+                },
+                "bar": {
+                "properties": {"b": {"type": "number"}}
+                }
+            },
+            "type": "object"
+        })""",
+        R"""(
+            alternative-0 ::= foo
+            alternative-1 ::= bar
+            bar ::= "{" space  (bar-b-kv )? "}" space
+            bar-b-kv ::= "\"b\"" space ":" space number
+            foo ::= "{" space  (foo-a-kv )? "}" space
+            foo-a-kv ::= "\"a\"" space ":" space number
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= alternative-0 | alternative-1
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "mix of allOf, anyOf and $ref (similar to https://json.schemastore.org/tsconfig.json)",
+        R"""({
+            "allOf": [
+                {"$ref": "#/definitions/foo"},
+                {"$ref": "#/definitions/bar"},
+                {
+                "anyOf": [
+                    {"$ref": "#/definitions/baz"},
+                    {"$ref": "#/definitions/bam"}
+                ]
+                }
+            ],
+            "definitions": {
+                "foo": {
+                "properties": {"a": {"type": "number"}}
+                },
+                "bar": {
+                "properties": {"b": {"type": "number"}}
+                },
+                "bam": {
+                "properties": {"c": {"type": "number"}}
+                },
+                "baz": {
+                "properties": {"d": {"type": "number"}}
+                }
+            },
+            "type": "object"
+        })""",
+        R"""(
+            a-kv ::= "\"a\"" space ":" space number
+            b-kv ::= "\"b\"" space ":" space number
+            c-kv ::= "\"c\"" space ":" space number
+            d-kv ::= "\"d\"" space ":" space number
+            d-rest ::= ( "," space c-kv )?
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            root ::= "{" space a-kv "," space b-kv ( "," space ( d-kv d-rest | c-kv ) )? "}" space
+            space ::= " "?
+        )"""
+    });
+
+    test({
+        SUCCESS,
+        "conflicting names",
+        R"""({
+            "type": "object",
+            "properties": {
+                "number": {
+                "type": "object",
+                "properties": {
+                    "number": {
+                    "type": "object",
+                    "properties": {
+                        "root": {
+                        "type": "number"
+                        }
+                    },
+                    "required": [
+                        "root"
+                    ],
+                    "additionalProperties": false
+                    }
+                },
+                "required": [
+                    "number"
+                ],
+                "additionalProperties": false
+                }
+            },
+            "required": [
+                "number"
+            ],
+            "additionalProperties": false,
+            "definitions": {}
+        })""",
+        R"""(
+            number ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space
+            number- ::= "{" space number-number-kv "}" space
+            number-kv ::= "\"number\"" space ":" space number-
+            number-number ::= "{" space number-number-root-kv "}" space
+            number-number-kv ::= "\"number\"" space ":" space number-number
+            number-number-root-kv ::= "\"root\"" space ":" space number
+            root ::= "{" space number-kv "}" space
+            space ::= " "?
+        )"""
+    });
+}
+
+int main() {
+    test_all("C++", [](const TestCase & tc) {
+        try {
+            tc.verify(json_schema_to_grammar(nlohmann::json::parse(tc.schema)));
+            tc.verify_status(SUCCESS);
+        } catch (const std::runtime_error & ex) {
+            fprintf(stderr, "Error: %s\n", ex.what());
+            tc.verify_status(FAILURE);
+        }
+    });
+    //test_all("Python", [](const TestCase & tc) {
+    //    write("test-json-schema-input.tmp", tc.schema);
+    //    tc.verify_status(std::system(
+    //        "python ./examples/json-schema-to-grammar.py test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE);
+    //    tc.verify(read("test-grammar-output.tmp"));
+    //});
+    //test_all("JavaScript", [](const TestCase & tc) {
+    //    write("test-json-schema-input.tmp", tc.schema);
+    //    tc.verify_status(std::system(
+    //        "node ./tests/run-json-schema-to-grammar.mjs test-json-schema-input.tmp > test-grammar-output.tmp") == 0 ? SUCCESS : FAILURE);
+    //    tc.verify(read("test-grammar-output.tmp"));
+    //});
+
+    test_all("Check Expectations Validity", [](const TestCase & tc) {
+        if (tc.expected_status == SUCCESS) {
+            tc.verify_expectation_parseable();
+        }
+    });
+}