Update

.github/workflows/bench.yml

@@ -32,7 +32,7 @@ on:
     -  cron: '04 2 * * *'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref || github.run_id }}-${{ github.event.inputs.sha }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}-${{ github.event.inputs.sha }}
   cancel-in-progress: true
 
 jobs:

.github/workflows/build.yml

@@ -32,8 +32,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -90,8 +88,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -210,8 +206,6 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -244,33 +238,6 @@ jobs:
           ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
-      - 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-ubuntu-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
-          name: llama-bin-ubuntu-x64.zip
-
 #  ubuntu-latest-cmake-sanitizer:
 #    runs-on: ubuntu-latest
 #

.github/workflows/docker.yml

@@ -91,12 +91,6 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
-      - name: Downcase github.repository_owner
-        run: |
-          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -104,7 +98,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -113,5 +107,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/server.yml

@@ -23,7 +23,7 @@ on:
     -  cron: '2 4 * * *'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref || github.run_id }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:

.gitignore

@@ -34,7 +34,6 @@ lcov-report/
 gcovr-report/
 
 build*
-!build.zig
 cmake-build-*
 out/
 tmp/
@@ -101,9 +100,6 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
-examples/server/*.html.hpp
-examples/server/*.js.hpp
-examples/server/*.mjs.hpp
 
 poetry.lock
 poetry.toml

CMakeLists.txt

@@ -43,12 +43,6 @@ else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()
 
-if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
-    set(LLAMA_LLAMAFILE_DEFAULT OFF)
-else()
-    set(LLAMA_LLAMAFILE_DEFAULT ON)
-endif()
-
 # general
 option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
@@ -94,7 +88,6 @@ endif()
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
-option(LLAMA_LLAMAFILE                       "llama: use llamafile SGEMM"                       ${LLAMA_LLAMAFILE_DEFAULT})
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
 option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
@@ -293,7 +286,6 @@ if (LLAMA_METAL)
         ${METALKIT_FRAMEWORK}
         )
 endif()
-
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -376,13 +368,6 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_LLAMAFILE)
-    add_compile_definitions(GGML_USE_LLAMAFILE)
-
-    set(GGML_HEADERS_LLAMAFILE sgemm.h)
-    set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
-endif()
-
 if (LLAMA_QKK_64)
     add_compile_definitions(GGML_QKK_64)
 endif()
@@ -1166,16 +1151,15 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
-            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
+            ${GGML_SOURCES_CUDA}    ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL}  ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL}   ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}     ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

Makefile

@@ -384,11 +384,6 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
-ifndef LLAMA_NO_LLAMAFILE
-	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJS        += sgemm.o
-endif
-
 ifdef LLAMA_BLIS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
@@ -485,9 +480,11 @@ ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/com
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(NVCC_COMPILE)
+
 endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
+
 	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
 	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
@@ -606,11 +603,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_LLAMAFILE
-sgemm.o: sgemm.cpp sgemm.h ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif
-
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -695,8 +687,8 @@ OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -764,7 +756,7 @@ batched: examples/batched/batched.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)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -796,19 +788,10 @@ 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 common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.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)
 
-# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-examples/server/%.hpp: examples/server/public/% Makefile
-	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
-		echo "unsigned char $${NAME}[] = {" && \
-		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
-		echo "};" && \
-		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
-	) > $@
-
 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)

Package.swift

@@ -2,45 +2,6 @@
 
 import PackageDescription
 
-var sources = [
-    "ggml.c",
-    "sgemm.cpp",
-    "llama.cpp",
-    "unicode.cpp",
-    "unicode-data.cpp",
-    "ggml-alloc.c",
-    "ggml-backend.c",
-    "ggml-quants.c",
-]
-
-var resources: [Resource] = []
-var linkerSettings: [LinkerSetting] = []
-var cSettings: [CSetting] =  [
-    .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-    .unsafeFlags(["-fno-objc-arc"]),
-    // NOTE: NEW_LAPACK will required iOS version 16.4+
-    // We should consider add this in the future when we drop support for iOS 14
-    // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-    // .define("ACCELERATE_NEW_LAPACK"),
-    // .define("ACCELERATE_LAPACK_ILP64")
-]
-
-#if canImport(Darwin)
-sources.append("ggml-metal.m")
-resources.append(.process("ggml-metal.metal"))
-linkerSettings.append(.linkedFramework("Accelerate"))
-cSettings.append(
-    contentsOf: [
-        .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL")
-    ]
-)
-#endif
-
-#if os(Linux)
-    cSettings.append(.define("_GNU_SOURCE"))
-#endif
-
 let package = Package(
     name: "llama",
     platforms: [
@@ -67,11 +28,34 @@ let package = Package(
                "ggml-cuda.h",
                "Makefile"
             ],
-            sources: sources,
-            resources: resources,
+            sources: [
+                "ggml.c",
+                "llama.cpp",
+                "unicode.cpp",
+                "unicode-data.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
+            ],
             publicHeadersPath: "spm-headers",
-            cSettings: cSettings,
-            linkerSettings: linkerSettings
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL"),
+                // NOTE: NEW_LAPACK will required iOS version 16.4+
+                // We should consider add this in the future when we drop support for iOS 14
+                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+                // .define("ACCELERATE_NEW_LAPACK"),
+                // .define("ACCELERATE_LAPACK_ILP64")
+            ],
+            linkerSettings: [
+                .linkedFramework("Accelerate")
+            ]
         )
     ],
     cxxLanguageStandard: .cxx11

README-sycl.md

@@ -68,7 +68,7 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
-| Intel Data Center Max Series  | Support | Max 1550, 1100                        |
+| Intel Data Center Max Series  | Support | Max 1550                              |
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M                         |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
@@ -84,7 +84,8 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 - **Execution Unit (EU)**
   - If the iGPU has less than 80 EUs, the inference speed will likely be too slow for practical use.
 
-### Other Vendor GPU
+### Nvidia GPU
+The BLAS acceleration on Nvidia GPU through oneAPI can be obtained using the Nvidia plugins for oneAPI and the cuBLAS backend of the upstream oneMKL library. Details and instructions on how to setup the runtime and library can be found in [this section](#i-setup-environment)
 
 **Verified devices**
 
@@ -93,9 +94,14 @@ It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS,
 | Ampere Series            | Support | A100, A4000    |
 | Ampere Series *(Mobile)* | Support | RTX 40 Series  |
 
+*Notes:*
+  - Support for Nvidia targets through oneAPI is currently limited to Linux platforms.
+
+  - Please make sure the native oneAPI MKL *(dedicated to intel CPUs and GPUs)* is not "visible" at this stage to properly setup and use the built-from-source oneMKL with cuBLAS backend in llama.cpp for Nvidia GPUs.
+
+
 ## Docker
 The docker build option is currently limited to *intel GPU* targets.
-
 ### Build image
 ```sh
 # Using FP16
@@ -162,10 +168,29 @@ Platform #0: Intel(R) OpenCL HD Graphics
 - **Nvidia GPU**
 
 In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
+Installation can be verified by running the following:
+```sh
+nvidia-smi
+```
+Please make sure at least one CUDA device is available, which can be displayed like this *(here an A100-40GB Nvidia GPU)*:
+```
++---------------------------------------------------------------------------------------+
+| NVIDIA-SMI 535.54.03              Driver Version: 535.54.03    CUDA Version: 12.2     |
+|-----------------------------------------+----------------------+----------------------+
+| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
+|                                         |                      |               MIG M. |
+|=========================================+======================+======================|
+|   0  NVIDIA A100-PCIE-40GB          On  | 00000000:8D:00.0 Off |                    0 |
+| N/A   36C    P0              57W / 250W |      4MiB / 40960MiB |      0%      Default |
+|                                         |                      |             Disabled |
++-----------------------------------------+----------------------+----------------------+
+```
+
 
 2. **Install Intel® oneAPI Base toolkit**
 
-- **For Intel GPU**
+- **Base installation**
 
 The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
@@ -177,10 +202,10 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 - **Adding support to Nvidia GPUs**
 
-**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
+**oneAPI**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
 
 
-**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
+**oneMKL**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
 
 ```sh
 git clone https://github.com/oneapi-src/oneMKL
@@ -212,7 +237,7 @@ When targeting an intel GPU, the user should expect one or more level-zero devic
 
 - **Nvidia GPU**
 
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+Similarly, user targetting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
 ```
 [opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
 [opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
@@ -229,14 +254,12 @@ source /opt/intel/oneapi/setvars.sh
 # Build LLAMA with MKL BLAS acceleration for intel GPU
 mkdir -p build && cd build
 
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Option 1: Use FP16 for better performance in long-prompt  inference
+cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+# Or without "--build", run "make" next
 
-# Option 2: Use FP16
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-#build all binary
-cmake --build . --config Release -j -v
+# Option 2: Use FP32 by default
+cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 ```
 
 #### Nvidia GPU
@@ -250,15 +273,11 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
 mkdir -p build && cd build
 
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# Option 2: Use FP16
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-#build all binary
-cmake --build . --config Release -j -v
+# Option 1: Use FP16 for better performance in long-prompt  inference
+cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
 
+# Option 2: Use FP32 by default
+cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 ```
 
 ### III. Run the inference
@@ -338,6 +357,7 @@ Otherwise, you can run the script:
 
 *Notes:*
 
+- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `/bin/main` if faced with the issue.
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
 ```sh
@@ -416,13 +436,9 @@ mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
-
-# Option 2: Or FP16
 cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
-make -j
+make
 ```
 
 Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
@@ -509,6 +525,7 @@ Otherwise, run the following wrapper script:
 
 Note:
 
+- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `main.exe` if faced with the issue.
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
 ```sh
@@ -540,6 +557,12 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Known Issues
 
+- Hanging during startup
+
+  llama.cpp uses *mmap* as the default mode for reading the model file and copying it to the GPU. In some systems, `memcpy` might behave abnormally and therefore hang.
+
+  - **Solution**: add `--no-mmap` or `--mmap 0` flag to the `main` executable.
+
 - `Split-mode:[row]` is not supported.
 
 ## Q&A
@@ -551,7 +574,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 - General compiler error:
 
-  - Remove **build** folder or try a clean-build.
+  - Remove build folder or try a clean-build.
 
 - I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
 

README.md

@@ -10,7 +10,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
-- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
 - [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
 - [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
 - [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
@@ -95,8 +94,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] LLaMA 2 🦙🦙
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
-- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
-- [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
+- [X] Falcon
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
@@ -123,7 +121,6 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
-- [x] [OLMo](https://allenai.org/olmo)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
@@ -191,8 +188,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [MindMac](https://mindmac.app) (proprietary)
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
-- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
-
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---
@@ -550,7 +545,7 @@ Building the program with BLAS support may lead to some performance improvements
   OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
 
   You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
-    - For Ubuntu, Debian, and Fedora the packages `opencl-headers`, `ocl-icd` may be needed.
+    - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
 
     - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
 
@@ -575,12 +570,6 @@ Building the program with BLAS support may lead to some performance improvements
 
   Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
 
-  Linux packaging:
-  Fedora Linux:
-  ```bash
-  sudo dnf install clblast
-  ```
-
   Alternatively, they may be built from source.
 
   - <details>
@@ -1117,9 +1106,7 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
-
-![matmul](media/matmul.png)
+- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
 
 ### Docs
 

build.zig

@@ -112,7 +112,6 @@ pub fn build(b: *std.build.Builder) !void {
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
     const ggml = make.obj("ggml", "ggml.c");
-    const sgemm = make.obj("sgemm", "sgemm.cpp");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
@@ -129,44 +128,15 @@ pub fn build(b: *std.build.Builder) !void {
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, 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, unicode_data, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }
-
-    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
-    for (server_assets) |asset| {
-        const input_path = b.fmt("examples/server/public/{s}", .{asset});
-        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
-
-        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
-
-        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
-        defer b.allocator.free(input);
-
-        var buf = std.ArrayList(u8).init(b.allocator);
-        defer buf.deinit();
-
-        for (input) |byte| {
-            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
-        }
-
-        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
-        defer b.allocator.free(name);
-        std.mem.replaceScalar(u8, name, '.', '_');
-
-        try std.fs.cwd().writeFile(output_path, b.fmt(
-            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
-            .{ name, buf.items, name, input.len },
-        ));
-
-        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
-    }
 }

ci/run.sh

@@ -153,54 +153,6 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
-# test_scripts_debug
-
-function gg_run_test_scripts_debug {
-    cd ${SRC}
-
-    set -e
-
-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-
-    set +e
-}
-
-function gg_sum_test_scripts_debug {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test scripts in debug mode\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
-    gg_printf '```\n'
-    gg_printf '\n'
-}
-
-# test_scripts_release
-
-function gg_run_test_scripts_release {
-    cd ${SRC}
-
-    set -e
-
-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-
-    set +e
-}
-
-function gg_sum_test_scripts_release {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test scripts in release mode\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
-    gg_printf '```\n'
-    gg_printf '\n'
-}
-
 function gg_get_model {
     local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
     local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
@@ -690,9 +642,6 @@ test $ret -eq 0 && gg_run ctest_release
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
 
-    test $ret -eq 0 && gg_run test_scripts_debug
-    test $ret -eq 0 && gg_run test_scripts_release
-
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/CMakeLists.txt

@@ -47,6 +47,9 @@ 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)
 
 add_library(${TARGET} STATIC
@@ -60,7 +63,6 @@ add_library(${TARGET} STATIC
     grammar-parser.h
     grammar-parser.cpp
     json.hpp
-    json-schema-to-grammar.cpp
     train.h
     train.cpp
     ngram-cache.h

common/common.cpp

@@ -1,6 +1,4 @@
 #include "common.h"
-#include "json.hpp"
-#include "json-schema-to-grammar.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -70,8 +68,6 @@
 #define LLAMA_CURL_MAX_HEADER_LENGTH 256
 #endif // LLAMA_USE_CURL
 
-using json = nlohmann::ordered_json;
-
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -108,79 +104,6 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
-#include <pthread.h>
-
-static void cpuid(unsigned leaf, unsigned subleaf,
-                  unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
-    __asm__("movq\t%%rbx,%%rsi\n\t"
-            "cpuid\n\t"
-            "xchgq\t%%rbx,%%rsi"
-            : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
-            : "0"(leaf), "2"(subleaf));
-}
-
-static int pin_cpu(int cpu) {
-    cpu_set_t mask;
-    CPU_ZERO(&mask);
-    CPU_SET(cpu, &mask);
-    return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
-}
-
-static bool is_hybrid_cpu(void) {
-    unsigned eax, ebx, ecx, edx;
-    cpuid(7, 0, &eax, &ebx, &ecx, &edx);
-    return !!(edx & (1u << 15));
-}
-
-static bool is_running_on_efficiency_core(void) {
-    unsigned eax, ebx, ecx, edx;
-    cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
-    int intel_atom = 0x20;
-    int core_type = (eax & 0xff000000u) >> 24;
-    return core_type == intel_atom;
-}
-
-static int count_math_cpus(int cpu_count) {
-    int result = 0;
-    for (int cpu = 0; cpu < cpu_count; ++cpu) {
-        if (pin_cpu(cpu)) {
-            return -1;
-        }
-        if (is_running_on_efficiency_core()) {
-            continue; // efficiency cores harm lockstep threading
-        }
-        ++cpu; // hyperthreading isn't useful for linear algebra
-        ++result;
-    }
-    return result;
-}
-
-#endif // __x86_64__ && __linux__
-
-/**
- * Returns number of CPUs on system that are useful for math.
- */
-int get_math_cpu_count() {
-#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
-    int cpu_count = sysconf(_SC_NPROCESSORS_ONLN);
-    if (cpu_count < 1) {
-        return get_num_physical_cores();
-    }
-    if (is_hybrid_cpu()) {
-        cpu_set_t affinity;
-        if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
-            int result = count_math_cpus(cpu_count);
-            pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
-            if (result > 0) {
-                return result;
-            }
-        }
-    }
-#endif
-    return get_num_physical_cores();
-}
-
 void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -242,9 +165,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
             invalid_param = true;
             return true;
         }
-        // This is temporary, in the future the samplign state will be moved fully to llama_sampling_context.
         params.seed = std::stoul(argv[i]);
-        sparams.seed = std::stoul(argv[i]);
         return true;
     }
     if (arg == "-t" || arg == "--threads") {
@@ -1227,14 +1148,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         );
         return true;
     }
-    if (arg == "-j" || arg == "--json-schema") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
-        return true;
-    }
     if (arg == "--override-kv") {
         if (++i >= argc) {
             invalid_param = true;
@@ -1440,9 +1353,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
     printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
     printf("  --grammar-file FNAME  file to read grammar from\n");
-    printf("  -j SCHEMA, --json-schema SCHEMA\n");
-    printf("                        JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object.\n");
-    printf("                        For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead\n");
     printf("  --cfg-negative-prompt PROMPT\n");
     printf("                        negative prompt to use for guidance. (default: empty)\n");
     printf("  --cfg-negative-prompt-file FNAME\n");
@@ -2328,12 +2238,12 @@ std::vector<llama_token> llama_tokenize(
     return result;
 }
 
-std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);

common/common.h

@@ -39,7 +39,6 @@ extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
-int get_math_cpu_count();
 int32_t get_num_physical_cores();
 
 //
@@ -49,7 +48,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = get_math_cpu_count();
+    int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
@@ -86,8 +85,8 @@ struct gpt_params {
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
-    enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-    enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
+    llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -237,12 +236,11 @@ std::vector<llama_token> llama_tokenize(
                         bool   add_special,
                         bool   parse_special = false);
 
-// tokenizes a token into a piece, optionally renders special/control tokens
+// tokenizes a token into a piece
 // should work similar to Python's `tokenizer.id_to_piece`
 std::string llama_token_to_piece(
         const struct llama_context * ctx,
-                       llama_token   token,
-                       bool          special = true);
+                       llama_token   token);
 
 // TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
 //       that takes into account the tokenizer type and decides how to handle the leading space

common/json-schema-to-grammar.cpp

@@ -11,101 +11,35 @@
 
 using json = nlohmann::ordered_json;
 
-template <typename Iterator>
-static std::string join(Iterator begin, Iterator end, const std::string & separator);
-
-static std::string repeat(const std::string & str, size_t n);
-
-static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
-    if (separator_rule.empty()) {
-        if (min_items == 0 && max_items == 1) {
-            return item_rule + "?";
-        } else if (min_items == 1 && max_items == std::numeric_limits<int>::max()) {
-            return item_rule + "+";
-        }
-    }
-
-    std::string result;
-    if (min_items > 0) {
-        if (item_rule_is_literal && separator_rule.empty()) {
-            result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
-        } else {
-            std::vector<std::string> items(min_items, item_rule);
-            result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
-        }
-    }
-
-    std::function<std::string(int, bool)> opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
-        auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
-
-        if (up_to_n == 0) {
-            return "";
-        } else if (up_to_n == 1) {
-            return "(" + content + ")?";
-        } else if (!separator_rule.empty() && !prefix_with_sep) {
-            return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
-        } else {
-            std::string res = repeat("(" + content + " ", up_to_n);
-            // strip trailing space
-            res = res.substr(0, res.length() - 1);
-            res += repeat(")?", up_to_n);
-            return res;
-        }
-    };
-
-    if (min_items > 0 && max_items != min_items) {
-        result += " ";
-    }
-
-    if (max_items != std::numeric_limits<int>::max()) {
-        result += opt_repetitions(max_items - min_items, min_items > 0);
-    } else {
-        std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
-        if (min_items == 0 && !separator_rule.empty()) {
-            result = "(" + item_rule + " " + item_operator + "*)?";
-        } else {
-            result += item_operator + "*";
-        }
-    }
-
-    return result;
-}
-
 const std::string SPACE_RULE = "\" \"?";
 
-struct BuiltinRule {
-    std::string content;
-    std::vector<std::string> deps;
+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"};
 
-const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
-
-std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
-    {"boolean", {"(\"true\" | \"false\") space", {}}},
-    {"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
-    {"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
-    {"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
-    {"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
-    {"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
-    {"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
-    {"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
-    {"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", {}}},
-    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
-    {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
-    {"null", {"\"null\" space", {}}},
-};
-
-std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_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", "time"}}},
-    {"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
-    {"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
-    {"date-time-string", {"\"\\\"\" date-time \"\\\"\" space", {"date-time"}}}
+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) {
@@ -113,7 +47,7 @@ static bool is_reserved_name(const std::string & name) {
     if (RESERVED_NAMES.empty()) {
         RESERVED_NAMES.insert("root");
         for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : STRING_FORMAT_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();
 }
@@ -258,7 +192,7 @@ private:
                 if (_dotall) {
                     rule = "[\\U00000000-\\U0010FFFF]";
                 } else {
-                    rule = "[^\\x0A\\x0D]";
+                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
                 }
                 return _add_rule("dot", rule);
             };
@@ -374,21 +308,47 @@ private:
                     auto &sub = last.first;
                     auto sub_is_literal = last.second;
 
-                    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);
+                    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;
                         }
-                        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;
                     }
-                    seq.back().first = build_repetition(
-                        sub_is_literal ? "\"" + sub + "\"" : sub,
-                        min_times,
-                        max_times,
-                        "",
-                        sub_is_literal
-                    );
-                    seq.back().second = false;
                 } else {
                     std::string literal;
                     auto is_non_literal = [&](char c) {
@@ -464,7 +424,7 @@ private:
         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_primitive("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            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("*");
         }
@@ -526,25 +486,6 @@ private:
         return rule;
     }
 
-    std::string _add_primitive(const std::string & name, const BuiltinRule & rule) {
-        auto n = _add_rule(name, rule.content);
-        for (const auto & dep : rule.deps) {
-            BuiltinRule dep_rule;
-            auto it = PRIMITIVE_RULES.find(dep);
-            if (it == PRIMITIVE_RULES.end()) {
-                it = STRING_FORMAT_RULES.find(dep);
-                if (it == STRING_FORMAT_RULES.end()) {
-                    _errors.push_back("Rule " + dep + " not known");
-                    continue;
-                }
-            }
-            if (_rules.find(dep) == _rules.end()) {
-                _add_primitive(dep, it->second);
-            }
-        }
-        return n;
-    }
-
 public:
     SchemaConverter(
         const std::function<json(const std::string &)> & fetch_json,
@@ -706,33 +647,49 @@ public:
                 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>() : std::numeric_limits<int>::max();
-
-                return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " \"]\" space");
+                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_primitive(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
-        } else if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
-            auto prim_name = schema_format + "-string";
-            return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
-        } else if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
-            std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
-            int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
-            int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
-            return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
+            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") {
-            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("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_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+            return _add_rule(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
         }
     }
 

common/sampling.cpp

@@ -1,6 +1,4 @@
-#define LLAMA_API_INTERNAL
 #include "sampling.h"
-#include <random>
 
 struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
     struct llama_sampling_context * result = new llama_sampling_context();
@@ -35,8 +33,6 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
-    llama_sampling_set_rng_seed(result, params.seed);
-
     return result;
 }
 
@@ -66,13 +62,6 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
     ctx->cur.clear();
 }
 
-void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
-    if (seed == LLAMA_DEFAULT_SEED) {
-        seed = time(NULL);
-    }
-    ctx->rng.seed(seed);
-}
-
 void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
     if (dst->grammar) {
         llama_grammar_free(dst->grammar);
@@ -214,7 +203,7 @@ static llama_token llama_sampling_sample_impl(
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 
-            id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
+            id = llama_sample_token(ctx_main, &cur_p);
 
             //{
             //    const int n_top = 10;

common/sampling.h

@@ -4,10 +4,9 @@
 
 #include "grammar-parser.h"
 
-#include <random>
 #include <string>
-#include <unordered_map>
 #include <vector>
+#include <unordered_map>
 
 // sampler types
 enum class llama_sampler_type : char {
@@ -21,26 +20,25 @@ enum class llama_sampler_type : char {
 
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t     n_prev                = 64;                 // number of previous tokens to remember
-    int32_t     n_probs               = 0;                  // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t     min_keep              = 0;                  // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t     top_k                 = 40;                 // <= 0 to use vocab size
-    float       top_p                 = 0.95f;              // 1.0 = disabled
-    float       min_p                 = 0.05f;              // 0.0 = disabled
-    float       tfs_z                 = 1.00f;              // 1.0 = disabled
-    float       typical_p             = 1.00f;              // 1.0 = disabled
-    float       temp                  = 0.80f;              // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    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        = 1.00f;              // 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           = false;              // consider newlines as a repeatable token
-    uint32_t    seed                  = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
+    int32_t     n_prev                = 64;       // number of previous tokens to remember
+    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t     top_k                 = 40;       // <= 0 to use vocab size
+    float       top_p                 = 0.95f;    // 1.0 = disabled
+    float       min_p                 = 0.05f;    // 0.0 = disabled
+    float       tfs_z                 = 1.00f;    // 1.0 = disabled
+    float       typical_p             = 1.00f;    // 1.0 = disabled
+    float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    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        = 1.00f;    // 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           = false;     // consider newlines as a repeatable token
 
     std::vector<llama_sampler_type> samplers_sequence = {
         llama_sampler_type::TOP_K,
@@ -81,8 +79,6 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
-
-    std::mt19937 rng;
 };
 
 #include "common.h"
@@ -97,9 +93,6 @@ void llama_sampling_free(struct llama_sampling_context * ctx);
 // - reset grammar
 void llama_sampling_reset(llama_sampling_context * ctx);
 
-// Set the sampler seed
-void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
-
 // Copy the sampler context
 void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
 

convert-hf-to-gguf.py

@@ -43,18 +43,17 @@ AnyModel = TypeVar("AnyModel", bound="type[Model]")
 class Model(ABC):
     _model_classes: dict[str, type[Model]] = {}
 
-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool):
+    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
         self.fname_out = fname_out
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
-        self.use_temp_file = use_temp_file
         self.is_safetensors = self._is_model_safetensors()
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
+        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
     @property
@@ -363,16 +362,6 @@ class Model(ABC):
                         scores.append(-1000.0)
                         toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
 
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            print(
-                f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]"
-            )
-            for i in range(1, pad_count + 1):
-                tokens.append(f"[PAD{i}]")
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
-
         assert len(tokens) == vocab_size
 
         self.gguf_writer.add_tokenizer_model("llama")
@@ -1217,91 +1206,9 @@ class StableLMModel(Model):
         rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
         self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
-        self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_head = self.hparams.get("num_attention_heads")
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        q_norms = dict()
-        k_norms = dict()
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-            n_dims = len(data.shape)
-            if name.find("q_layernorm.norms") != -1:
-                q_norms[name] = data
-                if len(q_norms) >= (block_count * n_head):
-                    self._stack_qk_norm(block_count, name, tensor_map, n_head, q_norms, n_dims, layer_name="q_layernorm")
-                continue
-            if name.find("k_layernorm.norms") != -1:
-                k_norms[name] = data
-                if len(k_norms) >= (block_count * n_kv_head):
-                    self._stack_qk_norm(block_count, name, tensor_map, n_kv_head, k_norms, n_dims, layer_name="k_layernorm")
-                continue
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-    def _stack_qk_norm(self, block_count, name, tensor_map, n_head, norms, n_dims, layer_name="q_layernorm"):
-        for bid in range(block_count):
-            datas = []
-            for xid in range(n_head):
-                ename = f"model.layers.{bid}.self_attn.{layer_name}.norms.{xid}.weight"
-                datas.append(norms[ename])
-                del norms[ename]
-            data = np.stack(datas, axis=0)
-            data_dtype = data.dtype
-            merged_name = f"model.layers.{bid}.self_attn.{layer_name}.weight"
-            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
 
 @Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
 class LlamaModel(Model):
@@ -1311,23 +1218,7 @@ class LlamaModel(Model):
         try:
             self. _set_vocab_sentencepiece()
         except FileNotFoundError:
-            try:
-                self._set_vocab_llama_hf()
-            except (FileNotFoundError, TypeError):
-                # Llama 3
-                self._set_vocab_gpt2()
-
-        # Apply to CodeLlama only (and ignore for Llama 3 with a vocab size of 128256)
-        if self.hparams.get("vocab_size", 32000) == 32016:
-            special_vocab = gguf.SpecialVocab(
-                self.dir_model, load_merges=False,
-                special_token_types = ['prefix', 'suffix', 'middle', 'eot']
-            )
-            special_vocab._set_special_token("prefix", 32007)
-            special_vocab._set_special_token("suffix", 32008)
-            special_vocab._set_special_token("middle", 32009)
-            special_vocab._set_special_token("eot",    32010)
-            special_vocab.add_to_gguf(self.gguf_writer)
+            self._set_vocab_llama_hf()
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -1536,102 +1427,6 @@ class GrokModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("DbrxForCausalLM")
-class DbrxModel(Model):
-    model_arch = gguf.MODEL_ARCH.DBRX
-
-    def set_gguf_parameters(self):
-        ffn_config = self.hparams["ffn_config"]
-        attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_name(self.hparams["model_type"])
-        self.gguf_writer.add_block_count(self.hparams["n_layers"])
-
-        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
-        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
-        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
-
-        self.gguf_writer.add_head_count(self.hparams["n_heads"])
-        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
-
-        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
-
-        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
-        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
-
-        self.gguf_writer.add_layer_norm_eps(1e-5)
-
-        self.gguf_writer.add_file_type(self.ftype)
-        print(f"gguf: file type = {self.ftype}")
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers")
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        for name, data_torch in self.get_tensors():
-            n_expert = self.hparams["ffn_config"]["moe_num_experts"]
-            n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
-            n_embd = self.hparams["d_model"]
-
-            # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
-            # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
-            # But llama.cpp moe graph works differently
-            # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
-            # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
-            exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
-                                "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
-                                "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
-            experts = False
-            for exp_tensor_name in exp_tensor_names.keys():
-                if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
-                    experts = True
-                    data_torch = data_torch.view(n_expert, n_ff, n_embd)
-                    if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
-                        data_torch = data_torch.permute(*permute_tensor)
-                    break
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            # In MoE models the ffn tensors are typically most of the model weights,
-            # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
-            # Every other model has the weight names ending in .weight,
-            # let's assume that is the convention which is not the case for dbrx:
-            # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
-            new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # Most of the codebase that takes in 1D tensors only handles F32 tensors
-            # and most of the outputs tensors are F32.
-            if data_dtype != np.float32 and n_dims == 1:
-                print(f"Can not map tensor {name!r}: all 1D tensors must be F32")
-                sys.exit()
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
 @Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
     model_arch = gguf.MODEL_ARCH.MINICPM
@@ -1799,111 +1594,6 @@ class QwenModel(Model):
 class Qwen2Model(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
-    def set_vocab(self):
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-
-@Model.register("Qwen2MoeForCausalLM")
-class Qwen2MoeModel(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN2MOE
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        if (n_experts := self.hparams.get("num_experts")) is not None:
-            self.gguf_writer.add_expert_count(n_experts)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_experts = self.hparams.get("num_experts")
-        experts = dict()
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.squeeze().numpy()
-
-            # process the experts separately
-            if name.find("experts") != -1:
-                experts[name] = data
-                if len(experts) >= n_experts * 3:
-                    # merge the experts into a single 3d tensor
-                    for bid in range(block_count):
-                        for w_name in ["down_proj", "gate_proj", "up_proj"]:
-                            full = True
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                if ename not in experts:
-                                    full = False
-                                    break
-                            if not full:
-                                continue
-
-                            datas = []
-                            for xid in range(n_experts):
-                                ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
-                                datas.append(experts[ename])
-                                del experts[ename]
-
-                            data = np.stack(datas, axis=0)
-                            data_dtype = data.dtype
-
-                            if self.ftype == 0 and data_dtype == np.float16:
-                                data = data.astype(np.float32)
-
-                            if self.ftype == 1 and data_dtype == np.float32:
-                                data = data.astype(np.float16)
-
-                            merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-
-                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
-                            if new_name is None:
-                                print(f"Can not map tensor {name!r}")
-                                sys.exit()
-
-                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
-
-                            self.gguf_writer.add_tensor(new_name, data)
-                continue
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-        if len(experts) > 0:
-            raise ValueError(f"Unprocessed experts: {experts.keys()}")
-
 
 @Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
@@ -1995,91 +1685,6 @@ class Phi2Model(Model):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@Model.register("Phi3ForCausalLM")
-class Phi3MiniModel(Model):
-    model_arch = gguf.MODEL_ARCH.PHI3
-
-    def set_vocab(self):
-        from sentencepiece import SentencePieceProcessor
-
-        tokenizer_path = self.dir_model / 'tokenizer.model'
-
-        if not tokenizer_path.is_file():
-            print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
-            sys.exit(1)
-
-        tokenizer = SentencePieceProcessor(str(tokenizer_path))
-
-        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
-
-        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
-        scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
-
-        for token_id in range(tokenizer.vocab_size()):
-
-            piece = tokenizer.id_to_piece(token_id)
-            text = piece.encode("utf-8")
-            score = tokenizer.get_score(token_id)
-
-            toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.is_unknown(token_id):
-                toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.is_control(token_id):
-                toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.is_unused(token_id):
-                toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.is_byte(token_id):
-                toktype = SentencePieceTokenTypes.BYTE
-
-            tokens[token_id] = text
-            scores[token_id] = score
-            toktypes[token_id] = toktype
-
-        added_tokens_file = self.dir_model / 'added_tokens.json'
-        if added_tokens_file.is_file():
-            with open(added_tokens_file, "r", encoding="utf-8") as f:
-                added_tokens_json = json.load(f)
-
-                for key in added_tokens_json:
-                    token_id = added_tokens_json[key]
-                    if (token_id >= vocab_size):
-                        print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
-                        continue
-
-                    tokens[token_id] = key.encode("utf-8")
-                    scores[token_id] = -1000.0
-                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
-
-        self.gguf_writer.add_tokenizer_model("llama")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-        special_vocab.add_to_gguf(self.gguf_writer)
-
-    def set_gguf_parameters(self):
-        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
-
-        rot_pct = 1.0
-        n_embd = self.find_hparam(["hidden_size", "n_embd"])
-        n_head = self.find_hparam(["num_attention_heads", "n_head"])
-        rms_eps = self.find_hparam(["rms_norm_eps"])
-
-        self.gguf_writer.add_name("Phi3")
-        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
-
-        self.gguf_writer.add_embedding_length(n_embd)
-        self.gguf_writer.add_feed_forward_length(8192)
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(n_head)
-        self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
-        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
-        self.gguf_writer.add_file_type(self.ftype)
-
-
 @Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
     model_arch = gguf.MODEL_ARCH.PLAMO
@@ -2303,8 +1908,6 @@ class InternLM2Model(Model):
         old_eos = special_vocab.special_token_ids["eos"]
         if "chat" in os.path.basename(self.dir_model.absolute()):
             # For the chat model, we replace the eos with '<|im_end|>'.
-            # TODO: this is a hack, should be fixed
-            #       https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
             special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer)
             print(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
 in chat mode so that the conversation can end normally.")
@@ -2541,16 +2144,6 @@ class GemmaModel(Model):
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
-        # TODO: these special tokens should be exported only for the CodeGemma family
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False,
-                                          special_token_types = ['prefix', 'suffix', 'middle', 'fsep', 'eot'])
-        special_vocab._set_special_token("prefix", 67)
-        special_vocab._set_special_token("suffix", 69)
-        special_vocab._set_special_token("middle", 68)
-        special_vocab._set_special_token("fsep",   70)
-        special_vocab._set_special_token("eot",    107)
-        special_vocab.add_to_gguf(self.gguf_writer)
-
     def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
@@ -2572,12 +2165,6 @@ class GemmaModel(Model):
         tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
 
         for name, data_torch in self.get_tensors():
-            # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
-            # To prevent errors, skip loading lm_head.weight.
-            if name == "lm_head.weight":
-                print(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
-                continue
-
             old_dtype = data_torch.dtype
 
             # convert any unsupported data types to float32
@@ -2637,34 +2224,28 @@ class MambaModel(Model):
 
             field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
             self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
             self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
             self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
             self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
             self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
             self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
-
             field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
             self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
 
     def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size",       "d_model"])
-        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
+        d_model = self.find_hparam(["hidden_size", "d_model"])
+        d_conv  = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
         d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 16
+        d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
         # ceiling division
         # ref: https://stackoverflow.com/a/17511341/22827863
         # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
-        dt_rank      = self.find_hparam(["time_step_rank",     "dt_rank"],      optional=True) or -(d_model // -16)
+        dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
 
         # Fail early for models which don't have a block expansion factor of 2
@@ -2756,66 +2337,6 @@ class CommandR2Model(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
 
 
-@Model.register("OlmoForCausalLM")
-@Model.register("OLMoForCausalLM")
-class OlmoModel(Model):
-    model_arch = gguf.MODEL_ARCH.OLMO
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_layer_norm_eps(1e-5)
-        if "clip_qkv" in self.hparams is not None:
-            self.gguf_writer.add_clamp_kqv(self.hparams["clip_qkv"])
-
-    # Same as super class, but permuting q_proj, k_proj
-    # Copied from: LlamaModel
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_head = self.hparams.get("num_attention_heads")
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.numpy()
-
-            if name.endswith("q_proj.weight"):
-                data = permute(data, n_head, n_head)
-            if name.endswith("k_proj.weight"):
-                data = permute(data, n_head, n_kv_head)
-
-            data = data.squeeze()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # 1d tensors need to be converted to float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
 ###### CONVERSION LOGIC ######
 
 
@@ -2842,7 +2363,6 @@ def parse_args() -> argparse.Namespace:
         "model", type=Path,
         help="directory containing model file",
     )
-    parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)")
 
     return parser.parse_args()
 
@@ -2886,7 +2406,7 @@ def main() -> None:
 
     with torch.inference_mode():
         model_class = Model.from_model_architecture(hparams["architectures"][0])
-        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file)
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
 
         print("Set model parameters")
         model_instance.set_gguf_parameters()

convert.py

@@ -525,14 +525,7 @@ class LlamaHfVocab(Vocab):
 
         # pre-check so we know if we need transformers
         tokenizer_model: dict[str, Any] = tokenizer_json['model']
-        is_llama3 = (
-            tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
-            and not tokenizer_model.get('byte_fallback', True)
-        )
-        if is_llama3:
-            raise TypeError('Llama 3 must be converted with BpeVocab')
-
-        if not is_llama3 and (
+        if (
             tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
             or tokenizer_json['decoder']['type'] != 'Sequence'
         ):

examples/CMakeLists.txt

@@ -13,6 +13,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 if (EMSCRIPTEN)
 else()
     add_subdirectory(baby-llama)
+    add_subdirectory(test-bench)
     add_subdirectory(batched)
     add_subdirectory(batched-bench)
     add_subdirectory(beam-search)

examples/batched.swift/Sources/main.swift

@@ -153,7 +153,7 @@ while n_cur <= n_len {
         // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
-        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {
@@ -229,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
 
 private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
     var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
     if nTokens < 0 {
         let actualTokensCount = -Int(nTokens)
         result = .init(repeating: 0, count: actualTokensCount)
@@ -237,8 +237,7 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
             model,
             token,
             &result,
-            Int32(result.count),
-            false
+            Int32(result.count)
         )
         assert(check == actualTokensCount)
     } else {

examples/batched/batched.cpp

@@ -191,8 +191,8 @@ int main(int argc, char ** argv) {
 
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
-            // is it an end of generation? -> mark the stream as finished
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            // is it an end of stream? -> mark the stream as finished
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && llama_token_is_eog(llama_get_model(callback_data.ctx), tokens[n_tokens-1]);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

examples/benchmark/benchmark-matmult.cpp

@@ -146,7 +146,7 @@ int main(int argc, char ** argv)  {
         /* no_alloc   =*/ 0
     };
 
-    ctx = ggml_init(params);
+    int main_gpu = 0;
     if (!ctx) {
         fprintf(stderr, "%s: ggml_init() failed\n", __func__);
         return 1;

examples/eval-callback/eval-callback.cpp

@@ -28,27 +28,14 @@ static std::string ggml_ne_string(const ggml_tensor * t) {
 }
 
 static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
-    GGML_ASSERT(n > 0);
     float sum = 0;
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
         printf("                                     [\n");
-        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
-            if (i2 == n && ne[2] > 2*n) {
-                printf("                                      ..., \n");
-                i2 = ne[2] - n;
-            }
+        for (int64_t i2 = 0; i2 < ne[2] && i2 < n; i2++) {
             printf("                                      [\n");
-            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
-                if (i1 == n && ne[1] > 2*n) {
-                    printf("                                       ..., \n");
-                    i1 = ne[1] - n;
-                }
+            for (int64_t i1 = 0; i1 < ne[1] && i1 < n; i1++) {
                 printf("                                       [");
-                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
-                    if (i0 == n && ne[0] > 2*n) {
-                        printf("..., ");
-                        i0 = ne[0] - n;
-                    }
+                for (int64_t i0 = 0; i0 < ne[0] && i0 < n; i0++) {
                     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
                     float v;
                     if (type == GGML_TYPE_F16) {
@@ -64,14 +51,17 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else {
                         GGML_ASSERT(false);
                     }
-                    printf("%12.4f", v);
+                    printf("%8.4f", v);
                     sum += v;
-                    if (i0 < ne[0] - 1) printf(", ");
+                    if (i0 < ne[0] - 1 && i0 < n - 1) printf(", ");
                 }
+                if (ne[0] > n) printf(", ...");
                 printf("],\n");
             }
+            if (ne[1] > n) printf("                                       ...\n");
             printf("                                      ],\n");
         }
+        if (ne[2] > n) printf("                                     ...\n");
         printf("                                     ]\n");
         printf("                                     sum = %f\n", sum);
     }

examples/gguf-split/README.md

@@ -5,6 +5,5 @@ CLI to split / merge GGUF files.
 **Command line options:**
 
 - `--split`: split GGUF to multiple GGUF, default operation.
-- `--split-max-size`: max size per split in `M` or `G`, f.ex. `500M` or `2G`.
 - `--split-max-tensors`: maximum tensors in each split: default(128)
 - `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -59,10 +59,10 @@ static size_t split_str_to_n_bytes(std::string str) {
     int n;
     if (str.back() == 'M') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = (size_t)n * 1024 * 1024; // megabytes
+        n_bytes = n * 1024 * 1024; // megabytes
     } else if (str.back() == 'G') {
         sscanf(str.c_str(), "%d", &n);
-        n_bytes = (size_t)n * 1024 * 1024 * 1024; // gigabytes
+        n_bytes = n * 1024 * 1024 * 1024; // gigabytes
     } else {
         throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
     }

examples/gguf-split/tests.sh

@@ -1,89 +0,0 @@
-#!/bin/bash
-
-set -eu
-
-if [ $# -lt 1 ]
-then
-    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
-    echo "example: $0 ../../build/bin ../../tmp"
-    exit 1
-fi
-
-if [ $# -gt 1 ]
-then
-    TMP_DIR=$2
-else
-    TMP_DIR=/tmp
-fi
-
-set -x
-
-SPLIT=$1/gguf-split
-MAIN=$1/main
-WORK_PATH=$TMP_DIR/gguf-split
-ROOT_DIR=$(realpath $(dirname $0)/../../)
-
-mkdir -p "$WORK_PATH"
-
-# Clean up in case of previously failed test
-rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf
-
-# 1. Get a model
-(
-cd $WORK_PATH
-"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/gemma-1.1-2b-it-Q8_0-GGUF --file gemma-1.1-2b-it.Q8_0.gguf
-)
-echo PASS
-
-# 2. Split with max tensors strategy
-$SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/ggml-model-split
-echo PASS
-echo
-
-# 2b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --random-prompt --n-predict 32
-echo PASS
-echo
-
-# 3. Merge
-$SPLIT --merge $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-merge.gguf
-echo PASS
-echo
-
-# 3b. Test the merged model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
-echo PASS
-echo
-
-# 4. Split with no tensor in metadata
-#$SPLIT --split-max-tensors 32 --no-tensor-in-metadata $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
-#echo PASS
-#echo
-
-# 4b. Test the sharded model is loading properly
-#$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf --random-prompt --n-predict 32
-#echo PASS
-#echo
-
-# 5. Merge
-#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf $WORK_PATH/ggml-model-merge-2.gguf
-#echo PASS
-#echo
-
-# 5b. Test the merged model is loading properly
-#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --random-prompt --n-predict 32
-#echo PASS
-#echo
-
-# 6. Split with size strategy
-$SPLIT --split-max-size 2G $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-2G
-echo PASS
-echo
-
-# 6b. Test the sharded model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --random-prompt --n-predict 32
-echo PASS
-echo
-
-# Clean up
-rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf

examples/gritlm/README.md

@@ -21,12 +21,12 @@ 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 --outdir models
+$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
 ```
 
 Run the example using the downloaded model:
 ```console
-$ ./gritlm -m models/gritlm-7b_q4_1.gguf
+$ ./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

examples/imatrix/imatrix.cpp

@@ -44,7 +44,7 @@ private:
     std::mutex                             m_mutex;
     int                                    m_last_call = 0;
     std::vector<float>                     m_src1_data;
-    std::vector<char>                      m_ids; // the expert ids from ggml_mul_mat_id
+    std::vector<int>                       m_ids; // the expert ids from ggml_mul_mat_id
                                                   //
     void save_imatrix(const char * file_name) const;
     void keep_imatrix(int ncall) const;
@@ -81,7 +81,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     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;
-        // why are small batches ignored (<16 tokens)?
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
         if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
         return true;
@@ -102,19 +101,14 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     // this has been adapted to the new format of storing merged experts in a single 3d tensor
     // ref: https://github.com/ggerganov/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
-        //   ids  -> [n_experts_used, n_tokens]
-        //   src1 -> [cols, n_expert_used, n_tokens]
+        const int idx  = ((int32_t *) t->op_params)[0];
         const ggml_tensor * ids = t->src[2];
         const int n_as = src0->ne[2];
-        const int n_ids = ids->ne[0];
 
         // the top-k selected expert ids are stored in the ids tensor
         // for simplicity, always copy ids to host, because it is small
-        // take into account that ids is not contiguous!
-
-        GGML_ASSERT(ids->ne[1] == src1->ne[2]);
-
-        m_ids.resize(ggml_nbytes(ids));
+        GGML_ASSERT(ids->ne[1] == src1->ne[1]);
+        m_ids.resize(ggml_nbytes(ids)/sizeof(int));
         ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
 
         auto & e = m_stats[wname];
@@ -124,35 +118,26 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         //       using the following line, we can correct for that if needed by replacing the line above with:
         //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
 
-        if (e.values.empty()) {
-            e.values.resize(src1->ne[0]*n_as, 0);
-        }
-        else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
-            exit(1); //GGML_ASSERT(false);
-        }
-        if (m_params.verbosity > 1) {
-            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[2], (int)src1->type);
-        }
         // loop over all possible experts, regardless if they are used or not in the batch
         for (int ex = 0; ex < n_as; ++ex) {
             size_t e_start = ex*src1->ne[0];
-
-            for (int idx = 0; idx < n_ids; ++idx) {
-                for (int row = 0; row < (int)src1->ne[2]; ++row) {
-                    const int excur = *(const int32_t *) (m_ids.data() + row*ids->nb[1] + idx*ids->nb[0]);
-
-                    GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
-
-                    if (excur != ex) continue;
-
-                    const int64_t i11 = idx % src1->ne[1];
-                    const int64_t i12 = row;
-                    const float * x = (const float *)((const char *)data + i11*src1->nb[1] + i12*src1->nb[2]);
-
-                    for (int j = 0; j < (int)src1->ne[0]; ++j) {
-                        e.values[e_start + j] += x[j]*x[j];
-                    }
+            if (e.values.empty()) {
+                e.values.resize(src1->ne[0]*n_as, 0);
+            }
+            else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
+                exit(1); //GGML_ASSERT(false);
+            }
+            if (m_params.verbosity > 1) {
+                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];
+                GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
+                if (excur != ex) continue;
+                const float * x = data + row * src1->ne[0];
+                for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                    e.values[e_start + j] += x[j]*x[j];
                 }
             }
             if (e.ncall > m_last_call) {

examples/infill/README.md

@@ -36,11 +36,6 @@ The `infill` program offers a seamless way to interact with LLaMA models, allowi
 
 ### Example
 
-Download a model that supports infill, for example CodeLlama:
-```console
-scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
-```
-
 ```bash
 ./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
 ```

examples/infill/infill.cpp

@@ -586,7 +586,7 @@ int main(int argc, char ** argv) {
 
             // deal with eot token in infill mode
             if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
-                if (is_interacting && !params.interactive_first) {
+                if(is_interacting && !params.interactive_first) {
                     // print an eot token
                     printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
@@ -651,8 +651,8 @@ int main(int argc, char ** argv) {
                 // LOG_TEE("took new input\n");
                 is_interacting = false;
             }
-            // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            // deal with end of text token in interactive mode
+            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -731,8 +731,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        // end of generation
-        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
             break;
         }
 

examples/json-schema-to-grammar.py

@@ -6,94 +6,37 @@ import re
 import sys
 from typing import Any, Dict, List, Set, Tuple, Union
 
-def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False):
-    if not separator_rule:
-        if min_items == 0 and max_items == 1:
-            return f'{item_rule}?'
-        elif min_items == 1 and max_items is None:
-            return f'{item_rule}+'
-
-    result = ''
-
-    if min_items > 0:
-        if item_rule_is_literal and separator_rule is None:
-            result = '"' + (item_rule[1:-1] * min_items) + '"'
-        else:
-            result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items)
-
-    def opt_repetitions(up_to_n, prefix_with_sep=False):
-        '''
-            - n=4, no sep:             '(a (a (a (a)?)?)?)?'
-            - n=4, sep=',', prefix:    '("," a ("," a ("," a ("," a)?)?)?)?'
-            - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?'
-        '''
-
-        content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule
-        if up_to_n == 0:
-            return ''
-        elif up_to_n == 1:
-            return f'({content})?'
-        elif separator_rule and not prefix_with_sep:
-            return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?'
-        else:
-            return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n)
-
-    if min_items > 0 and max_items != min_items:
-        result += ' '
-
-    if max_items is not None:
-        result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0)
-    else:
-        item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})'
-
-        if min_items == 0 and separator_rule:
-            result = f'({item_rule} {item_operator}*)?'
-        else:
-            result += f'{item_operator}*'
-
-    return result
-
-
-class BuiltinRule:
-    def __init__(self, content: str, deps: list = None):
-        self.content = content
-        self.deps = deps or []
-
-_up_to_15_digits = _build_repetition('[0-9]', 0, 15)
-
 # whitespace is constrained to a single space char to prevent model "running away" in
 # whitespace. Also maybe improves generation quality?
 SPACE_RULE = '" "?'
 
 PRIMITIVE_RULES = {
-    'boolean'      : BuiltinRule('("true" | "false") space', []),
-    'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []),
-    'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []),
-    'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
-    'integer'      : BuiltinRule('("-"? integral-part) space', ['integral-part']),
-    'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
-    'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
-    'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
-    'uuid'         : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []),
-    'char'         : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []),
-    'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
-    'null'         : BuiltinRule('"null" space', []),
+    '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''',
+    'null': '"null" space',
 }
+OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']
 
 # TODO: support "uri", "email" string formats
-STRING_FORMAT_RULES = {
-    'date'            : BuiltinRule('[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'            : BuiltinRule('([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'       : BuiltinRule('date "T" time', ['date', 'time']),
-    'date-string'     : BuiltinRule('"\\"" date "\\"" space', ['date']),
-    'time-string'     : BuiltinRule('"\\"" time "\\"" space', ['time']),
-    'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
+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',
 }
 
-DOTALL = '[\\U00000000-\\U0010FFFF]'
-DOT = '[^\\x0A\\x0D]'
-
-RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
+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"]')
@@ -103,6 +46,8 @@ 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, allow_fetch, dotall, raw_pattern):
@@ -110,9 +55,7 @@ class SchemaConverter:
         self._allow_fetch = allow_fetch
         self._dotall = dotall
         self._raw_pattern = raw_pattern
-        self._rules = {
-            'space': SPACE_RULE,
-        }
+        self._rules = {'space': SPACE_RULE}
         self._refs = {}
         self._refs_being_resolved = set()
 
@@ -122,29 +65,6 @@ class SchemaConverter:
         )
         return f'"{escaped}"'
 
-    def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str:
-        '''
-            not_literal('a') -> '[^a]'
-            not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?'
-        '''
-        assert len(literal) > 0, 'Empty literal not supported'
-        def recurse(i: int):
-            c = literal[i]
-            if maybe_escaped_underscores and c == '_':
-                yield f'[^{c}\\\\]'
-                yield ' | '
-                yield f'"\\\\"? "{c}"'
-            else:
-                yield f'[^{c}]'
-            if i < len(literal) - 1:
-                yield ' | '
-                yield self._format_literal(c)
-                yield ' ('
-                yield from recurse(i + 1)
-                yield ')?'
-
-        return ''.join(('(', *recurse(0), ')'))
-
     def _add_rule(self, name, rule):
         esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
         if esc_name not in self._rules or self._rules[esc_name] == rule:
@@ -249,10 +169,10 @@ class SchemaConverter:
 
             def get_dot():
                 if self._dotall:
-                    rule = DOTALL
+                    rule = '[\\U00000000-\\U0010FFFF]'
                 else:
                     # Accept any character... except \n and \r line break chars (\x0A and \xOD)
-                    rule = DOT
+                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
                 return self._add_rule(f'dot', rule)
 
             def join_seq():
@@ -326,14 +246,26 @@ class SchemaConverter:
 
                     (sub, sub_is_literal) = seq[-1]
 
-                    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
+                    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] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False)
+                        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:
@@ -441,47 +373,49 @@ class SchemaConverter:
                     ' "]" 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")
-                return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space')
+                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_primitive(
+            return self._add_rule(
                 'root' if rule_name == 'root' else schema_format,
                 PRIMITIVE_RULES['uuid']
             )
 
-        elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES:
-            prim_name = f'{schema_format}-string'
-            return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]))
-
-        elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema):
-            char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
-            min_len = schema.get('minLength', 0)
-            max_len = schema.get('maxLength')
-
-            return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" 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'
 
         elif (schema_type == 'object') or (len(schema) == 0):
-            return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
+            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_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type])
-
-    def _add_primitive(self, name: str, rule: BuiltinRule):
-        n = self._add_rule(name, rule.content)
-
-        for dep in rule.deps:
-            dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep)
-            assert dep_rule, f'Rule {dep} not known'
-            if dep not in self._rules:
-                self._add_primitive(dep, dep_rule)
-        return n
+            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
@@ -503,7 +437,7 @@ class SchemaConverter:
             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_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+                self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
             )
             optional_props.append("*")
 

examples/llama-bench/llama-bench.cpp

@@ -190,7 +190,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {get_math_cpu_count()},
+    /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},

examples/llama.android/app/src/main/cpp/llama-android.cpp

@@ -408,7 +408,7 @@ Java_com_example_llama_Llm_completion_1loop(
     const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
-    if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+    if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
         return env->NewStringUTF("");
     }
 

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

@@ -158,7 +158,7 @@ actor LlamaContext {
             new_token_id = llama_sample_token_greedy(context, &candidates_p)
         }
 
-        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             print("\n")
             let new_token_str = String(cString: temporary_invalid_cchars + [0])
             temporary_invalid_cchars.removeAll()
@@ -322,7 +322,7 @@ actor LlamaContext {
         defer {
             result.deallocate()
         }
-        let nTokens = llama_token_to_piece(model, token, result, 8, false)
+        let nTokens = llama_token_to_piece(model, token, result, 8)
 
         if nTokens < 0 {
             let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
@@ -330,7 +330,7 @@ actor LlamaContext {
             defer {
                 newResult.deallocate()
             }
-            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, false)
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
             let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
             return Array(bufferPointer)
         } else {

examples/llava/clip.cpp

@@ -3,7 +3,6 @@
 // I'll gradually clean and extend it
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
-#include "log.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -24,6 +23,7 @@
 #include <cstdlib>
 #include <cstring>
 #include <fstream>
+#include <iostream>
 #include <map>
 #include <regex>
 #include <stdexcept>
@@ -145,7 +145,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
 static int get_key_idx(const gguf_context * ctx, const char * key) {
     int i = gguf_find_key(ctx, key);
     if (i == -1) {
-        LOG_TEE("key %s not found in file\n", key);
+        fprintf(stderr, "key %s not found in file\n", key);
         throw std::runtime_error(format("Missing required key: %s", key));
     }
 
@@ -247,7 +247,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
 
 static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
     size_t tensor_size = ggml_nbytes(tensor);
-    LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+    printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
             prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
             tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
 }
@@ -265,7 +265,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
 static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        std::cerr << "Failed to open file for writing: " << filename << std::endl;
         return;
     }
 
@@ -284,7 +284,7 @@ static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::s
 static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        std::cerr << "Failed to open file for writing: " << filename << std::endl;
         return;
     }
 
@@ -515,7 +515,7 @@ struct clip_ctx {
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        printf("This gguf file seems to have no vision encoder\n");
         return nullptr;
     }
 
@@ -879,21 +879,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         const int idx_name = gguf_find_key(ctx, KEY_NAME);
         if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
             const std::string name = gguf_get_val_str(ctx, idx_name);
-            LOG_TEE("%s: model name:   %s\n", __func__, name.c_str());
+            printf("%s: model name:   %s\n", __func__, name.c_str());
         }
-        LOG_TEE("%s: description:  %s\n", __func__, description.c_str());
-        LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
-        LOG_TEE("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
-        LOG_TEE("%s: n_tensors:    %d\n", __func__, n_tensors);
-        LOG_TEE("%s: n_kv:         %d\n", __func__, n_kv);
-        LOG_TEE("%s: ftype:        %s\n", __func__, ftype_str.c_str());
-        LOG_TEE("\n");
+        printf("%s: description:  %s\n", __func__, description.c_str());
+        printf("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+        printf("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
+        printf("%s: n_tensors:    %d\n", __func__, n_tensors);
+        printf("%s: n_kv:         %d\n", __func__, n_kv);
+        printf("%s: ftype:        %s\n", __func__, ftype_str.c_str());
+        printf("\n");
     }
     const int n_tensors = gguf_get_n_tensors(ctx);
 
     // kv
     const int n_kv = gguf_get_n_kv(ctx);
-    LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+    printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
         __func__, n_kv, n_tensors, fname);
     {
         std::map<enum ggml_type, uint32_t> n_type;
@@ -904,7 +904,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             n_type[type]++;
         }
 
-        LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
         for (int i = 0; i < n_kv; i++) {
             const char * name           = gguf_get_key(ctx, i);
             const enum gguf_type type   = gguf_get_kv_type(ctx, i);
@@ -920,7 +920,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             }
             replace_all(value, "\n", "\\n");
 
-            LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+            printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
         }
 
         // print type counts
@@ -929,7 +929,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 continue;
             }
 
-            LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+            printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
         }
     }
 
@@ -944,7 +944,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             size_t tensor_size = ggml_nbytes(cur);
             model_size += tensor_size;
             if (verbosity >= 3) {
-                LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
                        __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
             }
         }
@@ -971,18 +971,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
 #ifdef GGML_USE_CUDA
     new_clip->backend = ggml_backend_cuda_init(0);
-    LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
+    printf("%s: CLIP using CUDA backend\n", __func__);
 #endif
 
 #ifdef GGML_USE_METAL
     new_clip->backend = ggml_backend_metal_init();
-    LOG_TEE("%s: CLIP using Metal backend\n", __func__);
+    printf("%s: CLIP using Metal backend\n", __func__);
 #endif
 
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
-        LOG_TEE("%s: CLIP using CPU backend\n", __func__);
+        printf("%s: CLIP using CPU backend\n", __func__);
     }
 
     // model size and capabilities
@@ -1006,15 +1006,15 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
 
         if (verbosity >= 1) {
-            LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
-            LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
-            LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
-            LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+            printf("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
+            printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+            printf("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            printf("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+            printf("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
-    LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
 
     // load tensors
     {
@@ -1027,7 +1027,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         new_clip->ctx_data = ggml_init(params);
         if (!new_clip->ctx_data) {
-            LOG_TEE("%s: ggml_init() failed\n", __func__);
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1035,7 +1035,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         auto fin = std::ifstream(fname, std::ios::binary);
         if (!fin) {
-            LOG_TEE("cannot open model file for loading tensors\n");
+            printf("cannot open model file for loading tensors\n");
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1057,7 +1057,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
-                LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
+                printf("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
                 gguf_free(ctx);
                 return nullptr;
@@ -1128,23 +1128,23 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         if (verbosity >= 2) {
-            LOG_TEE("\n%s: vision model hparams\n", __func__);
-            LOG_TEE("image_size         %d\n", hparams.image_size);
-            LOG_TEE("patch_size         %d\n", hparams.patch_size);
-            LOG_TEE("v_hidden_size      %d\n", hparams.hidden_size);
-            LOG_TEE("v_n_intermediate   %d\n", hparams.n_intermediate);
-            LOG_TEE("v_projection_dim   %d\n", hparams.projection_dim);
-            LOG_TEE("v_n_head           %d\n", hparams.n_head);
-            LOG_TEE("v_n_layer          %d\n", hparams.n_layer);
-            LOG_TEE("v_eps              %f\n", hparams.eps);
-            LOG_TEE("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
-            LOG_TEE("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
-            LOG_TEE("v_image_grid_pinpoints: ");
+            printf("\n%s: vision model hparams\n", __func__);
+            printf("image_size         %d\n", hparams.image_size);
+            printf("patch_size         %d\n", hparams.patch_size);
+            printf("v_hidden_size      %d\n", hparams.hidden_size);
+            printf("v_n_intermediate   %d\n", hparams.n_intermediate);
+            printf("v_projection_dim   %d\n", hparams.projection_dim);
+            printf("v_n_head           %d\n", hparams.n_head);
+            printf("v_n_layer          %d\n", hparams.n_layer);
+            printf("v_eps              %f\n", hparams.eps);
+            printf("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+            printf("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+            printf("v_image_grid_pinpoints: ");
             for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
-                LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
+                printf("%d ", hparams.image_grid_pinpoints[i]);
             }
-            LOG_TEE("\n");
-            LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+            printf("\n");
+            printf("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
 
         }
 
@@ -1155,7 +1155,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
             vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
         } catch(const std::exception& e) {
-            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
+            fprintf(stderr, "%s: failed to load vision model tensors\n", __func__);
         }
 
         // LLaVA projection
@@ -1184,7 +1184,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             } catch (std::runtime_error & e) {  }
             try {
                 vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
-                // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
+                // fprintf(stderr, "%s: image_newline tensor (llava-1.6) found\n", __func__);
             } catch (std::runtime_error & e) {  }
         } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
@@ -1264,7 +1264,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
-        LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+        printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
     return new_clip;
@@ -1304,7 +1304,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
     int nx, ny, nc;
     auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
     if (!data) {
-        LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
+        fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1316,7 +1316,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
     int nx, ny, nc;
     auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
     if (!data) {
-        LOG_TEE("%s: failed to decode image bytes\n", __func__);
+        fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1506,7 +1506,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int> & or
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -1545,7 +1545,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
 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) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        printf("This gguf file seems to have no vision encoder\n");
         return false;
     }
     auto & params = ctx->vision_model.hparams;
@@ -1622,7 +1622,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             }
 
             for (size_t i = 0; i < patches.size(); i++) {
-                // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+                // printf("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
                 clip_image_u8_free(patches[i]);
             }
 
@@ -1765,7 +1765,7 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        printf("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -1777,7 +1777,7 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
 
 bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        printf("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -1939,7 +1939,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             new_type = type;
             if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
                 new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
-                // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+                // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
             }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
@@ -1958,7 +1958,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 f32_data = (float *)conv_buf.data();
                 break;
             default:
-                LOG_TEE("Please use an input file in f32 or f16\n");
+                printf("Please use an input file in f32 or f16\n");
                 gguf_free(ctx_out);
                 return false;
             }
@@ -1985,7 +1985,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 
@@ -2001,8 +2001,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     gguf_free(ctx_out);
 
     {
-        LOG_TEE("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
-        LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+        printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
     }
 
     return true;

examples/llava/llava-cli.cpp

@@ -1,5 +1,4 @@
 #include "ggml.h"
-#include "log.h"
 #include "common.h"
 #include "clip.h"
 #include "llava.h"
@@ -19,7 +18,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
             n_eval = n_batch;
         }
         if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
-            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
         *n_past += n_eval;
@@ -46,7 +45,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
     const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
     llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
     static std::string ret;
-    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
     } else {
         ret = llama_token_to_piece(ctx_llama, id);
@@ -74,7 +73,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
     size_t img_base64_str_start, img_base64_str_end;
     find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
     if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        fprintf(stderr, "%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
         return NULL;
     }
 
@@ -88,7 +87,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
 
     auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
     if (!embed) {
-        LOG_TEE("%s: could not load image from base64 string.\n", __func__);
+        fprintf(stderr, "%s: could not load image from base64 string.\n", __func__);
         return NULL;
     }
 
@@ -113,8 +112,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    fprintf(stderr, "\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    fprintf(stderr, "  note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
@@ -124,18 +123,18 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     auto prompt = params->prompt;
     if (prompt_contains_image(prompt)) {
         if (!params->image.empty()) {
-            LOG_TEE("using base64 encoded image instead of command line image path\n");
+            fprintf(stderr, "using base64 encoded image instead of command line image path\n");
         }
         embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
         if (!embed) {
-            LOG_TEE("%s: can't load image from prompt\n", __func__);
+            fprintf(stderr, "%s: can't load image from prompt\n", __func__);
             return NULL;
         }
         params->prompt = remove_image_from_prompt(prompt);
     } else {
         embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
         if (!embed) {
-            LOG_TEE("%s: is %s really an image file?\n", __func__, params->image.c_str());
+            fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str());
             return NULL;
         }
     }
@@ -154,18 +153,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        LOG_TEE("system_prompt: %s\n", system_prompt.c_str());
+        printf("system_prompt: %s\n", system_prompt.c_str());
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
-        LOG_TEE("user_prompt: %s\n", user_prompt.c_str());
+        printf("user_prompt: %s\n", user_prompt.c_str());
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     } else {
@@ -175,7 +174,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     }
@@ -186,7 +185,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     // generate the response
 
-    LOG_TEE("\n");
+    fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
     std::string response = "";
@@ -225,7 +224,7 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
-        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
         return NULL;
     }
 
@@ -235,7 +234,7 @@ static struct llava_context * llava_init(gpt_params * params) {
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
-        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return NULL;
     }
 
@@ -258,12 +257,6 @@ static void llava_free(struct llava_context * ctx_llava) {
     llama_backend_free();
 }
 
-static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
-    (void) level;
-    (void) user_data;
-    LOG_TEE("%s", text);
-}
-
 int main(int argc, char ** argv) {
     ggml_time_init();
 
@@ -273,14 +266,6 @@ int main(int argc, char ** argv) {
         show_additional_info(argc, argv);
         return 1;
     }
-
-#ifndef LOG_DISABLE_LOGS
-    log_set_target(log_filename_generator("llava", "log"));
-    LOG_TEE("Log start\n");
-    log_dump_cmdline(argc, argv);
-    llama_log_set(llama_log_callback_logTee, nullptr);
-#endif // LOG_DISABLE_LOGS
-
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         gpt_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
@@ -289,7 +274,7 @@ int main(int argc, char ** argv) {
 
     auto ctx_llava = llava_init(&params);
     if (ctx_llava == NULL) {
-        LOG_TEE("%s: error: failed to init llava\n", __func__);
+        fprintf(stderr, "%s: error: failed to init llava\n", __func__);
         return 1;
     }
 

examples/llava/llava.cpp

@@ -54,7 +54,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int>& ori
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -154,13 +154,13 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
     if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
         if (newline_tmp->buffer == NULL) {
-            LOG_TEE("newline_tmp tensor buffer is NULL\n");
+            printf("newline_tmp tensor buffer is NULL\n");
         }
         ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
     } else {
         model.newline->data = newline_tmp->data;
         if (model.newline->data == NULL) {
-            LOG_TEE("newline_tmp tensor data is NULL\n");
+            printf("newline_tmp tensor data is NULL\n");
         }
     }
 
@@ -224,7 +224,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     img_res_v.size = 0;
     img_res_v.data = nullptr;
     if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
-        LOG_TEE("%s: unable to preprocess image\n", __func__);
+        fprintf(stderr, "%s: unable to preprocess image\n", __func__);
         delete[] img_res_v.data;
         return false;
     }
@@ -239,7 +239,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
         delete[] img_res_v.data;
         if (!encoded) {
-            LOG_TEE("Unable to encode image\n");
+            fprintf(stderr, "Unable to encode image\n");
 
             return false;
         }
@@ -252,12 +252,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
             const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_TEE("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
 
@@ -290,12 +290,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
-    LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
 
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
-    LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+    printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
 
     return true;
 }
@@ -305,7 +305,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
-        LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
         return false;
     }
     return true;
@@ -314,13 +314,13 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 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) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
-        LOG_TEE("Unable to allocate memory for image embeddings\n");
+        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
         return false;
     }
 
     int n_img_pos;
     if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
-        LOG_TEE("%s: cannot encode image, aborting\n", __func__);
+        fprintf(stderr, "%s: cannot encode image, aborting\n", __func__);
         free(image_embd);
         return false;
     }
@@ -340,7 +340,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         }
         llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
         if (llama_decode(ctx_llama, batch)) {
-            LOG_TEE("%s : failed to eval\n", __func__);
+            fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
         *n_past += n_eval;
@@ -352,7 +352,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
-        LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
+        fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);
         return NULL;
     }
 
@@ -361,7 +361,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        LOG_TEE("%s: coulnd't embed the image\n", __func__);
+        fprintf(stderr, "%s: coulnd't embed the image\n", __func__);
         return NULL;
     }
 
@@ -375,7 +375,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
 static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
     auto file = fopen(path, "rb");
     if (file == NULL) {
-        LOG_TEE("%s: can't read file %s\n", __func__, path);
+        fprintf(stderr, "%s: can't read file %s\n", __func__, path);
         return false;
     }
 
@@ -385,7 +385,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
 
     auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
     if (buffer == NULL) {
-        LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
         perror("Memory allocation error");
         fclose(file);
         return false;
@@ -410,7 +410,7 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
     if (!loaded) {
-        LOG_TEE("%s: failed to load %s\n", __func__, image_path);
+        fprintf(stderr, "%s: failed to load %s\n", __func__, image_path);
         return NULL;
     }
 

examples/lookahead/lookahead.cpp

@@ -299,7 +299,7 @@ int main(int argc, char ** argv) {
                 }
                 fflush(stdout);
 
-                if (llama_token_is_eog(model, id)) {
+                if (id == llama_token_eos(model)) {
                     has_eos = true;
                 }
 

examples/lookup/lookup-stats.cpp

@@ -30,6 +30,7 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
     GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt

examples/lookup/lookup.cpp

@@ -38,6 +38,7 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
     GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
@@ -140,7 +141,7 @@ int main(int argc, char ** argv){
                 printf("%s", token_str.c_str());
             }
 
-            if (llama_token_is_eog(model, id)) {
+            if (id == llama_token_eos(model)) {
                 has_eos = true;
             }
 

examples/main/README.md

@@ -304,12 +304,10 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 -   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
-### Grammars & JSON schemas
+### Grammars
 
 -   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
 
--   `--json-schema SCHEMA`: Specify a [JSON schema](https://json-schema.org/) to constrain model output to (e.g. `{}` for any JSON object, or `{"items": {"type": "string", "minLength": 10, "maxLength": 100}, "minItems": 10}` for a JSON array of strings with size constraints). If a schema uses external `$ref`s, you should use `--grammar "$( python examples/json_schema_to_grammar.py myschema.json )"` instead.
-
 ### Quantization
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).

examples/main/main.cpp

@@ -240,6 +240,7 @@ int main(int argc, char ** argv) {
                 return 1;
             }
             session_tokens.resize(n_token_count_out);
+            llama_set_rng_seed(ctx, params.seed);
             LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
         }
     }
@@ -794,8 +795,8 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            // deal with end of generation tokens in interactive mode
-            if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            // deal with end of text token in interactive mode
+            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -919,8 +920,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        // end of generation
-        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.instruct || params.interactive || params.chatml)) {
+        // end of text token
+        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive || params.chatml)) {
             LOG_TEE(" [end of text]\n");
             break;
         }

examples/parallel/parallel.cpp

@@ -359,7 +359,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (llama_token_is_eog(model, id) ||
+                        (id == llama_token_eos(model) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {

examples/passkey/passkey.cpp

@@ -252,8 +252,8 @@ int main(int argc, char ** argv) {
             // sample the most likely token
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
-            // is it an end of generation?
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;

examples/perplexity/perplexity.cpp

@@ -1852,20 +1852,12 @@ int main(int argc, char ** argv) {
 
     const int32_t n_ctx = params.n_ctx;
 
-    if (n_ctx <= 0) {
-        fprintf(stderr, "%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
-        return 1;
-    }
-
     const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence;
-
     if (ppl) {
-        const int32_t n_seq = std::max(1, params.n_batch / n_ctx);
-        const int32_t n_kv = n_seq * n_ctx;
-
+        int n_seq = std::max(1, params.n_batch / n_ctx);
+        int32_t n_kv = n_seq * n_ctx;
         params.n_parallel = n_seq;
-        params.n_ctx      = n_kv;
-
+        params.n_ctx = n_kv;
         params.n_batch = std::min(params.n_batch, n_kv);
     } else {
         params.n_batch = std::min(params.n_batch, params.n_ctx);

examples/quantize/quantize.cpp

@@ -97,7 +97,6 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
-    printf("  --keep-split: will generate quatized model in the same shards as input");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");
@@ -301,8 +300,6 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
-        } else if (strcmp(argv[arg_idx], "--keep-split")) {
-            params.keep_split = true;
         } else {
             usage(argv[0]);
         }
@@ -335,28 +332,20 @@ int main(int argc, char ** argv) {
     std::string fname_out;
 
     std::string ftype_str;
-    std::string suffix = ".gguf";
     if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
         std::string fpath;
         const size_t pos = fname_inp.find_last_of("/\\");
         if (pos != std::string::npos) {
             fpath = fname_inp.substr(0, pos + 1);
         }
-
-        // export as [inp path]/ggml-model-[ftype]. Only add extension if there is no splitting
-        fname_out = fpath + "ggml-model-" + ftype_str;
-        if (!params.keep_split) {
-            fname_out += suffix;
-        }
+        // export as [inp path]/ggml-model-[ftype].gguf
+        fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
         arg_idx++;
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
     } else {
         fname_out = argv[arg_idx];
-        if (params.keep_split && fname_out.find(suffix) != std::string::npos) {
-            fname_out = fname_out.substr(0, fname_out.length() - suffix.length());
-        }
         arg_idx++;
 
         if (argc <= arg_idx) {

examples/quantize/tests.sh

@@ -1,65 +0,0 @@
-#!/bin/bash
-
-set -eu
-
-if [ $# -lt 1 ]
-then
-    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
-    echo "example: $0 ../../build/bin ../../tmp"
-    exit 1
-fi
-
-if [ $# -gt 1 ]
-then
-    TMP_DIR=$2
-else
-    TMP_DIR=/tmp
-fi
-
-set -x
-
-SPLIT=$1/gguf-split
-QUANTIZE=$1/quantize
-MAIN=$1/main
-WORK_PATH=$TMP_DIR/quantize
-ROOT_DIR=$(realpath $(dirname $0)/../../)
-
-mkdir -p "$WORK_PATH"
-
-# Clean up in case of previously failed test
-rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf
-
-# 1. Get a model
-(
-cd $WORK_PATH
-"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/gemma-1.1-2b-it-Q8_0-GGUF --file gemma-1.1-2b-it.Q8_0.gguf
-)
-echo PASS
-
-# 2. Split model
-$SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/ggml-model-split
-echo PASS
-echo
-
-# 3. Requant model with '--keep_split'
-$QUANTIZE --allow-requantize --keep_split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
-echo PASS
-echo
-
-# 3a. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt --n-predict 32
-echo PASS
-echo
-
-# 4. Requant mode without '--keep_split'
-$QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
-echo PASS
-echo
-
-# 4b. Test the requanted model is loading properly
-$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --random-prompt --n-predict 32
-echo PASS
-echo
-
-# Clean up
-rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf

examples/regex-to-grammar.py

@@ -8,7 +8,7 @@ print(subprocess.check_output(
         "python",
         os.path.join(
         os.path.dirname(os.path.realpath(__file__)),
-        "json_schema_to_grammar.py"),
+        "json-schema-to-grammar.py"),
         *rest,
         "-",
         "--raw-pattern",

examples/server/CMakeLists.txt

@@ -1,34 +1,17 @@
 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} ${CMAKE_CURRENT_BINARY_DIR})
-set(TARGET_SRCS
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET}
     server.cpp
     utils.hpp
     httplib.h
 )
-set(PUBLIC_ASSETS
-    index.html
-    index.js
-    completion.js
-    json-schema-to-grammar.mjs
-)
-foreach(asset ${PUBLIC_ASSETS})
-    set(input "${CMAKE_CURRENT_SOURCE_DIR}/public/${asset}")
-    set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp")
-    list(APPEND TARGET_SRCS ${output})
-    add_custom_command(
-        DEPENDS "${input}"
-        OUTPUT "${output}"
-        COMMAND "${CMAKE_COMMAND}" "-DINPUT=${input}" "-DOUTPUT=${output}" -P "${PROJECT_SOURCE_DIR}/scripts/xxd.cmake"
-    )
-endforeach()
-add_executable(${TARGET} ${TARGET_SRCS})
 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

@@ -11,7 +11,6 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
  * Continuous batching
  * Multimodal (wip)
  * Monitoring endpoints
- * Schema-constrained JSON response format
 
 The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
 
@@ -251,8 +250,6 @@ node index.js
 
     `grammar`: Set grammar for grammar-based sampling.  Default: no grammar
 
-    `json_schema`: Set a JSON schema for grammar-based sampling (e.g. `{"items": {"type": "string"}, "minItems": 10, "maxItems": 100}` of a list of strings, or `{}` for any JSON). See [tests](../../tests/test-json-schema-to-grammar.cpp) for supported features.  Default: no JSON schema.
-
     `seed`: Set the random number generator (RNG) seed.  Default: `-1`, which is a random seed.
 
     `ignore_eos`: Ignore end of stream token and continue generating.  Default: `false`
@@ -368,8 +365,6 @@ Notice that each `probs` is an array of length `n_probs`.
 
     See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.
 
-    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}`), similar to other OpenAI-inspired API providers.
-
     *Examples:*
 
     You can use either Python `openai` library with appropriate checkpoints:

examples/server/completion.js.hpp

@@ -0,0 +1,496 @@
+unsigned char completion_js[] = {
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x44,
+  0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x3a, 0x20, 0x74, 0x72,
+  0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x3a, 0x20, 0x35, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20,
+  0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x3a,
+  0x20, 0x30, 0x2e, 0x32, 0x2c, 0x0a, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70,
+  0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x5d, 0x0a, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x6c, 0x65, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72,
+  0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e,
+  0x67, 0x73, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x0a,
+  0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65,
+  0x73, 0x20, 0x74, 0x68, 0x65, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74,
+  0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61,
+  0x74, 0x6f, 0x72, 0x2e, 0x20, 0x52, 0x65, 0x63, 0x6f, 0x6d, 0x6d, 0x65,
+  0x6e, 0x64, 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x6d, 0x6f, 0x73,
+  0x74, 0x20, 0x75, 0x73, 0x65, 0x20, 0x63, 0x61, 0x73, 0x65, 0x73, 0x2e,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70,
+  0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27,
+  0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e,
+  0x6a, 0x73, 0x27, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x71, 0x75, 0x65,
+  0x73, 0x74, 0x20, 0x3d, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x22,
+  0x54, 0x65, 0x6c, 0x6c, 0x20, 0x6d, 0x65, 0x20, 0x61, 0x20, 0x6a, 0x6f,
+  0x6b, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64,
+  0x69, 0x63, 0x74, 0x3a, 0x20, 0x38, 0x30, 0x30, 0x7d, 0x29, 0x0a, 0x2f,
+  0x2f, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61,
+  0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68,
+  0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x72, 0x65, 0x71, 0x75, 0x65,
+  0x73, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77,
+  0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64,
+  0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x2f, 0x2f, 0x0a,
+  0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63,
+  0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x2a, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c,
+  0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d,
+  0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b,
+  0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20,
+  0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x20, 0x3d,
+  0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x2e, 0x61, 0x70, 0x69, 0x5f,
+  0x75, 0x72, 0x6c, 0x20, 0x7c, 0x7c, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x0a,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x63, 0x6f, 0x6e, 0x74, 0x72,
+  0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20,
+  0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e,
+  0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e,
+  0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x44, 0x65, 0x66, 0x61, 0x75,
+  0x6c, 0x74, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2c, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72,
+  0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x77,
+  0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x60, 0x24,
+  0x7b, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x7d, 0x2f, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x60, 0x2c, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x74, 0x68, 0x6f, 0x64, 0x3a,
+  0x20, 0x27, 0x50, 0x4f, 0x53, 0x54, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x62, 0x6f, 0x64, 0x79, 0x3a, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e,
+  0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x50, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x61,
+  0x64, 0x65, 0x72, 0x73, 0x3a, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x27, 0x43, 0x6f, 0x6e, 0x6e, 0x65, 0x63, 0x74, 0x69, 0x6f,
+  0x6e, 0x27, 0x3a, 0x20, 0x27, 0x6b, 0x65, 0x65, 0x70, 0x2d, 0x61, 0x6c,
+  0x69, 0x76, 0x65, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x27, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2d, 0x54, 0x79, 0x70,
+  0x65, 0x27, 0x3a, 0x20, 0x27, 0x61, 0x70, 0x70, 0x6c, 0x69, 0x63, 0x61,
+  0x74, 0x69, 0x6f, 0x6e, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x27, 0x2c, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x27, 0x41, 0x63, 0x63, 0x65, 0x70,
+  0x74, 0x27, 0x3a, 0x20, 0x27, 0x74, 0x65, 0x78, 0x74, 0x2f, 0x65, 0x76,
+  0x65, 0x6e, 0x74, 0x2d, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x27, 0x2c,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x28, 0x70,
+  0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65,
+  0x79, 0x20, 0x3f, 0x20, 0x7b, 0x27, 0x41, 0x75, 0x74, 0x68, 0x6f, 0x72,
+  0x69, 0x7a, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x60, 0x42,
+  0x65, 0x61, 0x72, 0x65, 0x72, 0x20, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61,
+  0x6d, 0x73, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x7d, 0x60,
+  0x7d, 0x20, 0x3a, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61,
+  0x6c, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65,
+  0x72, 0x2e, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x0a, 0x20, 0x20,
+  0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74,
+  0x20, 0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x72, 0x65,
+  0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x2e,
+  0x67, 0x65, 0x74, 0x52, 0x65, 0x61, 0x64, 0x65, 0x72, 0x28, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63,
+  0x6f, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x54,
+  0x65, 0x78, 0x74, 0x44, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x72, 0x28, 0x29,
+  0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20,
+  0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65,
+  0x72, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x42,
+  0x75, 0x66, 0x66, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x61,
+  0x72, 0x74, 0x69, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x72, 0x65, 0x61, 0x64,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x74, 0x72,
+  0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b,
+  0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20,
+  0x28, 0x63, 0x6f, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20,
+  0x72, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2e, 0x72, 0x65, 0x61, 0x64, 0x28,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x6f, 0x6e, 0x65,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x62, 0x72, 0x65, 0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x41, 0x64, 0x64, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x6c, 0x65, 0x66,
+  0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x74,
+  0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e,
+  0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x6c,
+  0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x2b, 0x20, 0x64, 0x65,
+  0x63, 0x6f, 0x64, 0x65, 0x72, 0x2e, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75,
+  0x65, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x20, 0x69, 0x66, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x72,
+  0x61, 0x63, 0x74, 0x65, 0x72, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e,
+  0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65, 0x42, 0x72,
+  0x65, 0x61, 0x6b, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2e, 0x65,
+  0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, 0x5c, 0x6e, 0x27,
+  0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x53, 0x70, 0x6c, 0x69, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x6c, 0x69, 0x6e,
+  0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74,
+  0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x78,
+  0x74, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x5c, 0x6e, 0x27,
+  0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f,
+  0x20, 0x49, 0x66, 0x20, 0x74, 0x68, 0x65, 0x20, 0x74, 0x65, 0x78, 0x74,
+  0x20, 0x64, 0x6f, 0x65, 0x73, 0x6e, 0x27, 0x74, 0x20, 0x65, 0x6e, 0x64,
+  0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65,
+  0x20, 0x62, 0x72, 0x65, 0x61, 0x6b, 0x2c, 0x20, 0x74, 0x68, 0x65, 0x6e,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x6c, 0x69,
+  0x6e, 0x65, 0x20, 0x69, 0x73, 0x20, 0x69, 0x6e, 0x63, 0x6f, 0x6d, 0x70,
+  0x6c, 0x65, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f,
+  0x2f, 0x20, 0x53, 0x74, 0x6f, 0x72, 0x65, 0x20, 0x69, 0x74, 0x20, 0x69,
+  0x6e, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x74,
+  0x6f, 0x20, 0x62, 0x65, 0x20, 0x61, 0x64, 0x64, 0x65, 0x64, 0x20, 0x74,
+  0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6e, 0x65, 0x78, 0x74, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x61, 0x74, 0x61,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21,
+  0x65, 0x6e, 0x64, 0x73, 0x57, 0x69, 0x74, 0x68, 0x4c, 0x69, 0x6e, 0x65,
+  0x42, 0x72, 0x65, 0x61, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65,
+  0x72, 0x20, 0x3d, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73, 0x2e, 0x70, 0x6f,
+  0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x20, 0x2f, 0x2f, 0x20, 0x52, 0x65,
+  0x73, 0x65, 0x74, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x6f, 0x76, 0x65, 0x72,
+  0x20, 0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x68, 0x61, 0x76, 0x65, 0x20,
+  0x61, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x20, 0x62, 0x72, 0x65, 0x61, 0x6b,
+  0x20, 0x61, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x65, 0x6e, 0x64, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x50, 0x61, 0x72, 0x73, 0x65, 0x20,
+  0x61, 0x6c, 0x6c, 0x20, 0x73, 0x73, 0x65, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x64, 0x64, 0x20, 0x74,
+  0x68, 0x65, 0x6d, 0x20, 0x74, 0x6f, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x20, 0x3d, 0x20, 0x2f, 0x5e,
+  0x28, 0x5c, 0x53, 0x2b, 0x29, 0x3a, 0x5c, 0x73, 0x28, 0x2e, 0x2a, 0x29,
+  0x24, 0x2f, 0x67, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x66, 0x6f, 0x72, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x69, 0x6e, 0x65, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x73,
+  0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x20,
+  0x3d, 0x20, 0x72, 0x65, 0x67, 0x65, 0x78, 0x2e, 0x65, 0x78, 0x65, 0x63,
+  0x28, 0x6c, 0x69, 0x6e, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6d, 0x61, 0x74, 0x63,
+  0x68, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x5b, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x31, 0x5d, 0x5d, 0x20, 0x3d, 0x20, 0x6d, 0x61,
+  0x74, 0x63, 0x68, 0x5b, 0x32, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6e, 0x63,
+  0x65, 0x20, 0x77, 0x65, 0x20, 0x6b, 0x6e, 0x6f, 0x77, 0x20, 0x74, 0x68,
+  0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x2c, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x6a,
+  0x75, 0x73, 0x74, 0x20, 0x64, 0x65, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x74,
+  0x68, 0x65, 0x20, 0x6a, 0x73, 0x6f, 0x6e, 0x20, 0x69, 0x6e, 0x20, 0x64,
+  0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73,
+  0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x79, 0x69, 0x65, 0x6c, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x79, 0x69, 0x65, 0x6c, 0x64, 0x20,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20,
+  0x69, 0x66, 0x20, 0x77, 0x65, 0x20, 0x67, 0x6f, 0x74, 0x20, 0x61, 0x20,
+  0x73, 0x74, 0x6f, 0x70, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x20, 0x66,
+  0x72, 0x6f, 0x6d, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2c, 0x20,
+  0x77, 0x65, 0x20, 0x77, 0x69, 0x6c, 0x6c, 0x20, 0x62, 0x72, 0x65, 0x61,
+  0x6b, 0x20, 0x68, 0x65, 0x72, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73,
+  0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20,
+  0x28, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69,
+  0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20,
+  0x3d, 0x20, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+  0x5f, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x20, 0x3d, 0x20,
+  0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x72, 0x65,
+  0x61, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x73, 0x75,
+  0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x4a,
+  0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x72, 0x65,
+  0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x72, 0x65, 0x73, 0x75, 0x6c,
+  0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e, 0x6d, 0x65, 0x73, 0x73,
+  0x61, 0x67, 0x65, 0x2e, 0x69, 0x6e, 0x63, 0x6c, 0x75, 0x64, 0x65, 0x73,
+  0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61, 0x76, 0x61,
+  0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x54, 0x68, 0x72, 0x6f, 0x77,
+  0x20, 0x61, 0x6e, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x74, 0x6f,
+  0x20, 0x62, 0x65, 0x20, 0x63, 0x61, 0x75, 0x67, 0x68, 0x74, 0x20, 0x62,
+  0x79, 0x20, 0x75, 0x70, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x20, 0x63,
+  0x61, 0x6c, 0x6c, 0x65, 0x72, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74,
+  0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72,
+  0x6f, 0x72, 0x28, 0x27, 0x73, 0x6c, 0x6f, 0x74, 0x20, 0x75, 0x6e, 0x61,
+  0x76, 0x61, 0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x27, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e,
+  0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x20, 0x5b, 0x24,
+  0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x2e, 0x63, 0x6f, 0x64, 0x65, 0x7d, 0x20, 0x2d, 0x20, 0x24, 0x7b,
+  0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72,
+  0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x5d, 0x3a, 0x20, 0x24, 0x7b, 0x72,
+  0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x2e,
+  0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x28,
+  0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x60, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x20, 0x65, 0x72, 0x72, 0x6f,
+  0x72, 0x20, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x2e, 0x65,
+  0x72, 0x72, 0x6f, 0x72, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20,
+  0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x2e,
+  0x6e, 0x61, 0x6d, 0x65, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x27, 0x41, 0x62,
+  0x6f, 0x72, 0x74, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x27, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f,
+  0x6c, 0x65, 0x2e, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x6c, 0x6c,
+  0x61, 0x6d, 0x61, 0x20, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x3a, 0x20, 0x22,
+  0x2c, 0x20, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x65, 0x3b,
+  0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x61, 0x6c,
+  0x6c, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x61, 0x62, 0x6f, 0x72,
+  0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20,
+  0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43, 0x61,
+  0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72, 0x65,
+  0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e,
+  0x74, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x74, 0x68, 0x61,
+  0x74, 0x20, 0x79, 0x6f, 0x75, 0x20, 0x63, 0x61, 0x6e, 0x20, 0x73, 0x75,
+  0x62, 0x73, 0x63, 0x72, 0x69, 0x62, 0x65, 0x20, 0x74, 0x6f, 0x0a, 0x2f,
+  0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65,
+  0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74,
+  0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f,
+  0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x20, 0x3d, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a,
+  0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x6e, 0x2e, 0x61,
+  0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65,
+  0x6e, 0x65, 0x72, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+  0x22, 0x2c, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x72, 0x69,
+  0x74, 0x65, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x65, 0x74,
+  0x61, 0x69, 0x6c, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29,
+  0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x2f, 0x2f,
+  0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73,
+  0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72,
+  0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73,
+  0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69,
+  0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x76, 0x65,
+  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x6e,
+  0x65, 0x77, 0x20, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67,
+  0x65, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x28, 0x61, 0x73, 0x79,
+  0x6e, 0x63, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28,
+  0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20,
+  0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f,
+  0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c,
+  0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x29, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x68,
+  0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65,
+  0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64,
+  0x69, 0x73, 0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74,
+  0x28, 0x6e, 0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45,
+  0x76, 0x65, 0x6e, 0x74, 0x28, 0x22, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67,
+  0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c,
+  0x3a, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61,
+  0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67,
+  0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65,
+  0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f,
+  0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x22, 0x2c, 0x20, 0x7b,
+  0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69,
+  0x66, 0x20, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74,
+  0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b,
+  0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65,
+  0x6e, 0x74, 0x54, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73,
+  0x70, 0x61, 0x74, 0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e,
+  0x65, 0x77, 0x20, 0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65,
+  0x6e, 0x74, 0x28, 0x22, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x22,
+  0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74,
+  0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54,
+  0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x64, 0x69, 0x73, 0x70, 0x61, 0x74,
+  0x63, 0x68, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28, 0x6e, 0x65, 0x77, 0x20,
+  0x43, 0x75, 0x73, 0x74, 0x6f, 0x6d, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x28,
+  0x22, 0x64, 0x6f, 0x6e, 0x65, 0x22, 0x2c, 0x20, 0x7b, 0x20, 0x64, 0x65,
+  0x74, 0x61, 0x69, 0x6c, 0x3a, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x7d, 0x29, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74,
+  0x75, 0x72, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x54, 0x61, 0x72,
+  0x67, 0x65, 0x74, 0x3b, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f, 0x20, 0x43,
+  0x61, 0x6c, 0x6c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x61, 0x20, 0x70, 0x72, 0x6f, 0x6d,
+  0x69, 0x73, 0x65, 0x20, 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x73,
+  0x6f, 0x6c, 0x76, 0x65, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65,
+  0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0x64, 0x20, 0x74,
+  0x65, 0x78, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x64, 0x6f,
+  0x65, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x73, 0x75, 0x70, 0x70, 0x6f,
+  0x72, 0x74, 0x20, 0x73, 0x74, 0x72, 0x65, 0x61, 0x6d, 0x69, 0x6e, 0x67,
+  0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x45, 0x78, 0x61, 0x6d, 0x70,
+  0x6c, 0x65, 0x3a, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69,
+  0x73, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x2e, 0x74,
+  0x68, 0x65, 0x6e, 0x28, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+  0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74,
+  0x2e, 0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d,
+  0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x6f, 0x72, 0x0a, 0x2f, 0x2f, 0x0a, 0x2f, 0x2f, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65,
+  0x6e, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28,
+  0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x20, 0x20,
+  0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e,
+  0x77, 0x72, 0x69, 0x74, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x29, 0x0a, 0x2f, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74,
+  0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61,
+  0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x70,
+  0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d,
+  0x73, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66,
+  0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e,
+  0x65, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x69, 0x73, 0x65, 0x28, 0x61,
+  0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x72, 0x65, 0x73, 0x6f, 0x6c, 0x76,
+  0x65, 0x2c, 0x20, 0x72, 0x65, 0x6a, 0x65, 0x63, 0x74, 0x29, 0x20, 0x3d,
+  0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20,
+  0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x22, 0x22,
+  0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a,
+  0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77,
+  0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63,
+  0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d,
+  0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x70, 0x61,
+  0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67,
+  0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x2b, 0x3d, 0x20,
+  0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x63,
+  0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20,
+  0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65,
+  0x73, 0x6f, 0x6c, 0x76, 0x65, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+  0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61,
+  0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6a, 0x65,
+  0x63, 0x74, 0x28, 0x65, 0x72, 0x72, 0x6f, 0x72, 0x29, 0x3b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x7d,
+  0x3b, 0x0a, 0x0a, 0x2f, 0x2a, 0x2a, 0x0a, 0x20, 0x2a, 0x20, 0x28, 0x64,
+  0x65, 0x70, 0x72, 0x65, 0x63, 0x61, 0x74, 0x65, 0x64, 0x29, 0x0a, 0x20,
+  0x2a, 0x2f, 0x0a, 0x65, 0x78, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f,
+  0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x43, 0x6f, 0x6d,
+  0x70, 0x6c, 0x65, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e,
+  0x63, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63,
+  0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2c, 0x20, 0x63,
+  0x61, 0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x29, 0x20, 0x3d, 0x3e, 0x20,
+  0x7b, 0x0a, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69,
+  0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75,
+  0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28,
+  0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70,
+  0x74, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b,
+  0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20,
+  0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61,
+  0x6c, 0x6c, 0x62, 0x61, 0x63, 0x6b, 0x28, 0x63, 0x68, 0x75, 0x6e, 0x6b,
+  0x29, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x7d, 0x0a, 0x0a, 0x2f, 0x2f,
+  0x20, 0x47, 0x65, 0x74, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64,
+  0x65, 0x6c, 0x20, 0x69, 0x6e, 0x66, 0x6f, 0x20, 0x66, 0x72, 0x6f, 0x6d,
+  0x20, 0x74, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x2e,
+  0x20, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x75, 0x73, 0x65,
+  0x66, 0x75, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x67, 0x65, 0x74, 0x74,
+  0x69, 0x6e, 0x67, 0x20, 0x74, 0x68, 0x65, 0x20, 0x63, 0x6f, 0x6e, 0x74,
+  0x65, 0x78, 0x74, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x20, 0x61,
+  0x6e, 0x64, 0x20, 0x73, 0x6f, 0x20, 0x6f, 0x6e, 0x2e, 0x0a, 0x65, 0x78,
+  0x70, 0x6f, 0x72, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c,
+  0x6c, 0x61, 0x6d, 0x61, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x49, 0x6e, 0x66,
+  0x6f, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x63,
+  0x6f, 0x6e, 0x66, 0x69, 0x67, 0x20, 0x3d, 0x20, 0x7b, 0x7d, 0x29, 0x20,
+  0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21,
+  0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73,
+  0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20,
+  0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x61, 0x70, 0x69,
+  0x5f, 0x75, 0x72, 0x6c, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x66, 0x69,
+  0x67, 0x2e, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x20, 0x7c, 0x7c,
+  0x20, 0x22, 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e,
+  0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x20, 0x3d, 0x20, 0x61,
+  0x77, 0x61, 0x69, 0x74, 0x20, 0x66, 0x65, 0x74, 0x63, 0x68, 0x28, 0x60,
+  0x24, 0x7b, 0x61, 0x70, 0x69, 0x5f, 0x75, 0x72, 0x6c, 0x7d, 0x2f, 0x70,
+  0x72, 0x6f, 0x70, 0x73, 0x60, 0x29, 0x2e, 0x74, 0x68, 0x65, 0x6e, 0x28,
+  0x72, 0x20, 0x3d, 0x3e, 0x20, 0x72, 0x2e, 0x6a, 0x73, 0x6f, 0x6e, 0x28,
+  0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e,
+  0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x5f, 0x67, 0x65, 0x6e, 0x65,
+  0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x65, 0x74, 0x74, 0x69,
+  0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x72,
+  0x65, 0x74, 0x75, 0x72, 0x6e, 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 = 5909;

examples/server/deps.sh

@@ -8,3 +8,13 @@ PUBLIC=$DIR/public
 echo "download js bundle files"
 curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
 echo >> $PUBLIC/index.js # add newline
+
+FILES=$(ls $PUBLIC)
+
+cd $PUBLIC
+for FILE in $FILES; do
+  echo "generate $FILE.hpp"
+
+  # use simple flag for old version of xxd
+  xxd -i $FILE > $DIR/$FILE.hpp
+done

examples/server/public/index.html

@@ -51,6 +51,26 @@
       margin-bottom: 0.5em;
     }
 
+  button, input, textarea, .button, a.button, select {
+      color: #666;
+      border: 1px solid #ddd;
+      border-radius: 4px;
+      line-height: 1.5em;
+      padding: 0.25em 0.25em;
+      text-decoration: none;
+      font-size: 1.1rem;
+    }
+
+    button {
+      border: 1px solid #2a8aad;
+      background: #3584e4;
+      font-weight: normal;
+      color: #fff;
+    }
+    button:disabled {
+      background: #9cbce5;
+    }
+
     #write form {
       margin: 1em 0 0 0;
       display: flex;
@@ -567,7 +587,7 @@
         runCompletion();
       }
       return html`
-        <div>
+        <div class="right">
           <button onclick=${submit} type="button" disabled=${generating.value}>Start</button>
           <button onclick=${stop} disabled=${!generating.value}>Stop</button>
           <button onclick=${reset}>Reset</button>
@@ -881,11 +901,11 @@
         .replace(/&/g, '&amp;')
         .replace(/</g, '&lt;')
         .replace(/>/g, '&gt;')
-        .replace(/(^|\n)#{1,6} ([^\n]*)(?=([^`]*`[^`]*`)*[^`]*$)/g, '$1<h3>$2</h3>')
-        .replace(/\*\*(.*?)\*\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
-        .replace(/__(.*?)__(?=([^`]*`[^`]*`)*[^`]*$)/g, '<strong>$1</strong>')
-        .replace(/\*(.*?)\*(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
-        .replace(/_(.*?)_(?=([^`]*`[^`]*`)*[^`]*$)/g, '<em>$1</em>')
+        .replace(/^#{1,6} (.*)$/gim, '<h3>$1</h3>')
+        .replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
+        .replace(/__(.*?)__/g, '<strong>$1</strong>')
+        .replace(/\*(.*?)\*/g, '<em>$1</em>')
+        .replace(/_(.*?)_/g, '<em>$1</em>')
         .replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
         .replace(/`(.*?)`/g, '<code>$1</code>')
         .replace(/\n/gim, '<br />');

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

@@ -1,95 +1,33 @@
-// WARNING: This file was ported from json_schema_to_grammar.py, please fix bugs / add features there first.
+// WARNING: This file was ported from json-schema-to-grammar.py, please fix bugs / add features there first.
 const SPACE_RULE = '" "?';
 
-function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
-  const separatorRule = opts.separatorRule ?? '';
-  const itemRuleIsLiteral = opts.itemRuleIsLiteral ?? false
-
-  if (separatorRule === '') {
-    if (minItems === 0 && maxItems === 1) {
-      return `${itemRule}?`;
-    } else if (minItems === 1 && maxItems === undefined) {
-      return `${itemRule}+`;
-    }
-  }
-
-  let result = '';
-  if (minItems > 0) {
-    if (itemRuleIsLiteral && separatorRule === '') {
-      result = `"${itemRule.slice(1, -1).repeat(minItems)}"`;
-    } else {
-      result = Array.from({ length: minItems }, () => itemRule)
-        .join(separatorRule !== '' ? ` ${separatorRule} ` : ' ');
-    }
-  }
-
-  const optRepetitions = (upToN, prefixWithSep=false) => {
-    const content = separatorRule !== '' && prefixWithSep ? `${separatorRule} ${itemRule}` : itemRule;
-    if (upToN === 0) {
-      return '';
-    } else if (upToN === 1) {
-      return `(${content})?`;
-    } else if (separatorRule !== '' && !prefixWithSep) {
-      return `(${content} ${optRepetitions(upToN - 1, true)})?`;
-    } else {
-      return Array.from({ length: upToN }, () => `(${content}`).join(' ').trim() + Array.from({ length: upToN }, () => ')?').join('');
-    }
-  };
-
-  if (minItems > 0 && maxItems !== minItems) {
-    result += ' ';
-  }
-
-  if (maxItems !== undefined) {
-    result += optRepetitions(maxItems - minItems, minItems > 0);
-  } else {
-    const itemOperator = `(${separatorRule !== '' ? separatorRule + ' ' : ''}${itemRule})`;
-
-    if (minItems === 0 && separatorRule !== '') {
-      result = `(${itemRule} ${itemOperator}*)?`;
-    } else {
-      result += `${itemOperator}*`;
-    }
-  }
-
-  return result;
-}
-
-class BuiltinRule {
-  constructor(content, deps) {
-    this.content = content;
-    this.deps = deps || [];
-  }
-}
-
-const UP_TO_15_DIGITS = _buildRepetition('[0-9]', 0, 15);
-
 const PRIMITIVE_RULES = {
-  boolean        : new BuiltinRule('("true" | "false") space', []),
-  'decimal-part' : new BuiltinRule('[0-9] ' + UP_TO_15_DIGITS, []),
-  'integral-part': new BuiltinRule('[0-9] | [1-9] ' + UP_TO_15_DIGITS, []),
-  number         : new BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
-  integer        : new BuiltinRule('("-"? integral-part) space', ['integral-part']),
-  value          : new BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
-  object         : new BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
-  array          : new BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
-  uuid           : new BuiltinRule('"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space', []),
-  char           : new BuiltinRule(`[^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])`, []),
-  string         : new BuiltinRule(`"\\"" char* "\\"" space`, ['char']),
-  null           : new BuiltinRule('"null" space', []),
+  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 STRING_FORMAT_RULES = {
-  'date'            : new BuiltinRule('[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'            : new BuiltinRule('([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'       : new BuiltinRule('date "T" time', ['date', 'time']),
-  'date-string'     : new BuiltinRule('"\\"" date "\\"" space', ['date']),
-  'time-string'     : new BuiltinRule('"\\"" time "\\"" space', ['time']),
-  'date-time-string': new BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
-}
+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, ...STRING_FORMAT_RULES};
+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;
@@ -220,7 +158,7 @@ export class SchemaConverter {
         rule = '[\\U00000000-\\U0010FFFF]';
       } else {
         // Accept any character... except \n and \r line break chars (\x0A and \xOD)
-        rule = '[^\\x0A\\x0D]';
+        rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]';
       }
       return this._addRule('dot', rule);
     };
@@ -321,19 +259,26 @@ export class SchemaConverter {
 
           let [sub, subIsLiteral] = seq[seq.length - 1];
 
-          if (!subIsLiteral) {
-            let id = subRuleIds[sub];
-            if (id === undefined) {
-              id = this._addRule(`${name}-${Object.keys(subRuleIds).length + 1}`, sub);
-              subRuleIds[sub] = id;
+          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;
             }
-            sub = id;
-          }
 
-          seq[seq.length - 1] = [
-            _buildRepetition(subIsLiteral ? `"${sub}"` : sub, minTimes, maxTimes, {itemRuleIsLiteral: subIsLiteral}),
-            false
-          ];
+            const repeatedSub = Array.from({ length: minTimes }, () => subIsLiteral ? `"${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) {
@@ -449,50 +394,49 @@ export class SchemaConverter {
         );
       } else {
         const itemRuleName = this.visit(items, `${name ?? ''}${name ? '-' : ''}item`);
-        const minItems = schema.minItems || 0;
+        const listItemOperator = `( "," space ${itemRuleName} )`;
+        let successiveItems = '';
+        let minItems = schema.minItems || 0;
         const maxItems = schema.maxItems;
-        return this._addRule(ruleName, '"[" space ' + _buildRepetition(itemRuleName, minItems, maxItems, {separatorRule: '"," space'}) + ' "]" space');
+        if (minItems > 0) {
+          successiveItems = listItemOperator.repeat(minItems - 1);
+          minItems--;
+        }
+        if (maxItems !== undefined && maxItems > minItems) {
+          successiveItems += `${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._addPrimitive(
-        ruleName === 'root' ? 'root' : schemaFormat,
-        PRIMITIVE_RULES['uuid']
-      );
-    } else if ((schemaType === undefined || schemaType === 'string') && `${schema.format}-string` in STRING_FORMAT_RULES) {
-      const primName = `${schema.format}-string`
-      return this._addRule(ruleName, this._addPrimitive(primName, STRING_FORMAT_RULES[primName]));
-    } else if (schemaType === 'string' && ('minLength' in schema || 'maxLength' in schema)) {
-      const charRuleName = this._addPrimitive('char', PRIMITIVE_RULES['char']);
-      const minLen = schema.minLength || 0;
-      const maxLen = schema.maxLength;
-      return this._addRule(ruleName, '"\\\"" ' + _buildRepetition(charRuleName, minLen, maxLen) + ' "\\\"" space');
+      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)) {
-      return this._addRule(ruleName, this._addPrimitive('object', PRIMITIVE_RULES['object']));
+      for (const n of OBJECT_RULE_NAMES) {
+        this._addRule(n, PRIMITIVE_RULES[n]);
+      }
+      return this._addRule(ruleName, 'object');
     } else {
       if (!(schemaType in PRIMITIVE_RULES)) {
         throw new Error(`Unrecognized schema: ${JSON.stringify(schema)}`);
       }
       // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
-      return this._addPrimitive(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]);
+      return this._addRule(ruleName === 'root' ? 'root' : schemaType, PRIMITIVE_RULES[schemaType]);
     }
   }
 
-  _addPrimitive(name, rule) {
-    let n = this._addRule(name, rule.content);
-    for (const dep of rule.deps) {
-      const depRule = PRIMITIVE_RULES[dep] || STRING_FORMAT_RULES[dep];
-      if (!depRule) {
-        throw new Error(`Rule ${dep} not known`);
-      }
-      if (!(dep in this._rules)) {
-        this._addPrimitive(dep, depRule);
-      }
-    }
-    return n;
-  }
-
   _buildObjectRule(properties, required, name, additionalProperties) {
     const propOrder = this._propOrder;
     // sort by position in prop_order (if specified) then by original order
@@ -518,7 +462,7 @@ export class SchemaConverter {
       const valueRule = this.visit(additionalProperties === true ? {} : additionalProperties, `${subName}-value`);
       propKvRuleNames['*'] = this._addRule(
         `${subName}-kv`,
-        `${this._addPrimitive('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
+        `${this._addRule('string', PRIMITIVE_RULES['string'])} ":" space ${valueRule}`);
       optionalProps.push('*');
     }
 

examples/server/server.cpp

@@ -854,12 +854,12 @@ 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.n_discard          = json_value(data, "n_discard",         default_params.n_discard);
-        slot.sparams.seed              = json_value(data, "seed",              default_sparams.seed);
+        slot.params.seed               = json_value(data, "seed",              default_params.seed);
         slot.sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot.sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
 
         // process "json_schema" and "grammar"
-        if (data.contains("json_schema") && !data["json_schema"].is_null() && data.contains("grammar") && !data["grammar"].is_null()) {
+        if (data.contains("json_schema") && data.contains("grammar")) {
             send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST);
             return false;
         } else if (data.contains("json_schema") && !data.contains("grammar")) {
@@ -1028,6 +1028,7 @@ struct server_context {
                 send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
+            llama_set_rng_seed(ctx, slot.params.seed);
         }
 
         slot.command = SLOT_COMMAND_LOAD_PROMPT;
@@ -1082,7 +1083,7 @@ struct server_context {
                 };
 
                 if (llama_decode(ctx, batch_view) != 0) {
-                    LOG_ERROR("llama_decode() failed", {});
+                    LOG_TEE("%s: llama_decode() failed\n", __func__);
                     return;
                 }
             }
@@ -1117,7 +1118,7 @@ struct server_context {
 
     bool process_token(completion_token_output & result, server_slot & slot) {
         // remember which tokens were sampled - used for repetition penalties during sampling
-        const std::string token_str = llama_token_to_piece(ctx, result.tok, false);
+        const std::string token_str = llama_token_to_piece(ctx, result.tok);
         slot.sampled = result.tok;
 
         // search stop word and delete it
@@ -1200,7 +1201,7 @@ struct server_context {
             });
         }
 
-        if (llama_token_is_eog(model, result.tok)) {
+        if (result.tok == llama_token_eos(model)) {
             slot.stopped_eos    = true;
             slot.has_next_token = false;
 
@@ -1280,11 +1281,7 @@ struct server_context {
     }
 
     void send_error(const int id_task, const int id_multi, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) {
-        LOG_ERROR("task error", {
-            {"id_multi", id_multi},
-            {"id_task", id_task},
-            {"error", error},
-        });
+        LOG_TEE("task %i - error: %s\n", id_task, error.c_str());
 
         server_task_result res;
         res.id       = id_task;
@@ -2189,11 +2186,7 @@ struct server_context {
             if (ret != 0) {
                 if (n_batch == 1 || ret < 0) {
                     // if you get here, it means the KV cache is full - try increasing it via the context size
-                    LOG_ERROR("failed to decode the batch: KV cache is full - try increasing it via the context size", {
-                        {"i",   i},
-                        {"n_batch",  ret},
-                        {"ret",   ret},
-                    });
+                    LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
                     for (auto & slot : slots) {
                         slot.state = SLOT_STATE_PROCESSING;
                         slot.command = SLOT_COMMAND_NONE;
@@ -2203,16 +2196,12 @@ struct server_context {
                     break; // break loop of n_batch
                 }
 
+                LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2);
+
                 // retry with half the batch size to try to find a free slot in the KV cache
                 n_batch /= 2;
                 i -= n_batch;
 
-                LOG_WARNING("failed to find free space in the KV cache, retrying with smaller batch size - try increasing it via the context size or enable defragmentation", {
-                    {"i",   i},
-                    {"n_batch",  n_batch},
-                    {"ret",   ret},
-                });
-
                 continue; // continue loop of n_batch
             }
 

examples/server/tests/README.md

@@ -29,7 +29,7 @@ To mitigate it, you can increase values in `n_predict`, `kv_size`.
 cd ../../..
 mkdir build
 cd build
-cmake -DLLAMA_CURL=ON ../
+cmake ../
 cmake --build . --target server
 ```
 

examples/server/tests/features/results.feature

@@ -1,57 +0,0 @@
-@llama.cpp
-@results
-Feature: Results
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
-    And   a model file test-model-00001-of-00003.gguf
-    And   128 as batch size
-    And   256 KV cache size
-    And   128 max tokens to predict
-
-  Scenario Outline: Multi users completion
-    Given <n_slots> slots
-    And   continuous batching
-    Then  the server is starting
-    Then  the server is healthy
-
-    Given 42 as seed
-    And a prompt:
-      """
-      Write a very long story about AI.
-      """
-
-    Given 42 as seed
-    And a prompt:
-      """
-      Write a very long story about AI.
-      """
-
-    Given 42 as seed
-    And a prompt:
-      """
-      Write a very long story about AI.
-      """
-
-    Given 42 as seed
-    And a prompt:
-      """
-      Write a very long story about AI.
-      """
-
-    Given 42 as seed
-    And a prompt:
-      """
-      Write a very long story about AI.
-      """
-
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all predictions are equal
-    Examples:
-      | n_slots |
-      | 1       |
-      | 2       |

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

@@ -61,7 +61,6 @@ def step_server_config(context, server_fqdn, server_port):
     context.server_metrics = False
     context.server_process = None
     context.seed = None
-    context.draft = None
     context.server_seed = None
     context.user_api_key = None
     context.response_format = None
@@ -108,11 +107,6 @@ def step_n_gpu_layer(context, ngl):
     context.n_gpu_layer = ngl
 
 
-@step('{draft:d} as draft')
-def step_draft(context, draft):
-    context.draft = draft
-
-
 @step('{n_ctx:d} KV cache size')
 def step_n_ctx(context, n_ctx):
     context.n_ctx = n_ctx
@@ -260,15 +254,6 @@ def step_n_tokens_predicted(context, predicted_n):
     assert_n_tokens_predicted(context.completion, predicted_n)
 
 
-@step('all predictions are equal')
-@async_run_until_complete
-async def step_predictions_equal(context):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions >= 2, "need at least 2 completions"
-    assert_all_predictions_equal(context.tasks_result)
-    context.tasks_result = []
-
-
 @step('the completion is  truncated')
 def step_assert_completion_truncated(context):
     step_assert_completion_truncated(context, '')
@@ -1035,23 +1020,6 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
         assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
                                                      f' {n_predicted} <> {expected_predicted_n}')
 
-def assert_all_predictions_equal(completion_responses):
-    content_0 = completion_responses[0]['content']
-
-    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-        print(f"content 0: {content_0}")
-
-    i = 1
-    for response in completion_responses[1:]:
-        content = response['content']
-
-        if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-            print(f"content {i}: {content}")
-
-        assert content == content_0, "contents not equal"
-
-        i += 1
-
 
 async def gather_tasks_results(context):
     n_tasks = len(context.concurrent_tasks)
@@ -1180,8 +1148,6 @@ def start_server_background(context):
         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.draft is not None:
-        server_args.extend(['--draft', context.draft])
     if context.server_continuous_batching:
         server_args.append('--cont-batching')
     if context.server_embeddings:

examples/server/tests/tests.sh

@@ -4,8 +4,9 @@ set -eu
 
 if [ $# -lt 1 ]
 then
-    # Start @llama.cpp scenario
-    behave --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
+  # Start @llama.cpp scenario
+  behave --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 else
-    behave "$@"
+  behave "$@"
 fi
+

examples/server/utils.hpp

@@ -381,6 +381,10 @@ static json oaicompat_completion_params_parse(
     } else {
         llama_params["stop"] = json_value(body, "stop", json::array());
     }
+    // Some chat templates don't use EOS token to stop generation
+    // We must add their end sequences to list of stop words
+    llama_params["stop"].push_back("<|im_end|>"); // chatml
+    llama_params["stop"].push_back("<end_of_turn>"); // gemma
 
     // Handle "response_format" field
     if (body.contains("response_format")) {

examples/simple/simple.cpp

@@ -133,8 +133,8 @@ int main(int argc, char ** argv) {
             // sample the most likely token
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
-            // is it an end of generation?
-            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;

examples/speculative/speculative.cpp

@@ -360,7 +360,7 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                if (llama_token_is_eog(model_tgt, token_id)) {
+                if (token_id == llama_token_eos(model_tgt)) {
                     has_eos = true;
                 }
                 ++n_predict;

examples/sycl/build.sh

@@ -20,4 +20,4 @@ cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 #cmake --build . --config Release --target llama-bench
 
 #build all binary
-cmake --build . --config Release -j -v
+cmake --build . --config Release -v

examples/sycl/run-llama2.sh

@@ -12,7 +12,6 @@ if [ $# -gt 0 ]; then
     GGML_SYCL_SINGLE_GPU=1
 else
     GGML_SYCL_DEVICE=0
-    GGML_SYCL_SINGLE_GPU=0
 fi
 
 #export GGML_SYCL_DEBUG=1

examples/test-bench/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET test-bench)
+add_executable(${TARGET} test-bench.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common)
+#target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/test-bench/README.md

@@ -0,0 +1,278 @@
+# llama.cpp/example/llama-bench
+
+Performance testing tool for llama.cpp.
+
+## Table of contents
+
+1. [Syntax](#syntax)
+2. [Examples](#examples)
+    1. [Text generation with different models](#text-generation-with-different-models)
+    2. [Prompt processing with different batch sizes](#prompt-processing-with-different-batch-sizes)
+    3. [Different numbers of threads](#different-numbers-of-threads)
+    4. [Different numbers of layers offloaded to the GPU](#different-numbers-of-layers-offloaded-to-the-gpu)
+3. [Output formats](#output-formats)
+    1. [Markdown](#markdown)
+    2. [CSV](#csv)
+    3. [JSON](#json)
+    4. [SQL](#sql)
+
+## Syntax
+
+```
+usage: ./llama-bench [options]
+
+options:
+  -h, --help
+  -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                  (default: 512)
+  -n, --n-gen <n>                     (default: 128)
+  -b, --batch-size <n>                (default: 512)
+  -ctk <t>, --cache-type-k <t>        (default: f16)
+  -ctv <t>, --cache-type-v <t>        (default: f16)
+  -t, --threads <n>                   (default: 112)
+  -ngl, --n-gpu-layers <n>            (default: 99)
+  -sm, --split-mode <none|layer|row>  (default: layer)
+  -mg, --main-gpu <i>                 (default: 0)
+  -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -mmp, --mmap <0|1>                  (default: 1)
+  -ts, --tensor_split <ts0/ts1/..>    (default: 0)
+  -r, --repetitions <n>               (default: 5)
+  -o, --output <csv|json|md|sql>      (default: md)
+  -v, --verbose                       (default: 0)
+
+Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
+```
+
+llama-bench can perform two types of tests:
+
+- Prompt processing (pp): processing a prompt in batches (`-p`)
+- Text generation (tg): generating a sequence of tokens (`-n`)
+
+With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
+
+Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
+
+For a description of the other options, see the [main example](../main/README.md).
+
+Note:
+
+- When using SYCL backend, there would be hang issue in some cases. Please set `--mmp 0`.
+
+## Examples
+
+### Text generation with different models
+
+```sh
+$ ./llama-bench -m models/7B/ggml-model-q4_0.gguf -m models/13B/ggml-model-q4_0.gguf -p 0 -n 128,256,512
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    132.19 ± 0.55 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 256     |    129.37 ± 0.54 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 512     |    123.83 ± 0.25 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 128     |     82.17 ± 0.31 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 256     |     80.74 ± 0.23 |
+| llama 13B mostly Q4_0          |   6.86 GiB |    13.02 B | CUDA       |  99 | tg 512     |     78.08 ± 0.07 |
+
+### Prompt processing with different batch sizes
+
+```sh
+$ ./llama-bench -n 0 -p 1024 -b 128,256,512,1024
+```
+
+| model                          |       size |     params | backend    | ngl |    n_batch | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        128 | pp 1024    |   1436.51 ± 3.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        256 | pp 1024    |  1932.43 ± 23.48 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |        512 | pp 1024    |  2254.45 ± 15.59 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 |       1024 | pp 1024    |  2498.61 ± 13.58 |
+
+### Different numbers of threads
+
+```sh
+$ ./llama-bench -n 0 -n 16 -p 64 -t 1,2,4,8,16,32
+```
+
+| model                          |       size |     params | backend    |    threads | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ---------: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | pp 64      |      6.17 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          1 | tg 16      |      4.05 ± 0.02 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | pp 64      |     12.31 ± 0.13 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          2 | tg 16      |      7.80 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | pp 64      |     23.18 ± 0.06 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          4 | tg 16      |     12.22 ± 0.07 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | pp 64      |     32.29 ± 1.21 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |          8 | tg 16      |     16.71 ± 0.66 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | pp 64      |     33.52 ± 0.03 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         16 | tg 16      |     15.32 ± 0.05 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | pp 64      |     59.00 ± 1.11 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CPU        |         32 | tg 16      |     16.41 ± 0.79 ||
+
+### Different numbers of layers offloaded to the GPU
+
+```sh
+$ ./llama-bench -ngl 10,20,30,31,32,33,34,35
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | pp 512     |    373.36 ± 2.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  10 | tg 128     |     13.45 ± 0.93 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | pp 512     |    472.65 ± 1.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  20 | tg 128     |     21.36 ± 1.94 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | pp 512     |   631.87 ± 11.25 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  30 | tg 128     |     40.04 ± 1.82 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | pp 512     |    657.89 ± 5.08 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  31 | tg 128     |     48.19 ± 0.81 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | pp 512     |    688.26 ± 3.29 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  32 | tg 128     |     54.78 ± 0.65 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | pp 512     |    704.27 ± 2.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  33 | tg 128     |     60.62 ± 1.76 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | pp 512     |    881.34 ± 5.40 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  34 | tg 128     |     71.76 ± 0.23 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
+
+## Output formats
+
+By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
+
+### Markdown
+
+```sh
+$ ./llama-bench -o md
+```
+
+| model                          |       size |     params | backend    | ngl | test       |              t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ---------- | ---------------: |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | pp 512     |  2368.80 ± 93.24 |
+| llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  99 | tg 128     |    131.42 ± 0.59 |
+
+### CSV
+
+```sh
+$ ./llama-bench -o csv
+```
+
+```csv
+build_commit,build_number,cuda,opencl,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
+"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+```
+
+### JSON
+
+```sh
+$ ./llama-bench -o json
+```
+
+```json
+[
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 512,
+    "n_gen": 0,
+    "test_time": "2023-09-23T12:09:57Z",
+    "avg_ns": 212365953,
+    "stddev_ns": 985423,
+    "avg_ts": 2410.974041,
+    "stddev_ts": 11.163766,
+    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
+    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+  },
+  {
+    "build_commit": "3469684",
+    "build_number": 1275,
+    "cuda": true,
+    "opencl": false,
+    "metal": false,
+    "gpu_blas": true,
+    "blas": true,
+    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
+    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
+    "model_filename": "models/7B/ggml-model-q4_0.gguf",
+    "model_type": "llama 7B mostly Q4_0",
+    "model_size": 3825065984,
+    "model_n_params": 6738415616,
+    "n_batch": 512,
+    "n_threads": 16,
+    "f16_kv": true,
+    "n_gpu_layers": 99,
+    "main_gpu": 0,
+    "mul_mat_q": true,
+    "tensor_split": "0.00",
+    "n_prompt": 0,
+    "n_gen": 128,
+    "test_time": "2023-09-23T12:09:59Z",
+    "avg_ns": 977425219,
+    "stddev_ns": 9268593,
+    "avg_ts": 130.965708,
+    "stddev_ts": 1.238924,
+    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
+    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+  }
+]
+```
+
+### SQL
+
+SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.
+
+```sh
+$ ./llama-bench -o sql
+```
+
+```sql
+CREATE TABLE IF NOT EXISTS test (
+  build_commit TEXT,
+  build_number INTEGER,
+  cuda INTEGER,
+  opencl INTEGER,
+  metal INTEGER,
+  gpu_blas INTEGER,
+  blas INTEGER,
+  cpu_info TEXT,
+  gpu_info TEXT,
+  model_filename TEXT,
+  model_type TEXT,
+  model_size INTEGER,
+  model_n_params INTEGER,
+  n_batch INTEGER,
+  n_threads INTEGER,
+  f16_kv INTEGER,
+  n_gpu_layers INTEGER,
+  main_gpu INTEGER,
+  mul_mat_q INTEGER,
+  tensor_split TEXT,
+  n_prompt INTEGER,
+  n_gen INTEGER,
+  test_time TEXT,
+  avg_ns INTEGER,
+  stddev_ns INTEGER,
+  avg_ts REAL,
+  stddev_ts REAL
+);
+
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
+INSERT INTO test (build_commit, build_number, cuda, opencl, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+```

examples/test-bench/helpers.hpp

@@ -0,0 +1,196 @@
+
+static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
+    // static RNG initialization (revisit if n_threads stops being constant)
+    static const size_t n_threads = std::thread::hardware_concurrency();
+    static std::vector<std::default_random_engine> generators = []() {
+        std::random_device rd;
+        std::vector<std::default_random_engine> vec;
+        vec.reserve(n_threads);
+        //for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(1234 + i); } // fixed seed
+        for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(rd()); }
+        return vec;
+    }();
+
+    size_t size = ggml_nelements(tensor);
+    std::vector<float> data(size);
+
+    auto init_thread = [&](size_t ith, size_t start, size_t end) {
+        std::uniform_real_distribution<float> distribution(min, max);
+        for (size_t i = start; i < end; i++) {
+            data[i] = distribution(generators[ith]);
+        }
+    };
+
+    std::vector<std::thread> threads;
+    threads.reserve(n_threads);
+    for (size_t i = 0; i < n_threads; i++) {
+        size_t start =     i*size/n_threads;
+        size_t end   = (i+1)*size/n_threads;
+        threads.emplace_back(init_thread, i, start, end);
+    }
+    for (auto & t : threads) {
+        t.join();
+    }
+
+    if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
+        ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
+    } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) {
+        GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
+        std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
+        std::vector<float> imatrix(tensor->ne[0], 1.0f); // dummy importance matrix
+        const float * im = imatrix.data();
+        if (!ggml_quantize_requires_imatrix(tensor->type)) {
+            // when the imatrix is optional, we want to test both quantization with and without imatrix
+            // use one of the random numbers to decide
+            if (data[0] > 0.5f*(min + max)) {
+                im = nullptr;
+            }
+        }
+        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], im);
+        ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
+    } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
+        // This is going to create some weird integers though.
+        ggml_backend_tensor_set(tensor, data.data(), 0, ggml_nbytes(tensor));
+    } else {
+        GGML_ASSERT(false);
+    }
+}
+
+static std::vector<float> tensor_to_float(const ggml_tensor * t) {
+    std::vector<float> tv;
+    tv.reserve(ggml_nelements(t));
+
+    std::vector<uint8_t> buf(ggml_nbytes(t));
+    ggml_backend_tensor_get(t, buf.data(), 0, ggml_nbytes(t));
+
+    ggml_type_traits_t tt = ggml_internal_get_type_traits(t->type);
+    size_t bs = ggml_blck_size(t->type);
+    std::vector<float> vq(ggml_blck_size(t->type));
+    bool quantized = ggml_is_quantized(t->type);
+
+    // access elements by index to avoid gaps in views
+    for (int64_t i3 = 0; i3 < t->ne[3]; i3++) {
+        for (int64_t i2 = 0; i2 < t->ne[2]; i2++) {
+            for (int64_t i1 = 0; i1 < t->ne[1]; i1++) {
+                for (int64_t i0 = 0; i0 < t->ne[0]; i0 += bs) {
+                    size_t i = i3*t->nb[3] + i2*t->nb[2] + i1*t->nb[1] + i0/bs*t->nb[0];
+                    if (t->type == GGML_TYPE_F16) {
+                        tv.push_back(ggml_fp16_to_fp32(*(ggml_fp16_t*)&buf[i]));
+                    } else if (t->type == GGML_TYPE_F32) {
+                        tv.push_back(*(float *) &buf[i]);
+                    } else if (t->type == GGML_TYPE_I32) {
+                        tv.push_back((float)*(int32_t *) &buf[i]);
+                    } else if (t->type == GGML_TYPE_I16) {
+                        tv.push_back((float)*(int16_t *) &buf[i]);
+                    } else if (t->type == GGML_TYPE_I8) {
+                        tv.push_back((float)*(int8_t *) &buf[i]);
+                    } else if (quantized) {
+                        tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
+                        tv.insert(tv.end(), vq.begin(), vq.end());
+                    } else {
+                        GGML_ASSERT(false);
+                    }
+                }
+            }
+        }
+    }
+
+    return tv;
+}
+
+// normalized mean squared error = mse(a, b) / mse(a, 0)
+static double nmse(const float * a, const float * b, size_t n) {
+    double mse_a_b = 0.0;
+    double mse_a_0 = 0.0;
+
+    for (size_t i = 0; i < n; i++) {
+        float a_i = a[i];
+        float b_i = b[i];
+
+        mse_a_b += (a_i - b_i) * (a_i - b_i);
+        mse_a_0 += a_i * a_i;
+    }
+
+    return mse_a_b / mse_a_0;
+}
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// utils for printing the variables of the test cases
+#define VAR_TO_STR(x) (#x "=" + var_to_str(x))
+
+template<typename T>
+static std::string var_to_str(const T & x) {
+    return std::to_string(x);
+}
+
+template<typename T, size_t N>
+static std::string var_to_str(const T (&x)[N]) {
+    std::string s = "[";
+    for (size_t i = 0; i < N; i++) {
+        if (i > 0) {
+            s += ",";
+        }
+        s += var_to_str(x[i]);
+    }
+    s += "]";
+    return s;
+}
+
+template<typename T, size_t N>
+static std::string var_to_str(const std::array<T, N> & x) {
+    std::string s = "[";
+    for (size_t i = 0; i < N; i++) {
+        if (i > 0) {
+            s += ",";
+        }
+        s += var_to_str(x[i]);
+    }
+    s += "]";
+    return s;
+}
+
+//static std::string var_to_str(ggml_unary_op unary_op) {
+//    return ggml_unary_op_name(unary_op);
+//}
+
+static std::string var_to_str(ggml_type type) {
+    return ggml_type_name(type);
+}
+
+static std::string var_to_str(ggml_op_pool pool) {
+    switch (pool) {
+        case GGML_OP_POOL_AVG:  return "avg";
+        case GGML_OP_POOL_MAX:  return "max";
+        default:                return std::to_string(pool);
+    }
+}
+
+#define VARS_TO_STR1(a) VAR_TO_STR(a)
+#define VARS_TO_STR2(a, b) VAR_TO_STR(a) + "," + VAR_TO_STR(b)
+#define VARS_TO_STR3(a, b, c) VAR_TO_STR(a) + "," + VARS_TO_STR2(b, c)
+#define VARS_TO_STR4(a, b, c, d) VAR_TO_STR(a) + "," + VARS_TO_STR3(b, c, d)
+#define VARS_TO_STR5(a, b, c, d, e) VAR_TO_STR(a) + "," + VARS_TO_STR4(b, c, d, e)
+#define VARS_TO_STR6(a, b, c, d, e, f) VAR_TO_STR(a) + "," + VARS_TO_STR5(b, c, d, e, f)
+#define VARS_TO_STR7(a, b, c, d, e, f, g) VAR_TO_STR(a) + "," + VARS_TO_STR6(b, c, d, e, f, g)
+#define VARS_TO_STR8(a, b, c, d, e, f, g, h) VAR_TO_STR(a) + "," + VARS_TO_STR7(b, c, d, e, f, g, h)
+#define VARS_TO_STR9(a, b, c, d, e, f, g, h, i) VAR_TO_STR(a) + "," + VARS_TO_STR8(b, c, d, e, f, g, h, i)
+#define VARS_TO_STR10(a, b, c, d, e, f, g, h, i, j) VAR_TO_STR(a) + "," + VARS_TO_STR9(b, c, d, e, f, g, h, i, j)
+#define VARS_TO_STR11(a, b, c, d, e, f, g, h, i, j, k) VAR_TO_STR(a) + "," + VARS_TO_STR10(b, c, d, e, f, g, h, i, j, k)
+#define VARS_TO_STR12(a, b, c, d, e, f, g, h, i, j, k, l) VAR_TO_STR(a) + "," + VARS_TO_STR11(b, c, d, e, f, g, h, i, j, k, l)
+
+#ifdef GGML_USE_SYCL
+static bool inline _isinf(float f) {
+    return (*(uint32_t *)&f & 0x7fffffff) == 0x7f800000;
+}
+#else
+static bool inline _isinf(float f) { return std::isinf(f); }
+#endif
+
+// accept FLT_MAX as infinity
+static bool isinf_or_max(float f) {
+    return _isinf(f) || f == FLT_MAX || f == -FLT_MAX;
+}
+

examples/test-bench/test-bench.cpp

@@ -0,0 +1,443 @@
+#include <ggml.h>
+#include <ggml-alloc.h>
+#include <ggml-backend.h>
+#include <ggml-backend-impl.h>
+
+#include <iostream>
+#include <algorithm>
+#include <array>
+#include <cfloat>
+#include <cstring>
+#include <functional>
+#include <memory>
+#include <random>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <thread>
+#include <vector>
+
+#include "helpers.hpp"
+
+enum test_mode {
+    MODE_TEST,
+    MODE_PERF,
+};
+
+struct test_case {
+    virtual ~test_case() {}
+
+    virtual std::string op_desc(ggml_tensor * t) {
+        return ggml_op_desc(t);
+    }
+
+    virtual std::string vars() {
+        return "";
+    }
+
+    virtual ggml_tensor * build_graph(ggml_context * ctx) = 0;
+
+    virtual double max_nmse_err() {
+        return 1e-7;
+    }
+
+    virtual void initialize_tensors(ggml_context * ctx) {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t);
+        }
+    }
+
+    virtual size_t op_size(ggml_tensor * t) {
+        size_t size = ggml_nbytes(t);
+        // add source tensors
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (t->src[i] != NULL) {
+                size += ggml_nbytes(t->src[i]);
+            }
+        }
+        return size;
+    }
+
+    ggml_cgraph * gf = nullptr;
+
+    static const int sentinel_size = 1024;
+
+    test_mode mode;
+
+    std::vector<ggml_tensor *> sentinels;
+
+    void add_sentinel(ggml_context * ctx) {
+        if (mode == MODE_PERF) {
+            return;
+        }
+        ggml_tensor * sentinel = ::ggml_new_tensor_1d(ctx, GGML_TYPE_F32, sentinel_size);
+        ggml_format_name(sentinel, "sent_%zu", sentinels.size());
+        sentinels.push_back(sentinel);
+    }
+
+    // hijack ggml_new_tensor to add sentinels after each tensor to check for overflows in the backend
+
+    ggml_tensor * ggml_new_tensor(ggml_context * ctx, ggml_type type, int n_dims, const int64_t * ne) {
+        ggml_tensor * t = ::ggml_new_tensor(ctx, type, n_dims, ne);
+        add_sentinel(ctx);
+        return t;
+    }
+
+    ggml_tensor * ggml_new_tensor_1d(ggml_context * ctx, ggml_type type, int64_t ne0) {
+        ggml_tensor * t = ::ggml_new_tensor_1d(ctx, type, ne0);
+        add_sentinel(ctx);
+        return t;
+    }
+
+    ggml_tensor * ggml_new_tensor_2d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1) {
+        ggml_tensor * t = ::ggml_new_tensor_2d(ctx, type, ne0, ne1);
+        add_sentinel(ctx);
+        return t;
+    }
+
+    ggml_tensor * ggml_new_tensor_3d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2) {
+        ggml_tensor * t = ::ggml_new_tensor_3d(ctx, type, ne0, ne1, ne2);
+        add_sentinel(ctx);
+        return t;
+    }
+
+    ggml_tensor * ggml_new_tensor_4d(ggml_context * ctx, ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+        ggml_tensor * t = ::ggml_new_tensor_4d(ctx, type, ne0, ne1, ne2, ne3);
+        add_sentinel(ctx);
+        return t;
+    }
+
+    bool eval(ggml_backend_t backend1, ggml_backend_t backend2, const char * op_name) {
+        mode = MODE_TEST;
+
+        ggml_init_params params = {
+            /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
+            /* .mem_base = */ NULL,
+            /* .no_alloc = */ true,
+        };
+        ggml_context * ctx = ggml_init(params);
+
+        gf = ggml_new_graph(ctx);
+
+        // pre-graph sentinel
+        add_sentinel(ctx);
+
+        ggml_tensor * out = build_graph(ctx);
+
+        if (op_name != nullptr && op_desc(out) != op_name) {
+            //printf("  %s: skipping\n", op_desc(out).c_str());
+            ggml_free(ctx);
+            return true;
+        }
+
+        printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
+        fflush(stdout);
+
+        // check if the backends support the ops
+        bool supported = true;
+        for (ggml_backend_t backend : {backend1, backend2}) {
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+                if (!ggml_backend_supports_op(backend, t)) {
+                    printf("not supported [%s] ", ggml_backend_name(backend));
+                    supported = false;
+                    break;
+                }
+            }
+        }
+        if (!supported) {
+            printf("\n");
+            ggml_free(ctx);
+            return true;
+        }
+
+        // post-graph sentinel
+        add_sentinel(ctx);
+
+        // allocate
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
+        if (buf == NULL) {
+            printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
+            ggml_free(ctx);
+            return false;
+        }
+
+        // build graph
+        ggml_build_forward_expand(gf, out);
+
+        // add sentinels as graph nodes so that they are checked in the callback
+        for (ggml_tensor * sentinel : sentinels) {
+            gf->nodes[gf->n_nodes++] = sentinel;
+        }
+
+        // randomize tensors
+        initialize_tensors(ctx);
+
+        // compare
+        struct callback_userdata {
+            bool   ok;
+            double max_err;
+            ggml_backend_t backend1;
+            ggml_backend_t backend2;
+        };
+
+        callback_userdata ud {
+            true,
+            max_nmse_err(),
+            backend1,
+            backend2
+        };
+
+        auto callback = [](int index, ggml_tensor * t1, ggml_tensor * t2, void * user_data) -> bool {
+            callback_userdata * ud = (callback_userdata *) user_data;
+            const char * bn1 = ggml_backend_name(ud->backend1);
+            const char * bn2 = ggml_backend_name(ud->backend2);
+
+            if (t1->op == GGML_OP_NONE) {
+                // sentinels must be unchanged
+                std::vector<uint8_t> t1_data(ggml_nbytes(t1));
+                std::vector<uint8_t> t2_data(ggml_nbytes(t2));
+                ggml_backend_tensor_get(t1, t1_data.data(), 0, ggml_nbytes(t1));
+                ggml_backend_tensor_get(t2, t2_data.data(), 0, ggml_nbytes(t2));
+
+                if (memcmp(t1_data.data(), t2_data.data(), ggml_nbytes(t1)) != 0) {
+                    printf("sentinel mismatch: %s ", t1->name);
+                    ud->ok = false;
+                    return true;
+                }
+            }
+
+            std::vector<float> f1 = tensor_to_float(t1);
+            std::vector<float> f2 = tensor_to_float(t2);
+
+            for (size_t i = 0; i < f1.size(); i++) {
+                // check for nans
+                if (std::isnan(f1[i]) || std::isnan(f2[i])) {
+                    printf("[%s] NaN at index %zu (%s=%f %s=%f) ", ggml_op_desc(t1), i, bn1, f1[i], bn2, f2[i]);
+                    ud->ok = false;
+                    return true;
+                }
+                // check for infs: both must be inf of the same sign, or both must be finite
+                if (isinf_or_max(f1[i]) || isinf_or_max(f2[i])) {
+                    if (isinf_or_max(f1[i]) && isinf_or_max(f2[i])) {
+                        if (std::signbit(f1[i]) != std::signbit(f2[i])) {
+                            printf("[%s] inf sign mismatch: %s=%f %s=%f ", ggml_op_desc(t1), bn1, f1[i], bn2, f2[i]);
+                            ud->ok = false;
+                            return true;
+                        }
+                    } else {
+                        printf("[%s] inf mismatch: %s=%f %s=%f ", ggml_op_desc(t1), bn1, f1[i], bn2, f2[i]);
+                        ud->ok = false;
+                        return true;
+                    }
+                }
+            }
+
+            double err = nmse(f1.data(), f2.data(), f1.size());
+            if (err > ud->max_err) {
+                printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
+                //for (int i = 0; i < (int) f1.size(); i++) {
+                //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
+                //}
+                //printf("\n");
+                //exit(1);
+                ud->ok = false;
+            }
+            return true;
+
+            GGML_UNUSED(index);
+        };
+
+        const bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
+
+        if (!cmp_ok) {
+            printf("compare failed ");
+        }
+
+        ggml_backend_buffer_free(buf);
+
+        ggml_free(ctx);
+
+        if (ud.ok && cmp_ok) {
+            printf("\033[1;32mOK\033[0m\n");
+            return true;
+        }
+
+        printf("\033[1;31mFAIL\033[0m\n");
+        return false;
+    }
+
+    bool eval_perf(ggml_backend_t backend, const char * op_name, int n_runs) {
+        mode = MODE_PERF;
+
+        static const size_t graph_nodes = 8192;
+
+        ggml_init_params params = {
+            /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead_custom(graph_nodes, false),
+            /* .mem_base = */ NULL,
+            /* .no_alloc = */ true,
+        };
+        ggml_context * ctx = ggml_init(params);
+
+        ggml_tensor * out = build_graph(ctx);
+
+        if (op_name != nullptr && op_desc(out) != op_name) {
+            //printf("  %s: skipping\n", op_desc(out).c_str());
+            ggml_free(ctx);
+            return true;
+        }
+
+        int len = printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
+        fflush(stdout);
+
+        // check if backends support op
+        if (!ggml_backend_supports_op(backend, out)) {
+            printf("not supported\n");
+            ggml_free(ctx);
+            return true;
+        }
+
+        // align while also leaving some margin for variations in parameters
+        int align = 20;
+        int last = (len + align - 1) / align * align;
+        if (last - len < 5) {
+            last += align;
+        }
+        last = std::max(last, 60);
+        printf("%*s", last - len, "");
+
+        // allocate
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        if (buf == NULL) {
+            printf("failed to allocate tensors\n");
+            ggml_free(ctx);
+            return false;
+        }
+
+        // randomize tensors
+        initialize_tensors(ctx);
+
+        // build graph
+        ggml_cgraph * gf = ggml_new_graph_custom(ctx, graph_nodes, false);
+        ggml_build_forward_expand(gf, out);
+
+        // warmup run
+        ggml_backend_graph_compute(backend, gf);
+
+        // duplicate the op
+        size_t target_size = ggml_backend_is_cpu(backend) ? 1ULL << 33 : 1ULL << 35; // 8 GB CPU, 32 GB GPU
+        //int n_runs = std::min((size_t)gf->size - gf->n_nodes, target_size / op_size(out)) + 1;
+        for (int i = 1; i < n_runs; i++) {
+            gf->nodes[gf->n_nodes++] = out;
+        }
+
+        // calculate memory
+        size_t mem = n_runs * op_size(out);
+        auto tensor_op_size = [](ggml_tensor * t) {
+            size_t size = ggml_nbytes(t);
+            // add source tensors
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (t->src[i] != NULL) {
+                    size += ggml_nbytes(t->src[i]);
+                }
+            }
+            return size;
+        };
+        for (int i = 0; i < gf->n_nodes; i++) {
+            if (ggml_is_view_op(gf->nodes[i]->op) || gf->nodes[i] == out) {
+                continue;
+            }
+            mem += tensor_op_size(gf->nodes[i]);
+        }
+
+        // run
+        ggml_backend_synchronize(backend);
+
+        int64_t start_time = ggml_time_us();
+        ggml_backend_graph_compute(backend, gf);
+        ggml_backend_synchronize(backend);
+        int64_t end_time = ggml_time_us();
+        double time_us = end_time - start_time;
+
+        printf("    %5d runs - %8.2f us/run - %8zu kB/run - \033[1;34m%7.2f GB/s\033[0m\n",
+            n_runs,
+            time_us / n_runs,
+            op_size(out) / 1024,
+            mem / (time_us/1e6) / 1024.0 / 1024.0 / 1024.0);
+
+        ggml_backend_buffer_free(buf);
+
+        ggml_free(ctx);
+
+        return true;
+    }
+};
+
+// GGML_OP_MUL_MAT
+struct test_mul_mat : public test_case {
+    const ggml_type type_a;
+    const ggml_type type_b;
+    const int64_t m;
+    const int64_t n;
+    const int64_t k;
+    const std::array<int64_t, 2> bs; // dims 3 and 4
+    const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
+
+    std::string vars() override {
+        return VARS_TO_STR7(type_a, type_b, m, n, k, bs, nr); 
+    }    
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }    
+
+    size_t op_size(ggml_tensor * t) override {
+        size_t a = ggml_nbytes(t->src[0]) * n * nr[0] * nr[1];
+        size_t b = ggml_nbytes(t->src[1]) * m; 
+        size_t c  = ggml_nbytes(t);
+        return a + b + c; 
+
+        GGML_UNUSED(t);
+    }    
+
+    test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
+            int64_t m = 32, int64_t n = 32, int64_t k = 32,
+            std::array<int64_t, 2> bs = {10, 10}, 
+            std::array<int64_t, 2> nr = {2, 2})
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        // C^T = A * B^T: (k, m) * (k, n) => (m, n)
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0]      , bs[1]);
+        ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
+        ggml_tensor * out = ggml_mul_mat(ctx, a, b);
+        return out; 
+    }    
+};
+
+void bench_mul_mat(ggml_backend_t backend) {
+    auto test1 = test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F16, 11008, 1, 4096, { 1,  1}, {1, 1});
+    auto test2 = test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0, 11008, 1, 4096, { 1,  1}, {1, 1});
+
+    auto test3 = test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F16, 512, 256, 1024, { 1,  1}, {1, 1});
+    auto test4 = test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0, 512, 256, 1024, { 1,  1}, {1, 1});
+
+    test1.eval_perf(backend, "MUL_MAT", 8000 /*n_runs*/);
+    test2.eval_perf(backend, "MUL_MAT", 8000 /*n_runs*/);
+    test3.eval_perf(backend, "MUL_MAT", 8000 /*n_runs*/);
+    test4.eval_perf(backend, "MUL_MAT", 8000 /*n_runs*/);
+}
+
+int main() {
+    // enumerate backends
+    std::cout << "num_backends:" << ggml_backend_reg_get_count() << std::endl;
+    int backend_id = 1;
+
+    {
+        ggml_backend_t backend = ggml_backend_reg_init_backend(backend_id, NULL);
+        std::cout << "Using backend:" << ggml_backend_name(backend) << std::endl;
+
+        bench_mul_mat(backend);
+
+        ggml_backend_free(backend);
+    }
+}

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

@@ -73,7 +73,6 @@ struct my_llama_model {
 static const char * LLM_KV_TRAINING_TYPE_TRAIN_MODEL     = "train_model";
 static const char * LLM_KV_TRAINING_TYPE                 = "training.type";
 
-static const char * LLM_KV_GENERAL_NAME                = "general.name";
 static const char * LLM_KV_GENERAL_ARCHITECTURE        = "general.architecture";
 static const char * LLM_KV_GENERAL_FILE_TYPE           = "general.file_type";
 
@@ -530,7 +529,6 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
 
 static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model) {
     const char * arch = "llama";
-
     enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
 
     std::vector<char> keybuf;
@@ -542,7 +540,6 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
 
     // set arch
     gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch);
-    gguf_set_val_str(fctx, LLM_KV_GENERAL_NAME, arch);
     gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype);
 
     // set hparams

examples/ts-type-to-grammar.sh

@@ -1,7 +1,7 @@
 #!/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
+# python examples/json-schema-to-grammar.py https://json.schemastore.org/tsconfig.json
 #
 set -euo pipefail
 
@@ -25,4 +25,4 @@ npx ts-json-schema-generator --unstable --no-top-ref --path "$DTS_FILE" --type M
 # https://github.com/YousefED/typescript-json-schema
 # npx typescript-json-schema --defaultProps --required "$DTS_FILE" MyType | tee "$SCHEMA_FILE" >&2
 
-./examples/json_schema_to_grammar.py "$SCHEMA_FILE"
+./examples/json-schema-to-grammar.py "$SCHEMA_FILE"

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1713537308,
-        "narHash": "sha256-XtTSSIB2DA6tOv+l0FhvfDMiyCmhoRbNB+0SeInZkbk=",
+        "lastModified": 1712163089,
+        "narHash": "sha256-Um+8kTIrC19vD4/lUCN9/cU9kcOsD1O1m+axJqQPyMM=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5c24cf2f0a12ad855f444c30b2421d044120c66f",
+        "rev": "fd281bd6b7d3e32ddfa399853946f782553163b5",
         "type": "github"
       },
       "original": {

ggml-alloc.c

@@ -371,16 +371,16 @@ struct ggml_gallocr {
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
     GGML_ASSERT(galloc != NULL);
 
-    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
+    galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
     GGML_ASSERT(galloc->bufts != NULL);
 
-    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
+    galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
     GGML_ASSERT(galloc->buffers != NULL);
 
-    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+    galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
     GGML_ASSERT(galloc->buf_tallocs != NULL);
 
     for (int i = 0; i < n_bufs; i++) {
@@ -646,8 +646,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         free(galloc->hash_set.keys);
         free(galloc->hash_values);
         galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
-        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
+        galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
+        galloc->hash_values   = calloc(sizeof(struct hash_node), hash_size);
         GGML_ASSERT(galloc->hash_set.keys != NULL);
         GGML_ASSERT(galloc->hash_values != NULL);
     } else {
@@ -667,7 +667,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     // set the node_allocs from the hash table
     if (galloc->n_nodes < graph->n_nodes) {
         free(galloc->node_allocs);
-        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
         GGML_ASSERT(galloc->node_allocs != NULL);
     }
     galloc->n_nodes = graph->n_nodes;
@@ -697,7 +697,7 @@ 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(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
         GGML_ASSERT(galloc->leaf_allocs != NULL);
     }
     galloc->n_leafs = graph->n_leafs;

ggml-backend.c

@@ -822,11 +822,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return
-                op->type != GGML_TYPE_IQ2_XXS &&
-                op->type != GGML_TYPE_IQ2_XS  &&
-                op->type != GGML_TYPE_IQ1_S   &&
-                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1725,23 +1721,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
 
-    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
+    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
 
     // initialize hash table
     sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
-    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
+    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
+    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
 
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
-    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
+    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
 
     sched->n_backends = n_backends;
 
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     const int initial_splits_capacity = 16;
-    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
+    sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
     sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
@@ -1972,10 +1968,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
     struct ggml_hash_set hash_set = {
         /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
+        /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
     };
-    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
-    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
+    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
+    bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);
 
     struct ggml_init_params params = {
         /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),

ggml-cuda.cu

@@ -1231,7 +1231,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
-        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
+        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool());
         if (src0->type != GGML_TYPE_F16) {
             const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
             GGML_ASSERT(to_fp16_cuda != nullptr);
@@ -1241,7 +1241,7 @@ static void ggml_cuda_op_mul_mat_cublas(
         }
         const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
 
-        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool(id));
+        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool());
         if (src1->type != GGML_TYPE_F16) {
             const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
             GGML_ASSERT(to_fp16_cuda != nullptr);
@@ -1250,7 +1250,7 @@ static void ggml_cuda_op_mul_mat_cublas(
             to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
         }
         const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
-        ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(id), row_diff*src1_ncols);
+        ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(), row_diff*src1_ncols);
 
         const half alpha_f16 = 1.0f;
         const half beta_f16 = 0.0f;
@@ -1946,7 +1946,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     } else if (!split && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || fp16_performance_good) && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
@@ -1960,73 +1960,20 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
-    GGML_TENSOR_BINARY_OP_LOCALS
-
     GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    const size_t nb11 = src1->nb[1];
+    const size_t nb1  =  dst->nb[1];
+
+    const int32_t id = ((int32_t *) dst->op_params)[0];
+    const int32_t n_as = src0->ne[2];
 
     std::vector<char> ids_host(ggml_nbytes(ids));
     const char * ids_dev = (const char *) ids->data;
@@ -2035,7 +1982,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     ggml_tensor src0_row = *src0;
     ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
+    ggml_tensor dst_row = *dst;
 
     char * src0_original = (char *) src0->data;
     char * src1_original = (char *) src1->data;
@@ -2043,39 +1990,19 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     src0_row.ne[2] = 1;
     src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
+    src0_row.nb[3] = src0->nb[2];
 
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
+    if (src1->ne[1] == 1) {
+        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+            const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
 
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
+            GGML_ASSERT(row_id >= 0 && row_id < n_as);
 
-    if (ne12 == 1) {
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-            for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            src0_row.data = src0_original + row_id*src0->nb[2];
+            src1_row.data = src1_original + i01*src1->nb[1];
+            dst_row.data  =  dst_original + i01*dst->nb[1];
 
-                GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-                const int64_t i11 = id % ne11;
-                const int64_t i12 = iid1;
-
-                const int64_t i1 = id;
-                const int64_t i2 = i12;
-
-                src0_row.data = src0_original + i02*nb02;
-                src1_row.data = src1_original + i11*nb11 + i12*nb12;
-                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
-
-                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-            }
+            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
         }
     } else {
         ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
@@ -2084,69 +2011,54 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         src1_row.data = src1_contiguous.get();
         dst_row.data  =  dst_contiguous.get();
 
-        for (int64_t i02 = 0; i02 < n_as; i02++) {
+        for (int32_t row_id = 0; row_id < n_as; ++row_id) {
             int64_t num_src1_rows = 0;
+            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
 
-            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-                for (int64_t id = 0; id < n_ids; id++) {
-                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                    if (row_id_i != i02) {
-                        continue;
-                    }
-
-                    num_src1_rows++;
+                if (row_id_i != row_id) {
+                    continue;
                 }
+
+                GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+                CUDA_CHECK(cudaMemcpyAsync(src1_contiguous.get() + num_src1_rows*nb11, src1_original + i01*nb11,
+                                        nb11, cudaMemcpyDeviceToDevice, stream));
+                num_src1_rows++;
             }
 
             if (num_src1_rows == 0) {
                 continue;
             }
 
-            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne10, 768u));
-                dim3 grid_dims(ids->ne[1], n_ids);
-                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        src1_original, src1_contiguous.get(),
-                        dev_cur_src1_row.get(), dev_row_mapping.get(),
-                        ids_dev, i02, ids->nb[1], ids->nb[0],
-                        ne11, ne10,
-                        nb11, nb12);
-                CUDA_CHECK(cudaGetLastError());
-            }
-
-            src0_row.data = src0_original + i02*nb02;
-
-            GGML_ASSERT(nb11 == sizeof(float)*ne10);
-            GGML_ASSERT(nb1 == sizeof(float)*ne0);
+            src0_row.data = src0_original + row_id*src0->nb[2];
 
             src1_row.ne[1] = num_src1_rows;
+            dst_row.ne[1] = num_src1_rows;
+
             src1_row.nb[1] = nb11;
             src1_row.nb[2] = num_src1_rows*nb11;
             src1_row.nb[3] = num_src1_rows*nb11;
 
-            dst_row.ne[1] = num_src1_rows;
             dst_row.nb[1] = nb1;
             dst_row.nb[2] = num_src1_rows*nb1;
             dst_row.nb[3] = num_src1_rows*nb1;
 
             ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
-            {
-                dim3 block_dims(std::min((unsigned int)ne0, 768u));
-                dim3 grid_dims(num_src1_rows);
-                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        dst_original, dst_contiguous.get(),
-                        dev_row_mapping.get(),
-                        ne0,
-                        nb1, nb2);
-                CUDA_CHECK(cudaGetLastError());
+            num_src1_rows = 0;
+            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
+
+                if (row_id_i != row_id) {
+                    continue;
+                }
+
+                GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+                CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous.get() + num_src1_rows*nb1,
+                                        nb1, cudaMemcpyDeviceToDevice, stream));
+                num_src1_rows++;
             }
         }
     }
@@ -2575,8 +2487,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
 GGML_CALL static bool ggml_backend_cuda_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
 
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
 
     GGML_UNUSED(backend);
 }

ggml-cuda/binbcast.cu

@@ -22,7 +22,6 @@ static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
         /*int s10,*/ int s11, int s12, int s13) {
     const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
     const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
@@ -37,9 +36,9 @@ static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst
     const int i12 = i2 % ne12;
     const int i13 = i3 % ne13;
 
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
     const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_dst  = i_src0;
 
     const src0_t * src0_row = src0 + i_src0;
     const src1_t * src1_row = src1 + i_src1;
@@ -56,7 +55,6 @@ static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * s
         int ne0, int ne1, int ne2, int ne3,
         int ne10, int ne11, int ne12, int ne13,
         /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
         /*int s10,*/ int s11, int s12, int s13) {
 
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -74,9 +72,9 @@ static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * s
     const int i12 = i2 % ne12;
     const int i13 = i3 % ne13;
 
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
     const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_dst  = i_src0;
 
     const src0_t * src0_row = src0 + i_src0;
     const src1_t * src1_row = src1 + i_src1;
@@ -103,14 +101,10 @@ struct bin_bcast_cuda {
         int nr[4] = { nr0, nr1, nr2, nr3 };
 
         // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
+        int64_t cne0[] = {ne0, ne1, ne2, ne3};
         int64_t cne1[] = {ne10, ne11, ne12, ne13};
-
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
+        size_t cnb0[] = {nb0, nb1, nb2, nb3};
         size_t cnb1[] = {nb10, nb11, nb12, nb13};
-
         auto collapse = [](int64_t cne[]) {
             cne[0] *= cne[1];
             cne[1] = cne[2];
@@ -124,47 +118,32 @@ struct bin_bcast_cuda {
             cnb[3] *= cne[3];
         };
 
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
-                }
+        for (int i = 0; i < 4; i++) {
+            if (nr[i] != 1) {
+                break;
+            }
+            if (i > 0) {
+                collapse_nb(cnb0, cne0);
+                collapse_nb(cnb1, cne1);
+                collapse(cne0);
+                collapse(cne1);
             }
         }
-
         {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            //int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
-            //int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
-            //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
-            //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
+            int64_t ne0 = cne0[0];
+            int64_t ne1 = cne0[1];
+            int64_t ne2 = cne0[2];
+            int64_t ne3 = cne0[3];
 
             int64_t ne10 = cne1[0];
             int64_t ne11 = cne1[1];
             int64_t ne12 = cne1[2];
             int64_t ne13 = cne1[3];
 
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
+            size_t nb0 = cnb0[0];
+            size_t nb1 = cnb0[1];
+            size_t nb2 = cnb0[2];
+            size_t nb3 = cnb0[3];
 
             size_t nb10 = cnb1[0];
             size_t nb11 = cnb1[1];
@@ -181,28 +160,7 @@ struct bin_bcast_cuda {
             size_t s12 = nb12 / sizeof(src1_t);
             size_t s13 = nb13 / sizeof(src1_t);
 
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
             GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s00 == 1);
             GGML_ASSERT(s10 == 1);
 
             const int block_size = 128;
@@ -221,14 +179,13 @@ struct bin_bcast_cuda {
             );
 
             if (block_nums.z > 65535) {
-                // this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
+                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
                 int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
                 k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
                     src0_dd, src1_dd, dst_dd,
                     ne0, ne1, ne2, ne3,
                     ne10, ne11, ne12, ne13,
                     /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
                     /* s10, */ s11, s12, s13);
             } else {
                 k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
@@ -236,7 +193,6 @@ struct bin_bcast_cuda {
                     ne0, ne1, ne2, ne3,
                     ne10, ne11, ne12, ne13,
                     /* s0, */ s1, s2, s3,
-                    /* s00, */ s01, s02, s03,
                     /* s10, */ s11, s12, s13);
             }
         }

ggml-cuda/convert.cu

@@ -45,8 +45,6 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
         vals[ix] = x0[ix];
     }
 
-    __syncthreads();
-
 #pragma unroll
     for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
         if (need_check && i0 + iy + 2*threadIdx.x >= k) {

ggml-impl.h

@@ -45,7 +45,7 @@ extern "C" {
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
-#if defined(__ARM_NEON)
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
 
 // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
 //
@@ -53,262 +53,8 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
 typedef __fp16 ggml_fp16_internal_t;
 
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddvq_s16(int16x8_t v) {
-    return
-        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
-        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
-        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
-        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(__MSC_VER)
-
 #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
@@ -329,6 +75,8 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #else
 
+typedef uint16_t ggml_fp16_internal_t;
+
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -340,7 +88,7 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <intrin.h>
 #else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
 #if !defined(__riscv)
 #include <immintrin.h>
 #endif
@@ -473,7 +221,7 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #endif // __F16C__
 
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+#endif // __ARM_NEON
 
 // precomputed f32 table for f16 (256 KB)
 // defined in ggml.c, initialized in ggml_init()

ggml-metal.m

@@ -37,15 +37,11 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_DIV_ROW,
     GGML_METAL_KERNEL_TYPE_SCALE,
     GGML_METAL_KERNEL_TYPE_SCALE_4,
-    GGML_METAL_KERNEL_TYPE_CLAMP,
     GGML_METAL_KERNEL_TYPE_TANH,
     GGML_METAL_KERNEL_TYPE_RELU,
     GGML_METAL_KERNEL_TYPE_GELU,
-    GGML_METAL_KERNEL_TYPE_GELU_4,
     GGML_METAL_KERNEL_TYPE_GELU_QUICK,
-    GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
     GGML_METAL_KERNEL_TYPE_SILU,
-    GGML_METAL_KERNEL_TYPE_SILU_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,
     GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
@@ -472,15 +468,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                     clamp,                  true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                    gelu_4,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,              gelu_quick_4,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                    silu_4,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,               ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,             ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,          true);
@@ -721,7 +713,6 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SCALE:
-        case GGML_OP_CLAMP:
         case GGML_OP_SQR:
         case GGML_OP_SUM_ROWS:
             return true;
@@ -1163,30 +1154,8 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
-                case GGML_OP_CLAMP:
-                {
-                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
-
-                    float min;
-                    float max;
-                    memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
-                    memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
-
-                    [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                    [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                    [encoder setBytes:&max length:sizeof(max) atIndex:3];
-
-                    const int64_t n = ggml_nelements(dst);
-
-                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                } break;
                 case GGML_OP_UNARY:
                     switch (ggml_get_unary_op(gf->nodes[i])) {
-                        // we are not taking into account the strides, so for now require contiguous tensors
-                        GGML_ASSERT(ggml_is_contiguous(src0));
-
                         case GGML_UNARY_OP_TANH:
                             {
                                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline;
@@ -1213,60 +1182,42 @@ static enum ggml_status ggml_metal_graph_compute(
                             } break;
                         case GGML_UNARY_OP_GELU:
                             {
-                                int64_t n = ggml_nelements(dst);
-
-                                id<MTLComputePipelineState> pipeline = nil;
-
-                                if (n % 4 == 0) {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline;
-                                    n /= 4;
-                                } else {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
-                                }
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
 
                                 [encoder setComputePipelineState:pipeline];
                                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                const int64_t n = ggml_nelements(dst);
+                                GGML_ASSERT(n % 4 == 0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } break;
                         case GGML_UNARY_OP_GELU_QUICK:
                             {
-                                int64_t n = ggml_nelements(dst);
-
-                                id<MTLComputePipelineState> pipeline = nil;
-
-                                if (n % 4 == 0) {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline;
-                                    n /= 4;
-                                } else {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
-                                }
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
 
                                 [encoder setComputePipelineState:pipeline];
                                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                const int64_t n = ggml_nelements(dst);
+                                GGML_ASSERT(n % 4 == 0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } break;
                         case GGML_UNARY_OP_SILU:
                             {
-                                int64_t n = ggml_nelements(dst);
-
-                                id<MTLComputePipelineState> pipeline = nil;
-
-                                if (n % 4 == 0) {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline;
-                                    n /= 4;
-                                } else {
-                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
-                                }
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
 
                                 [encoder setComputePipelineState:pipeline];
                                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
 
-                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                const int64_t n = ggml_nelements(dst);
+                                GGML_ASSERT(n % 4 == 0);
+
+                                [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } break;
                         default:
                             {
@@ -1732,10 +1683,15 @@ static enum ggml_status ggml_metal_graph_compute(
                     } break;
                 case GGML_OP_MUL_MAT_ID:
                     {
+                        //GGML_ASSERT(ne00 == ne10);
+                        //GGML_ASSERT(ne03 == ne13);
                         const int n_as = src0->ne[2];
 
+                        // max size of the src1ids array in the kernel shared buffer
+                        GGML_ASSERT(ne11 <= 4096);
+
                         // src2 = ids
-                        const int64_t  ne20 = src2->ne[0];
+                        const int64_t  ne20 = src2->ne[0]; GGML_UNUSED(ne20);
                         const int64_t  ne21 = src2->ne[1];
                         const int64_t  ne22 = src2->ne[2]; GGML_UNUSED(ne22);
                         const int64_t  ne23 = src2->ne[3]; GGML_UNUSED(ne23);
@@ -1756,13 +1712,15 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                         // to the matrix-vector kernel
-                        // ne20 = n_used_experts
-                        // ne21 = n_rows
-                        const int dst_rows = ne20*ne21;
-                        const int dst_rows_min = n_as;
+                        int ne11_mm_min = n_as;
 
-                        // max size of the rowids array in the kernel shared buffer
-                        GGML_ASSERT(dst_rows <= 2048);
+                        const int idx = ((int32_t *) dst->op_params)[0];
+
+                        // batch size
+                        GGML_ASSERT(ne21 == ne11); // ?
+                        GGML_ASSERT(ne12 == 1 && ne13 == 1); // no broadcasting
+                        const uint r2 = 1;
+                        const uint r3 = 1;
 
                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
@@ -1772,7 +1730,7 @@ static enum ggml_status ggml_metal_graph_compute(
                         // !!!
                         if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
                             ne00 % 32 == 0 && ne00 >= 64 &&
-                            dst_rows > dst_rows_min) {
+                            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)
@@ -1814,26 +1772,26 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
                             [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
-                            [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
-                            [encoder setBytes:&ne21    length:sizeof(ne21) atIndex:5];
-                            [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
-                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:7];
-                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:8];
-                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:9];
-                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:10];
-                            [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:11];
-                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:13];
-                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:14];
-                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:15];
-                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:16];
-                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:17];
-                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:18];
-                            [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:19];
+                            [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:4];
+                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:5];
+                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
+                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:9];
+                            [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:10];
+                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
+                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
+                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:13];
+                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:14];
+                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:15];
+                            [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:16];
+                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:17];
+                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:18];
+                            [encoder setBytes:&idx     length:sizeof(idx)  atIndex:19];
 
-                            [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
+                            [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + 2*ne11, 16) atIndex:0];
 
-                            [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne01 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
                             int nth0 = 32;
                             int nth1 = 1;
@@ -1968,12 +1926,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
-                                    #if QK_K == 64
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    #else
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
-                                    #endif
-
                                     } break;
                                 default:
                                     {
@@ -1986,72 +1939,72 @@ static enum ggml_status ggml_metal_graph_compute(
                                 GGML_ASSERT(ne00 >= nth0*nth1);
                             }
 
+                            const int64_t _ne1 = 1; // kernels needs a reference in constant memory
+
                             [encoder setComputePipelineState:pipeline];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                             [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                             [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
                             [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
-                            [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
-                            [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
-                            [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:20];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:21];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:22];
-
-                            const int64_t _ne1 = 1;
-                            const int tgz = dst_rows;
+                            [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:4];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:5];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:6];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:7];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:8];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:12];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:18];
+                            [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:19];
+                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:20];
+                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:21];
+                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:22];
+                            [encoder setBytes:&idx  length:sizeof(idx)  atIndex:23];
 
                             if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
                                 src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
                                 src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
                                 const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
                                 const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
                                 const int mem_size = 32*sizeof(float);
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q4_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q3_K) {
 #ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #else
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
 #endif
                             }
                             else if (src0t == GGML_TYPE_Q5_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q6_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             } else {
-                                const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                const int64_t ny = (_ne1 + nrows - 1)/nrows;
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                         }
                     } break;

ggml-quants.c

@@ -14,6 +14,41 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT
 
+#ifdef __ARM_NEON
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#if defined(__POWER9_VECTOR__) || defined(__powerpc64__)
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
 #undef MIN
 #undef MAX
 
@@ -97,7 +132,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) {
 }
 
 static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVXVNNI__) || defined(__AVX512VNNI__)
     const __m256i zero = _mm256_setzero_si256();
     const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
     return _mm256_cvtepi32_ps(summed_pairs);
@@ -241,6 +276,258 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 #endif // __AVX__ || __AVX2__ || __AVX512F__
 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 
+#if defined(__ARM_NEON)
+
+#ifdef _MSC_VER
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif
+
+#if !defined(__aarch64__)
+
+// 64-bit compatibility
+
+// vaddvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif
+
+#endif
+
 #if defined(__ARM_NEON) || defined(__wasm_simd128__)
 #define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
 #define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)

ggml-sycl.cpp

@@ -3154,6 +3154,7 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
 #define SYCL_SCALE_BLOCK_SIZE 256
 #define SYCL_CLAMP_BLOCK_SIZE 256
 #define SYCL_ROPE_BLOCK_SIZE 256
+#define SYCL_SOFT_MAX_BLOCK_SIZE 1024
 #define SYCL_ALIBI_BLOCK_SIZE 32
 #define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
 #define SYCL_QUANTIZE_BLOCK_SIZE 256
@@ -13079,13 +13080,11 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
                               const int nrows_y, const float scale, const float max_bias,
                               dpct::queue_ptr stream) {
     int nth = WARP_SIZE;
-    int max_block_size = g_work_group_size;
-    while (nth < ncols_x && nth < max_block_size) nth *= 2;
-    if (nth>max_block_size) nth = max_block_size;
-
+    while (nth < ncols_x && nth < SYCL_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, 1, nrows_x);
     const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE);
+    static_assert(SYCL_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
     const uint32_t n_head_kv   = nrows_x/nrows_y;
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
@@ -13095,12 +13094,6 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
 
     const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
     if (n_local_scratch*sizeof(float) < local_mem_size) {
-        if (ncols_x > max_block_size) {
-            soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
-                                               max_bias, m0, m1, n_head_log2, block_nums,
-                                               block_dims, n_local_scratch, stream);
-            return;
-        }
         switch (ncols_x) {
             case 32:
                 soft_max_f32_submitter<true, 32, 32>(x, mask, pos, dst, ncols_x, nrows_y, scale,
@@ -15996,76 +15989,73 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
 static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
                                  const ggml_tensor *src1,
                                  ggml_tensor *dst) try {
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT &&
-                "mul_mat_id does not support split buffers");
-    const ggml_tensor *ids = dst->src[2];
-    const dpct::queue_ptr stream = g_syclStreams[g_main_device][0];
+#if 0
+    ggml_sycl_mul_mat_id_sycl(dst);
+    // TODO: mmq/mmv support
+#endif
 
-    const size_t nb11 = src1->nb[1];
-    const size_t nb1 = dst->nb[1];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb1  =  dst->nb[1];
 
-    const int32_t id = ((int32_t *)dst->op_params)[0];
-    const int32_t n_as = src0->ne[2];
+    const struct ggml_tensor * ids = src0;
+    const int32_t id = ((int32_t *) dst->op_params)[0];
+    const int32_t n_as = ((int32_t *) dst->op_params)[1];
 
     std::vector<char> ids_host(ggml_nbytes(ids));
-    const char *ids_dev = (const char *)ids->data;
 
-    SYCL_CHECK(CHECK_TRY_ERROR(
-        stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
-    SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
+    const dpct::queue_ptr stream = g_syclStreams[g_main_device][0];
 
-    const ggml_tensor_extra_gpu *src0_extra =
-        (const ggml_tensor_extra_gpu *)src0->extra;
-    const ggml_tensor_extra_gpu *src1_extra =
-        (const ggml_tensor_extra_gpu *)src1->extra;
-    const ggml_tensor_extra_gpu *dst_extra =
-        (const ggml_tensor_extra_gpu *)dst->extra;
+    if (ids->backend == GGML_BACKEND_TYPE_GPU) {
+        const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)).wait()));
+        // SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
+    } else {
+        memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
+    }
+
+    const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra;
+    const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra;
 
-    ggml_tensor_extra_gpu src0_row_extra;
     ggml_tensor_extra_gpu src1_row_extra;
     ggml_tensor_extra_gpu dst_row_extra;
 
-    ggml_tensor src0_row = *src0;
     ggml_tensor src1_row = *src1;
     ggml_tensor dst_row = *dst;
 
     src1_row.backend = GGML_BACKEND_TYPE_GPU;
     dst_row.backend  = GGML_BACKEND_TYPE_GPU;
 
-    src0_row.extra = &src0_row_extra;
     src1_row.extra = &src1_row_extra;
     dst_row.extra = &dst_row_extra;
 
-    char *src0_original = src1->backend == GGML_BACKEND_TYPE_CPU
-                              ? (char *)src0->data
-                              : (char *)src0_extra->data_device[g_main_device];
-    char *src1_original = src1->backend == GGML_BACKEND_TYPE_CPU
-                              ? (char *)src1->data
-                              : (char *)src1_extra->data_device[g_main_device];
-    char *dst_original = dst->backend == GGML_BACKEND_TYPE_CPU
-                             ? (char *)dst->data
-                             : (char *)dst_extra->data_device[g_main_device];
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = src0->nb[2];
+    char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ?
+        (char *) src1->data : (char *) src1_extra->data_device[g_main_device];
+    char * dst_original  =  dst->backend == GGML_BACKEND_TYPE_CPU ?
+        (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device];
 
     if (src1->ne[1] == 1) {
+        GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+        GGML_ASSERT(dst->backend  == GGML_BACKEND_TYPE_GPU);
+
         for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-            const int32_t row_id =
-                *(const int32_t *)(ids_host.data() + i01 * ids->nb[1] +
-                                   id * ids->nb[0]);
+            //int32_t row_id;
+            //SYCL_CHECK(syclMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), syclMemcpyDeviceToHost, g_syclStreams[g_main_device][0]));
+            //SYCL_CHECK(syclStreamSynchronize(g_syclStreams[g_main_device][0]));
+
+            const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
 
             GGML_ASSERT(row_id >= 0 && row_id < n_as);
 
-            src0_row_extra.data_device[g_main_device] =
-                src0_original + row_id * src0->nb[2];
-            src1_row_extra.data_device[g_main_device] =
-                src1_original + i01 * src1->nb[1];
-            dst_row_extra.data_device[g_main_device] =
-                dst_original + i01 * dst->nb[1];
+            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
 
-            ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
+            src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1];
+            src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set?
+
+            dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1];
+            dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set?
+
+            ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row);
         }
     } else {
         sycl_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1));
@@ -16075,6 +16065,8 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
         dst_row_extra.data_device[g_main_device]  =  dst_contiguous.get();
 
         for (int32_t row_id = 0; row_id < n_as; ++row_id) {
+            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
+
             int64_t num_src1_rows = 0;
             for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
                 const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
@@ -16087,7 +16079,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
 
                 SYCL_CHECK(CHECK_TRY_ERROR(
                     stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11,
-                                   src1_original + i01 * nb11, nb11)));
+                                   src1_original + i01 * nb11, nb11).wait()));
                 num_src1_rows++;
             }
 
@@ -16095,9 +16087,6 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
                 continue;
             }
 
-            src0_row_extra.data_device[g_main_device] =
-                src0_original + row_id * src0->nb[2];
-
             src1_row.ne[1] = num_src1_rows;
             dst_row.ne[1] = num_src1_rows;
 
@@ -16109,7 +16098,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
             dst_row.nb[2] = num_src1_rows*nb1;
             dst_row.nb[3] = num_src1_rows*nb1;
 
-            ggml_sycl_mul_mat(&src0_row, &src1_row, &dst_row);
+            ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row);
 
             num_src1_rows = 0;
             for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
@@ -16123,7 +16112,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
 
                 SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
                     dst_original + i01 * nb1,
-                    dst_contiguous.get() + num_src1_rows * nb1, nb1)));
+                    dst_contiguous.get() + num_src1_rows * nb1, nb1).wait()));
                 num_src1_rows++;
             }
         }
@@ -16825,13 +16814,11 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
     const dpct::queue_ptr stream = g_syclStreams[ctx->device][0];
     SYCL_CHECK(
         CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
-    char* host_buf = (char*)malloc(size);
-    memcpy(host_buf, data, size);
+
     SYCL_CHECK(
         CHECK_TRY_ERROR((*stream)
-                             .memcpy((char *)tensor->data + offset, host_buf, size)
+                             .memcpy((char *)tensor->data + offset, data, size)
                              .wait()));
-    free(host_buf);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -17752,7 +17739,7 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
 
 GGML_CALL static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
     const int min_batch_size = 32;
-    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS && op->op != GGML_OP_MUL_MAT_ID;
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
     GGML_UNUSED(backend);
 }
 

ggml.c

@@ -4,7 +4,6 @@
 #include "ggml-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
-#include "sgemm.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -33,10 +32,6 @@
 #include <unistd.h>
 #endif
 
-#ifdef __ARM_FEATURE_MATMUL_INT8
-#undef GGML_USE_LLAMAFILE
-#endif
-
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -4578,32 +4573,21 @@ void ggml_mul_mat_set_prec(
 
 // ggml_mul_mat_id
 
-/*
-    c = ggml_mul_mat_id(ctx, as, b, ids);
-
-    as  -> [cols, rows, n_expert]
-    ids -> [n_experts_used, n_tokens] (i32)
-    b   -> [cols, n_expert_used, n_tokens]
-    c   -> [cols, n_expert_used, n_tokens]
-
-    in b, n_experts_used can be broadcasted to match the n_expert_used of ids
-
-    c ~= as[:,:,i] @ b[:,i%r,t], i = ids[e,t] for all e,t in ids
-*/
+// NOTE: id will be removed in the future and instead all the experts listed in ids will be computed
+//       this will allow computing all the used experts in a single matrix multiplication
 struct ggml_tensor * ggml_mul_mat_id(
         struct ggml_context * ctx,
         struct ggml_tensor  * as,
-        struct ggml_tensor  * b,
-        struct ggml_tensor  * ids) {
-    GGML_ASSERT(!ggml_is_transposed(as));
-    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+        struct ggml_tensor  * ids,
+        int                   id,
+        struct ggml_tensor  * b) {
 
-    GGML_ASSERT(as->ne[3] == 1); // as is 3d (one matrix per expert)
-    GGML_ASSERT(b->ne[3] == 1); // b is 3d
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
     GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1); // ids is 2d
-    GGML_ASSERT(ids->ne[1] == b->ne[2]); // must have an expert list per b row
+    GGML_ASSERT(ids->ne[1] == b->ne[1]); // must have an expert per b row
+    GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
+    GGML_ASSERT(id >= 0 && id < ids->ne[0]); // valid id
     GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat
-    GGML_ASSERT(ids->ne[0] % b->ne[1] == 0); // can broadcast
 
     bool is_node = false;
 
@@ -4611,9 +4595,11 @@ struct ggml_tensor * ggml_mul_mat_id(
         is_node = true;
     }
 
-    const int64_t ne[4] = { as->ne[1], ids->ne[0], b->ne[2], 1 };
+    const int64_t ne[4] = { as->ne[1], b->ne[1], b->ne[2], b->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
+    ggml_set_op_params_i32(result, 0, id);
+
     result->op   = GGML_OP_MUL_MAT_ID;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = as;
@@ -10824,28 +10810,6 @@ static void ggml_compute_forward_mul_mat(
     }
 #endif
 
-#if GGML_USE_LLAMAFILE
-    if (src1_cont) {
-        for (int64_t i13 = 0; i13 < ne13; i13++)
-            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
-                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(src0->type),
-                                     (const char *)src1->data + i12*nb12 + i13*nb13,
-                                     nb11/ggml_type_size(src1->type),
-                                     (char *)dst->data + i12*nb2 + i13*nb3,
-                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
-                                     params->type,
-                                     src0->type,
-                                     src1->type,
-                                     dst->type))
-                    goto UseGgmlGemm1;
-        return;
-    }
-UseGgmlGemm1:;
-#endif
-
     if (params->type == GGML_TASK_TYPE_INIT) {
         if (ith != 0) {
             return;
@@ -10877,28 +10841,6 @@ UseGgmlGemm1:;
     const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
     const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-#if GGML_USE_LLAMAFILE
-    if (src1->type != vec_dot_type) {
-        for (int64_t i13 = 0; i13 < ne13; i13++)
-            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
-                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(src0->type),
-                                     (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
-                                     row_size/ggml_type_size(vec_dot_type),
-                                     (char *)dst->data + i12*nb2 + i13*nb3,
-                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
-                                     params->type,
-                                     src0->type,
-                                     vec_dot_type,
-                                     dst->type))
-                    goto UseGgmlGemm2;
-        return;
-    }
-UseGgmlGemm2:;
-#endif
-
     const int64_t nr0 = ne01;          // src0 rows
     const int64_t nr1 = ne1*ne12*ne13; // src1 rows
 
@@ -11016,6 +10958,11 @@ static void ggml_compute_forward_mul_mat_id(
     enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
     ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
 
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
     GGML_ASSERT(nb10 == ggml_type_size(src1->type));
@@ -11026,21 +10973,22 @@ static void ggml_compute_forward_mul_mat_id(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
+    // broadcast is not supported with mmid
+    assert(ne12 == 1);
+    assert(ne13 == 1);
+
     // row groups
-    const int n_ids = ids->ne[0]; // n_expert_used
-    const int n_as  = ne02;       // n_expert
+    const int id   = ggml_get_op_params_i32(dst, 0);
+    const int n_as = src0->ne[2];
 
     char * wdata_src1_end = (src1->type == vec_dot_type) ?
             (char *) params->wdata :
             (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
 
-    struct mmid_row_mapping {
-        int32_t i1;
-        int32_t i2;
-    };
-
     int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
-    struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *)(matrix_row_counts + n_as); // [n_as][ne11]
+    int64_t * matrix_rows       = matrix_row_counts + n_as;     // [n_as][ne11]
+
+    #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
 
    if (params->type == GGML_TASK_TYPE_INIT) {
         if (ith != 0) {
@@ -11064,20 +11012,16 @@ static void ggml_compute_forward_mul_mat_id(
         }
 
         // initialize matrix_row_counts
+        GGML_ASSERT(wdata == wdata_src1_end);
         memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
 
-#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne12 + (i1)]
-
         // group rows by src0 matrix
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
-            for (int id = 0; id < n_ids; ++id) {
-                const int32_t i02 = *(const int32_t *) ((const char *) ids->data + iid1*ids->nb[1] + id*ids->nb[0]);
+        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+            const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
 
-                assert(i02 >= 0 && i02 < n_as);
-
-                MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = (struct mmid_row_mapping) {id, iid1};
-                matrix_row_counts[i02] += 1;
-            }
+            GGML_ASSERT(row_id >= 0 && row_id < n_as);
+            MMID_MATRIX_ROW(row_id, matrix_row_counts[row_id]) = i01;
+            matrix_row_counts[row_id] += 1;
         }
 
         return;
@@ -11095,13 +11039,15 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        const char * src0_cur = (const char *) src0->data + cur_a*nb02;
+        size_t src0_offset = cur_a*src0->nb[2];
 
         const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-        const int64_t nr0 = ne01; // src0 rows
-        const int64_t nr1 = cne1; // src1 rows
+        const int64_t nr0 = ne01;           // src0 rows
+        const int64_t nr1 = cne1*ne12*ne13; // src1 rows
+
+        //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
         // distribute the thread work across the inner or outer loop based on which one is larger
 
@@ -11120,11 +11066,13 @@ static void ggml_compute_forward_mul_mat_id(
         const int64_t ir110 = dr1*ith1;
         const int64_t ir111 = MIN(ir110 + dr1, nr1);
 
+        //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
         // threads with no work simply yield (not sure if it helps)
-        //if (ir010 >= ir011 || ir110 >= ir111) {
-        //    sched_yield();
-        //    continue;
-        //}
+        if (ir010 >= ir011 || ir110 >= ir111) {
+            sched_yield();
+            continue;
+        }
 
         // block-tiling attempt
         const int64_t blck_0 = 16;
@@ -11136,16 +11084,20 @@ static void ggml_compute_forward_mul_mat_id(
         for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
             for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
                 for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                    const int64_t _i12 = ir1; // logical row index for this expert
+                    const int64_t  i13 = (ir1/(ne12*cne1)); // Note: currently, src1 is always a matrix
+                    const int64_t  i12 = (ir1 - i13*ne12*cne1)/cne1;
+                    const int64_t _i11 = (ir1 - i13*ne12*cne1 - i12*cne1);
+                    const int64_t  i11 = MMID_MATRIX_ROW(cur_a, _i11);
 
-                    struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12);
-                    const int id       = row_mapping.i1; // selected expert index
+                    // broadcast src0 into src1
+                    //const int64_t i03 = i13/r3;
+                    //const int64_t i02 = i12/r2;
 
-                    const int64_t  i11 = id % ne11;
-                    const int64_t  i12 = row_mapping.i2; // row index in src1
+                    const int64_t i1 = i11;
+                    const int64_t i2 = i12;
+                    const int64_t i3 = i13;
 
-                    const int64_t  i1 = id;  // selected expert index
-                    const int64_t  i2 = i12; // row
+                    const char * src0_row = (const char *) src0->data + src0_offset;
 
                     // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
                     //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
@@ -11153,26 +11105,25 @@ static void ggml_compute_forward_mul_mat_id(
                     // TODO: this is a bit of a hack, we should probably have a better way to handle this
                     const char * src1_col = (const char *) wdata +
                         (src1_cont || src1->type != vec_dot_type
-                        ? (i11      + i12*ne11)*row_size
-                        : (i11*nb11 + i12*nb12));
+                        ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+                        : (i11*nb11 + i12*nb12 + i13*nb13));
 
-                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2));
+                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
 
                     //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
                     //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
                     //}
 
                     for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                        vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1);
+                        vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_row + ir0*nb01, 0, src1_col, 0, 1);
                     }
-
                     memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
                 }
             }
         }
     }
 
-#undef MMID_MATRIX_ROW
+    #undef MMID_MATRIX_ROW
 }
 
 // ggml_compute_forward_out_prod
@@ -18511,7 +18462,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                     const int n_as = src0->ne[2];
                     cur += GGML_PAD(cur, sizeof(int64_t));       // align
                     cur += n_as * sizeof(int64_t);               // matrix_row_counts
-                    cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
+                    cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
                 } break;
             case GGML_OP_OUT_PROD:
                 {
@@ -20599,32 +20550,6 @@ static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
     return ok;
 }
 
-static void gguf_free_kv(struct gguf_kv * kv) {
-    if (kv->key.data) {
-        GGML_FREE(kv->key.data);
-    }
-
-    if (kv->type == GGUF_TYPE_STRING) {
-        if (kv->value.str.data) {
-            GGML_FREE(kv->value.str.data);
-        }
-    }
-
-    if (kv->type == GGUF_TYPE_ARRAY) {
-        if (kv->value.arr.data) {
-            if (kv->value.arr.type == GGUF_TYPE_STRING) {
-                for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
-                    struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
-                    if (str->data) {
-                        GGML_FREE(str->data);
-                    }
-                }
-            }
-            GGML_FREE(kv->value.arr.data);
-        }
-    }
-}
-
 struct gguf_context * gguf_init_empty(void) {
     struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
 
@@ -20937,12 +20862,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
             ok = ok && cur != NULL;
 
+            ggml_set_name(cur, ctx->infos[i].name.data);
+
             if (!ok) {
                 break;
             }
 
-            ggml_set_name(cur, ctx->infos[i].name.data);
-
             // point the data member to the appropriate location in the binary blob using the tensor infos
             if (!params.no_alloc) {
               //cur->data = (char *) data->data + ctx->infos[i].offset - ctx->offset; // offset from start of file
@@ -20974,7 +20899,31 @@ void gguf_free(struct gguf_context * ctx) {
     if (ctx->kv) {
         // free string memory - not great..
         for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
-            gguf_free_kv(&ctx->kv[i]);
+            struct gguf_kv * kv = &ctx->kv[i];
+
+            if (kv->key.data) {
+                GGML_FREE(kv->key.data);
+            }
+
+            if (kv->type == GGUF_TYPE_STRING) {
+                if (kv->value.str.data) {
+                    GGML_FREE(kv->value.str.data);
+                }
+            }
+
+            if (kv->type == GGUF_TYPE_ARRAY) {
+                if (kv->value.arr.data) {
+                    if (kv->value.arr.type == GGUF_TYPE_STRING) {
+                        for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
+                            struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
+                            if (str->data) {
+                                GGML_FREE(str->data);
+                            }
+                        }
+                    }
+                    GGML_FREE(kv->value.arr.data);
+                }
+            }
         }
 
         GGML_FREE(ctx->kv);
@@ -21199,19 +21148,6 @@ static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     return n_kv;
 }
 
-void gguf_remove_key(struct gguf_context * ctx, const char * key) {
-    const int idx = gguf_find_key(ctx, key);
-    if (idx >= 0) {
-        const int n_kv = gguf_get_n_kv(ctx);
-        gguf_free_kv(&ctx->kv[idx]);
-        for (int i = idx; i < n_kv-1; ++i) {
-            ctx->kv[i] = ctx->kv[i+1];
-        }
-        ctx->kv = realloc(ctx->kv, (n_kv - 1) * sizeof(struct gguf_kv));
-        ctx->header.n_kv--;
-    }
-}
-
 void gguf_set_val_u8(struct gguf_context * ctx, const char * key, uint8_t val) {
     const int idx = gguf_get_or_add_key(ctx, key);
 

ggml.h

@@ -1161,11 +1161,13 @@ extern "C" {
             enum ggml_prec       prec);
 
     // indirect matrix multiplication
+    //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
     GGML_API struct ggml_tensor * ggml_mul_mat_id(
             struct ggml_context * ctx,
             struct ggml_tensor  * as,
-            struct ggml_tensor  * b,
-            struct ggml_tensor  * ids);
+            struct ggml_tensor  * ids,
+            int                   id,
+            struct ggml_tensor  * b);
 
     // A: m columns, n rows,
     // B: p columns, n rows,
@@ -2287,9 +2289,6 @@ extern "C" {
     GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
     GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
 
-    // removes key if it exists
-    GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
-
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
     GGML_API void gguf_set_val_i8  (struct gguf_context * ctx, const char * key, int8_t   val);

gguf-py/README.md

@@ -21,8 +21,6 @@ pip install gguf
 
 [scripts/gguf-convert-endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-convert-endian.py) — Allows converting the endianness of GGUF files.
 
-[scripts/gguf-new-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-new-metadata.py) — Copies a GGUF file with added/modified/removed metadata values.
-
 ## Development
 Maintainers who participate in development of this package are advised to install it in editable mode:
 

gguf-py/gguf/constants.py

@@ -90,13 +90,6 @@ class Keys:
         HF_JSON          = "tokenizer.huggingface.json"
         RWKV             = "tokenizer.rwkv.world"
         CHAT_TEMPLATE    = "tokenizer.chat_template"
-        CHAT_TEMPLATE_N  = "tokenizer.chat_template.{name}"
-        CHAT_TEMPLATES   = "tokenizer.chat_templates"
-        # FIM/Infill special tokens constants
-        PREFIX_ID        = "tokenizer.ggml.prefix_token_id"
-        SUFFIX_ID        = "tokenizer.ggml.suffix_token_id"
-        MIDDLE_ID        = "tokenizer.ggml.middle_token_id"
-        EOT_ID           = "tokenizer.ggml.eot_token_id"
 
 
 #
@@ -122,9 +115,7 @@ class MODEL_ARCH(IntEnum):
     STABLELM   = auto()
     QWEN       = auto()
     QWEN2      = auto()
-    QWEN2MOE   = auto()
     PHI2       = auto()
-    PHI3       = auto()
     PLAMO      = auto()
     CODESHELL  = auto()
     ORION      = auto()
@@ -135,50 +126,44 @@ class MODEL_ARCH(IntEnum):
     MAMBA      = auto()
     XVERSE     = auto()
     COMMAND_R  = auto()
-    DBRX       = auto()
-    OLMO       = auto()
 
 
 class MODEL_TENSOR(IntEnum):
-    TOKEN_EMBD         = auto()
-    TOKEN_EMBD_NORM    = auto()
-    TOKEN_TYPES        = auto()
-    POS_EMBD           = auto()
-    OUTPUT             = auto()
-    OUTPUT_NORM        = auto()
-    ROPE_FREQS         = auto()
-    ATTN_Q             = auto()
-    ATTN_K             = auto()
-    ATTN_V             = auto()
-    ATTN_QKV           = auto()
-    ATTN_OUT           = auto()
-    ATTN_NORM          = auto()
-    ATTN_NORM_2        = auto()
-    ATTN_OUT_NORM      = auto()
-    ATTN_ROT_EMBD      = auto()
-    FFN_GATE_INP       = auto()
-    FFN_GATE_INP_SHEXP = auto()
-    FFN_NORM           = auto()
-    FFN_GATE           = auto()
-    FFN_DOWN           = auto()
-    FFN_UP             = auto()
-    FFN_ACT            = auto()
-    FFN_GATE_EXP       = auto()
-    FFN_DOWN_EXP       = auto()
-    FFN_UP_EXP         = auto()
-    FFN_GATE_SHEXP     = auto()
-    FFN_DOWN_SHEXP     = auto()
-    FFN_UP_SHEXP       = auto()
-    ATTN_Q_NORM        = auto()
-    ATTN_K_NORM        = auto()
-    LAYER_OUT_NORM     = auto()
-    SSM_IN             = auto()
-    SSM_CONV1D         = auto()
-    SSM_X              = auto()
-    SSM_DT             = auto()
-    SSM_A              = auto()
-    SSM_D              = auto()
-    SSM_OUT            = auto()
+    TOKEN_EMBD      = auto()
+    TOKEN_EMBD_NORM = auto()
+    TOKEN_TYPES     = auto()
+    POS_EMBD        = auto()
+    OUTPUT          = auto()
+    OUTPUT_NORM     = auto()
+    ROPE_FREQS      = auto()
+    ATTN_Q          = auto()
+    ATTN_K          = auto()
+    ATTN_V          = auto()
+    ATTN_QKV        = auto()
+    ATTN_OUT        = auto()
+    ATTN_NORM       = auto()
+    ATTN_NORM_2     = auto()
+    ATTN_OUT_NORM   = auto()
+    ATTN_ROT_EMBD   = auto()
+    FFN_GATE_INP    = auto()
+    FFN_NORM        = auto()
+    FFN_GATE        = auto()
+    FFN_DOWN        = auto()
+    FFN_UP          = auto()
+    FFN_ACT         = auto()
+    FFN_GATE_EXP    = auto()
+    FFN_DOWN_EXP    = auto()
+    FFN_UP_EXP      = auto()
+    ATTN_Q_NORM     = auto()
+    ATTN_K_NORM     = auto()
+    LAYER_OUT_NORM  = auto()
+    SSM_IN          = auto()
+    SSM_CONV1D      = auto()
+    SSM_X           = auto()
+    SSM_DT          = auto()
+    SSM_A           = auto()
+    SSM_D           = auto()
+    SSM_OUT         = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -199,9 +184,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
     MODEL_ARCH.QWEN2:          "qwen2",
-    MODEL_ARCH.QWEN2MOE:       "qwen2moe",
     MODEL_ARCH.PHI2:           "phi2",
-    MODEL_ARCH.PHI3:           "phi3",
     MODEL_ARCH.PLAMO:          "plamo",
     MODEL_ARCH.CODESHELL:      "codeshell",
     MODEL_ARCH.ORION:          "orion",
@@ -212,50 +195,44 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.MAMBA:          "mamba",
     MODEL_ARCH.XVERSE:         "xverse",
     MODEL_ARCH.COMMAND_R:      "command-r",
-    MODEL_ARCH.DBRX:           "dbrx",
-    MODEL_ARCH.OLMO:           "olmo",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
-    MODEL_TENSOR.TOKEN_EMBD:         "token_embd",
-    MODEL_TENSOR.TOKEN_EMBD_NORM:    "token_embd_norm",
-    MODEL_TENSOR.TOKEN_TYPES:        "token_types",
-    MODEL_TENSOR.POS_EMBD:           "position_embd",
-    MODEL_TENSOR.OUTPUT_NORM:        "output_norm",
-    MODEL_TENSOR.OUTPUT:             "output",
-    MODEL_TENSOR.ROPE_FREQS:         "rope_freqs",
-    MODEL_TENSOR.ATTN_NORM:          "blk.{bid}.attn_norm",
-    MODEL_TENSOR.ATTN_NORM_2:        "blk.{bid}.attn_norm_2",
-    MODEL_TENSOR.ATTN_QKV:           "blk.{bid}.attn_qkv",
-    MODEL_TENSOR.ATTN_Q:             "blk.{bid}.attn_q",
-    MODEL_TENSOR.ATTN_K:             "blk.{bid}.attn_k",
-    MODEL_TENSOR.ATTN_V:             "blk.{bid}.attn_v",
-    MODEL_TENSOR.ATTN_OUT:           "blk.{bid}.attn_output",
-    MODEL_TENSOR.ATTN_ROT_EMBD:      "blk.{bid}.attn_rot_embd",
-    MODEL_TENSOR.ATTN_Q_NORM:        "blk.{bid}.attn_q_norm",
-    MODEL_TENSOR.ATTN_K_NORM:        "blk.{bid}.attn_k_norm",
-    MODEL_TENSOR.ATTN_OUT_NORM:      "blk.{bid}.attn_output_norm",
-    MODEL_TENSOR.FFN_GATE_INP:       "blk.{bid}.ffn_gate_inp",
-    MODEL_TENSOR.FFN_GATE_INP_SHEXP: "blk.{bid}.ffn_gate_inp_shexp",
-    MODEL_TENSOR.FFN_NORM:           "blk.{bid}.ffn_norm",
-    MODEL_TENSOR.FFN_GATE:           "blk.{bid}.ffn_gate",
-    MODEL_TENSOR.FFN_DOWN:           "blk.{bid}.ffn_down",
-    MODEL_TENSOR.FFN_UP:             "blk.{bid}.ffn_up",
-    MODEL_TENSOR.FFN_GATE_SHEXP:     "blk.{bid}.ffn_gate_shexp",
-    MODEL_TENSOR.FFN_DOWN_SHEXP:     "blk.{bid}.ffn_down_shexp",
-    MODEL_TENSOR.FFN_UP_SHEXP:       "blk.{bid}.ffn_up_shexp",
-    MODEL_TENSOR.FFN_ACT:            "blk.{bid}.ffn",
-    MODEL_TENSOR.FFN_GATE_EXP:       "blk.{bid}.ffn_gate_exps",
-    MODEL_TENSOR.FFN_DOWN_EXP:       "blk.{bid}.ffn_down_exps",
-    MODEL_TENSOR.FFN_UP_EXP:         "blk.{bid}.ffn_up_exps",
-    MODEL_TENSOR.LAYER_OUT_NORM:     "blk.{bid}.layer_output_norm",
-    MODEL_TENSOR.SSM_IN:             "blk.{bid}.ssm_in",
-    MODEL_TENSOR.SSM_CONV1D:         "blk.{bid}.ssm_conv1d",
-    MODEL_TENSOR.SSM_X:              "blk.{bid}.ssm_x",
-    MODEL_TENSOR.SSM_DT:             "blk.{bid}.ssm_dt",
-    MODEL_TENSOR.SSM_A:              "blk.{bid}.ssm_a",
-    MODEL_TENSOR.SSM_D:              "blk.{bid}.ssm_d",
-    MODEL_TENSOR.SSM_OUT:            "blk.{bid}.ssm_out",
+    MODEL_TENSOR.TOKEN_EMBD:      "token_embd",
+    MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm",
+    MODEL_TENSOR.TOKEN_TYPES:     "token_types",
+    MODEL_TENSOR.POS_EMBD:        "position_embd",
+    MODEL_TENSOR.OUTPUT_NORM:     "output_norm",
+    MODEL_TENSOR.OUTPUT:          "output",
+    MODEL_TENSOR.ROPE_FREQS:      "rope_freqs",
+    MODEL_TENSOR.ATTN_NORM:       "blk.{bid}.attn_norm",
+    MODEL_TENSOR.ATTN_NORM_2:     "blk.{bid}.attn_norm_2",
+    MODEL_TENSOR.ATTN_QKV:        "blk.{bid}.attn_qkv",
+    MODEL_TENSOR.ATTN_Q:          "blk.{bid}.attn_q",
+    MODEL_TENSOR.ATTN_K:          "blk.{bid}.attn_k",
+    MODEL_TENSOR.ATTN_V:          "blk.{bid}.attn_v",
+    MODEL_TENSOR.ATTN_OUT:        "blk.{bid}.attn_output",
+    MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
+    MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
+    MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.ATTN_OUT_NORM:   "blk.{bid}.attn_output_norm",
+    MODEL_TENSOR.FFN_GATE_INP:    "blk.{bid}.ffn_gate_inp",
+    MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
+    MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
+    MODEL_TENSOR.FFN_DOWN:        "blk.{bid}.ffn_down",
+    MODEL_TENSOR.FFN_UP:          "blk.{bid}.ffn_up",
+    MODEL_TENSOR.FFN_ACT:         "blk.{bid}.ffn",
+    MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate_exps",
+    MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down_exps",
+    MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up_exps",
+    MODEL_TENSOR.LAYER_OUT_NORM:  "blk.{bid}.layer_output_norm",
+    MODEL_TENSOR.SSM_IN:          "blk.{bid}.ssm_in",
+    MODEL_TENSOR.SSM_CONV1D:      "blk.{bid}.ssm_conv1d",
+    MODEL_TENSOR.SSM_X:           "blk.{bid}.ssm_x",
+    MODEL_TENSOR.SSM_DT:          "blk.{bid}.ssm_dt",
+    MODEL_TENSOR.SSM_A:           "blk.{bid}.ssm_a",
+    MODEL_TENSOR.SSM_D:           "blk.{bid}.ssm_d",
+    MODEL_TENSOR.SSM_OUT:         "blk.{bid}.ssm_out",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -461,8 +438,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
-        MODEL_TENSOR.ATTN_Q_NORM,
-        MODEL_TENSOR.ATTN_K_NORM,
     ],
     MODEL_ARCH.QWEN: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -492,25 +467,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.QWEN2MOE: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_GATE_INP,
-        MODEL_TENSOR.FFN_GATE_EXP,
-        MODEL_TENSOR.FFN_DOWN_EXP,
-        MODEL_TENSOR.FFN_UP_EXP,
-        MODEL_TENSOR.FFN_GATE_INP_SHEXP,
-        MODEL_TENSOR.FFN_GATE_SHEXP,
-        MODEL_TENSOR.FFN_DOWN_SHEXP,
-        MODEL_TENSOR.FFN_UP_SHEXP,
-    ],
     MODEL_ARCH.PLAMO: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -552,20 +508,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.PHI3: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_QKV,
-        MODEL_TENSOR.ATTN_Q,
-        MODEL_TENSOR.ATTN_K,
-        MODEL_TENSOR.ATTN_V,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_DOWN,
-        MODEL_TENSOR.FFN_UP,
-    ],
     MODEL_ARCH.CODESHELL: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.POS_EMBD,
@@ -700,30 +642,6 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_K_NORM,
         MODEL_TENSOR.ATTN_Q_NORM,
     ],
-    MODEL_ARCH.DBRX: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.OUTPUT,
-        MODEL_TENSOR.ATTN_NORM,
-        MODEL_TENSOR.ATTN_QKV,
-        MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.ATTN_OUT_NORM,
-        MODEL_TENSOR.FFN_GATE_INP,
-        MODEL_TENSOR.FFN_GATE_EXP,
-        MODEL_TENSOR.FFN_DOWN_EXP,
-        MODEL_TENSOR.FFN_UP_EXP,
-    ],
-    MODEL_ARCH.OLMO: [
-        MODEL_TENSOR.TOKEN_EMBD,
-        MODEL_TENSOR.OUTPUT,
-        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
 }
 
@@ -888,7 +806,6 @@ GGML_QUANT_SIZES = {
     GGMLQuantizationType.I32:     (1, 4),
     GGMLQuantizationType.I64:     (1, 8),
     GGMLQuantizationType.F64:     (1, 8),
-    GGMLQuantizationType.IQ1_M:   (256, QK_K // 8 + QK_K // 16  + QK_K // 32),
 }
 
 
@@ -953,7 +870,3 @@ KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
 KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV
-KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
-KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
-KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
-KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID

gguf-py/gguf/gguf_writer.py

@@ -6,8 +6,7 @@ import struct
 import tempfile
 from enum import Enum, auto
 from io import BufferedWriter
-from typing import IO, Any, Sequence, Mapping
-from string import ascii_letters, digits
+from typing import IO, Any, Sequence
 
 import numpy as np
 
@@ -467,47 +466,9 @@ class GGUFWriter:
     def add_add_space_prefix(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
 
-    def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
-        if isinstance(value, list):
-            template_default = None
-            template_names = set()
-
-            for choice in value:
-                name = choice.get('name', '')
-                template = choice.get('template')
-
-                # Allowing non-alphanumerical characters in template name is probably not a good idea, so filter it
-                name = ''.join((c if c in ascii_letters + digits else '_' for c in name))
-
-                if name and template is not None:
-                    if name == 'default':
-                        template_default = template
-                    else:
-                        template_names.add(name)
-                        self.add_string(Keys.Tokenizer.CHAT_TEMPLATE_N.format(name=name), template)
-
-            if template_names:
-                self.add_array(Keys.Tokenizer.CHAT_TEMPLATES, list(template_names))
-
-            if template_default is None:
-                return
-
-            value = template_default
-
+    def add_chat_template(self, value: str) -> None:
         self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
 
-    def add_prefix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
-
-    def add_suffix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
-
-    def add_middle_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
-
-    def add_eot_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.EOT_ID, id)
-
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -10,7 +10,7 @@ class TensorNameMap:
         # Token embeddings
         MODEL_TENSOR.TOKEN_EMBD: (
             "gpt_neox.embed_in",                         # gptneox
-            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx
+            "transformer.wte",                           # gpt2 gpt-j mpt refact qwen
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
@@ -48,7 +48,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -60,7 +60,7 @@ class TensorNameMap:
             "transformer.ln_f",                        # gpt2 gpt-j falcon
             "model.norm",                              # llama-hf baichuan internlm2
             "norm",                                    # llama-pth
-            "transformer.norm_f",                      # mpt dbrx
+            "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
             "model.final_layernorm",                   # persimmon
@@ -96,7 +96,6 @@ class TensorNameMap:
             "model.layers.{bid}.norm",                              # mamba-qbert
             "backbone.layers.{bid}.norm",                           # mamba
             "transformer.decoder_layer.{bid}.rms_norm",             # Grok
-            "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
         ),
 
         # Attention norm 2
@@ -109,7 +108,6 @@ class TensorNameMap:
             "gpt_neox.layers.{bid}.attention.query_key_value",                     # gptneox
             "transformer.h.{bid}.attn.c_attn",                                     # gpt2 qwen
             "transformer.blocks.{bid}.attn.Wqkv",                                  # mpt
-            "transformer.blocks.{bid}.norm_attn_norm.attn.Wqkv",                   # dbrx
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
@@ -117,7 +115,6 @@ class TensorNameMap:
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
             "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
-            "model.layers.{bid}.self_attn.qkv_proj"                                # phi3
         ),
 
         # Attention query
@@ -155,24 +152,23 @@ class TensorNameMap:
 
         # Attention output
         MODEL_TENSOR.ATTN_OUT: (
-            "gpt_neox.layers.{bid}.attention.dense",                        # gptneox
-            "transformer.h.{bid}.attn.c_proj",                              # gpt2 refact qwen
-            "transformer.blocks.{bid}.attn.out_proj",                       # mpt
-            "transformer.h.{bid}.self_attention.dense",                     # falcon
-            "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf
-            "layers.{bid}.attention.wo",                                    # llama-pth
-            "encoder.layer.{bid}.attention.output.dense",                   # bert
-            "transformer.h.{bid}.attn.out_proj",                            # gpt-j
-            "language_model.encoder.layers.{bid}.self_attention.dense",     # persimmon
-            "model.layers.{bid}.self_attn.dense",                           # persimmon
-            "h.{bid}.attn.c_proj",                                          # gpt2
-            "transformer.h.{bid}.mixer.out_proj",                           # phi2
-            "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
-            "model.layers.{bid}.attention.wo",                              # internlm2
-            "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
-            "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
-            "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
+            "gpt_neox.layers.{bid}.attention.dense",                     # gptneox
+            "transformer.h.{bid}.attn.c_proj",                           # gpt2 refact qwen
+            "transformer.blocks.{bid}.attn.out_proj",                    # mpt
+            "transformer.h.{bid}.self_attention.dense",                  # falcon
+            "h.{bid}.self_attention.dense",                              # bloom
+            "model.layers.{bid}.self_attn.o_proj",                       # llama-hf
+            "layers.{bid}.attention.wo",                                 # llama-pth
+            "encoder.layer.{bid}.attention.output.dense",                # bert
+            "transformer.h.{bid}.attn.out_proj",                         # gpt-j
+            "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
+            "model.layers.{bid}.self_attn.dense",                        # persimmon
+            "h.{bid}.attn.c_proj",                                       # gpt2
+            "transformer.h.{bid}.mixer.out_proj",                        # phi2
+            "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
+            "model.layers.{bid}.attention.wo",                           # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                        # nomic-bert
+            "transformer.decoder_layer.{bid}.multi_head_attention.linear"# Grok
         ),
 
         # Attention output norm
@@ -180,7 +176,6 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
             "encoder.layers.{bid}.norm1",                      # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_1",      # Grok
-            "transformer.blocks.{bid}.norm_attn_norm.norm_2",  # dbrx
         ),
 
         # Rotary embeddings
@@ -207,15 +202,9 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
-            "layers.{bid}.feed_forward.gate",             # mixtral
-            "model.layers.{bid}.block_sparse_moe.gate",   # mixtral
-            "model.layers.{bid}.mlp.gate",                # qwen2moe
-            "transformer.decoder_layer.{bid}.router",     # Grok
-            "transformer.blocks.{bid}.ffn.router.layer",  # dbrx
-        ),
-
-        MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert_gate", # qwen2moe
+            "layers.{bid}.feed_forward.gate",           # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate", # mixtral
+            "transformer.decoder_layer.{bid}.router"    # Grok
         ),
 
         # Feed-forward up
@@ -235,7 +224,6 @@ class TensorNameMap:
             "h.{bid}.mlp.c_fc",                                       # gpt2
             "transformer.h.{bid}.mlp.fc1",                            # phi2
             "model.layers.{bid}.mlp.fc1",                             # phi2
-            "model.layers.{bid}.mlp.gate_up_proj",                    # phi3
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
@@ -243,14 +231,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
-            "layers.{bid}.feed_forward.experts.w3",          # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear_v",  # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.v1",   # dbrx
-            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe (merged)
-        ),
-
-        MODEL_TENSOR.FFN_UP_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.up_proj",  # qwen2moe
+            "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
         ),
 
         # AWQ-activation gate
@@ -269,14 +251,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
-            "layers.{bid}.feed_forward.experts.w1",         # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear",   # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w1",  # dbrx
-            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe (merged)
-        ),
-
-        MODEL_TENSOR.FFN_GATE_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.gate_proj",  # qwen2moe
+            "layers.{bid}.feed_forward.experts.w1",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear"            # Grok (merged)
         ),
 
         # Feed-forward down
@@ -302,14 +278,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
-            "layers.{bid}.feed_forward.experts.w2",          # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear_1",  # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w2",   # dbrx
-            "model.layers.{bid}.mlp.experts.down_proj",      # qwen2moe (merged)
-        ),
-
-        MODEL_TENSOR.FFN_DOWN_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.down_proj",  # qwen2moe
+            "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (
@@ -388,7 +358,7 @@ class TensorNameMap:
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
                 # TODO: make this configurable
-                n_experts = 60
+                n_experts = 8
                 for xid in range(n_experts):
                     tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
                     self.mapping[tensor_name] = (tensor, tensor_name)

gguf-py/gguf/vocab.py

@@ -141,7 +141,7 @@ class SpecialVocab:
         with open(tokenizer_config_file, encoding = 'utf-8') as f:
             tokenizer_config = json.load(f)
         chat_template = tokenizer_config.get('chat_template')
-        if chat_template is None or isinstance(chat_template, (str, list)):
+        if chat_template is None or isinstance(chat_template, str):
             self.chat_template = chat_template
         else:
             print(

gguf-py/pyproject.toml

@@ -33,4 +33,3 @@ build-backend = "poetry.core.masonry.api"
 gguf-convert-endian = "scripts:gguf_convert_endian_entrypoint"
 gguf-dump = "scripts:gguf_dump_entrypoint"
 gguf-set-metadata = "scripts:gguf_set_metadata_entrypoint"
-gguf-new-metadata = "scripts:gguf_new_metadata_entrypoint"

gguf-py/scripts/__init__.py

@@ -8,6 +8,5 @@ os.environ["NO_LOCAL_GGUF"] = "TRUE"
 gguf_convert_endian_entrypoint = import_module("scripts.gguf-convert-endian").main
 gguf_dump_entrypoint           = import_module("scripts.gguf-dump").main
 gguf_set_metadata_entrypoint   = import_module("scripts.gguf-set-metadata").main
-gguf_new_metadata_entrypoint   = import_module("scripts.gguf-new-metadata").main
 
 del import_module, os

gguf-py/scripts/gguf-new-metadata.py

@@ -1,190 +0,0 @@
-#!/usr/bin/env python3
-import logging
-import argparse
-import os
-import sys
-import json
-from pathlib import Path
-
-import numpy as np
-from typing import Any, Mapping, Sequence
-
-# Necessary to load the local gguf package
-if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
-    sys.path.insert(0, str(Path(__file__).parent.parent))
-
-import gguf
-
-logger = logging.getLogger("gguf-new-metadata")
-
-
-def get_byteorder(reader: gguf.GGUFReader) -> gguf.GGUFEndian:
-    if np.uint32(1) == np.uint32(1).newbyteorder("<"):
-        # Host is little endian
-        host_endian = gguf.GGUFEndian.LITTLE
-        swapped_endian = gguf.GGUFEndian.BIG
-    else:
-        # Sorry PDP or other weird systems that don't use BE or LE.
-        host_endian = gguf.GGUFEndian.BIG
-        swapped_endian = gguf.GGUFEndian.LITTLE
-
-    if reader.byte_order == "S":
-        return swapped_endian
-    else:
-        return host_endian
-
-
-def decode_field(field: gguf.ReaderField) -> Any:
-    if field and field.types:
-        main_type = field.types[0]
-
-        if main_type == gguf.GGUFValueType.ARRAY:
-            sub_type = field.types[-1]
-
-            if sub_type == gguf.GGUFValueType.STRING:
-                return [str(bytes(field.parts[idx]), encoding='utf8') for idx in field.data]
-            else:
-                return [pv for idx in field.data for pv in field.parts[idx].tolist()]
-        if main_type == gguf.GGUFValueType.STRING:
-            return str(bytes(field.parts[-1]), encoding='utf8')
-        else:
-            return field.parts[-1][0]
-
-    return None
-
-
-def get_field_data(reader: gguf.GGUFReader, key: str) -> Any:
-    field = reader.get_field(key)
-
-    return decode_field(field)
-
-
-def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: Mapping[str, str], remove_metadata: Sequence[str]) -> None:
-    for field in reader.fields.values():
-        # Suppress virtual fields and fields written by GGUFWriter
-        if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'):
-            logger.debug(f'Suppressing {field.name}')
-            continue
-
-        # Skip old chat templates if we have new ones
-        if field.name.startswith(gguf.Keys.Tokenizer.CHAT_TEMPLATE) and gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
-            logger.debug(f'Skipping {field.name}')
-            continue
-
-        if field.name in remove_metadata:
-            logger.debug(f'Removing {field.name}')
-            continue
-
-        old_val = decode_field(field)
-        val = new_metadata.get(field.name, old_val)
-
-        if field.name in new_metadata:
-            logger.debug(f'Modifying {field.name}: "{old_val}" -> "{val}"')
-            del new_metadata[field.name]
-        elif val is not None:
-            logger.debug(f'Copying {field.name}')
-
-        if val is not None:
-            writer.add_key(field.name)
-            writer.add_val(val, field.types[0])
-
-    if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
-        logger.debug('Adding chat template(s)')
-        writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE])
-        del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE]
-
-    # TODO: Support other types than string?
-    for key, val in new_metadata.items():
-        logger.debug(f'Adding {key}: {val}')
-        writer.add_key(key)
-        writer.add_val(val, gguf.GGUFValueType.STRING)
-
-    for tensor in reader.tensors:
-        # Dimensions are written in reverse order, so flip them first
-        shape = np.flipud(tensor.shape)
-        writer.add_tensor_info(tensor.name, shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
-
-    writer.write_header_to_file()
-    writer.write_kv_data_to_file()
-    writer.write_ti_data_to_file()
-
-    for tensor in reader.tensors:
-        writer.write_tensor_data(tensor.data)
-
-    writer.close()
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser(description="Make a copy of a GGUF file with new metadata")
-    parser.add_argument("input",                                       type=Path, help="GGUF format model input filename")
-    parser.add_argument("output",                                      type=Path, help="GGUF format model output filename")
-    parser.add_argument("--general-name",                              type=str,  help="The models general.name")
-    parser.add_argument("--general-description",                       type=str,  help="The models general.description")
-    parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)")
-    parser.add_argument("--chat-template-config",                      type=Path, help="Config file (tokenizer_config.json) containing chat template(s)")
-    parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model")
-    parser.add_argument("--force",                action="store_true",            help="Bypass warnings without confirmation")
-    parser.add_argument("--verbose",              action="store_true",            help="Increase output verbosity")
-    args = parser.parse_args(None if len(sys.argv) > 2 else ["--help"])
-
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-
-    new_metadata = {}
-    remove_metadata = args.remove_metadata or []
-
-    if args.general_name:
-        new_metadata[gguf.Keys.General.NAME] = args.general_name
-
-    if args.general_description:
-        new_metadata[gguf.Keys.General.DESCRIPTION] = args.general_description
-
-    if args.chat_template:
-        new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template
-
-    if args.chat_template_config:
-        with open(args.chat_template_config, 'r') as fp:
-            config = json.load(fp)
-            template = config.get('chat_template')
-            if template:
-                new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = template
-
-    if remove_metadata:
-        logger.warning('*** Warning *** Warning *** Warning **')
-        logger.warning('* Most metadata is required for a fully functional GGUF file,')
-        logger.warning('* removing crucial metadata may result in a corrupt output file!')
-
-        if not args.force:
-            logger.warning('* Enter exactly YES if you are positive you want to proceed:')
-            response = input('YES, I am sure> ')
-            if response != 'YES':
-                logger.info("You didn't enter YES. Okay then, see ya!")
-                sys.exit(0)
-
-    logger.info(f'* Loading: {args.input}')
-    reader = gguf.GGUFReader(args.input, 'r')
-
-    arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE)
-    endianess = get_byteorder(reader)
-
-    if os.path.isfile(args.output) and not args.force:
-        logger.warning('*** Warning *** Warning *** Warning **')
-        logger.warning(f'* The "{args.output}" GGUF file already exists, it will be overwritten!')
-        logger.warning('* Enter exactly YES if you are positive you want to proceed:')
-        response = input('YES, I am sure> ')
-        if response != 'YES':
-            logger.info("You didn't enter YES. Okay then, see ya!")
-            sys.exit(0)
-
-    logger.info(f'* Writing: {args.output}')
-    writer = gguf.GGUFWriter(args.output, arch=arch, endianess=endianess)
-
-    alignment = get_field_data(reader, gguf.Keys.General.ALIGNMENT)
-    if alignment is not None:
-        logger.debug(f'Setting custom alignment: {alignment}')
-        writer.data_alignment = alignment
-
-    copy_with_new_metadata(reader, writer, new_metadata, remove_metadata)
-
-
-if __name__ == '__main__':
-    main()

grammars/README.md

@@ -89,13 +89,3 @@ This guide provides a brief overview. Check out the GBNF files in this directory
 ```
 ./main -m <model> --grammar-file grammars/some-grammar.gbnf -p 'Some prompt'
 ```
-
-## Troubleshooting
-
-Grammars currently have performance gotchas (see https://github.com/ggerganov/llama.cpp/issues/4218).
-
-### Efficient optional repetitions
-
-A common pattern is to allow repetitions of a pattern `x` up to N times.
-
-While semantically correct, the syntax `x? x? x?.... x?` (with N repetitions) will result in extremely slow inference. Instead, you can write `(x (x (x ... (x)?...)?)?)?` (w/ N-deep nesting)