common : disable repeat penalties by default (#6127 )

ci : exempt some labels from being tagged as stale (#6140 )
common : print usage on '-h' and '--help' (#6145 )
2026-07-31 00:28:04 +02:00 · 2024-03-19 10:21:54 +02:00 · 2024-03-19 10:06:54 +02:00 · 2024-03-19 07:59:36 +02:00 · 2024-03-18 18:51:30 +00:00 · 2024-03-18 12:49:02 -04:00
72 changed files with 5524 additions and 3580 deletions
@@ -48,6 +48,28 @@ jobs:
          CC=gcc-8 make tests -j $(nproc)
          make test -j $(nproc)

+  ubuntu-focal-make-curl:
+    runs-on: ubuntu-20.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
+
+      - name: Build
+        id: make_build
+        env:
+          LLAMA_FATAL_WARNINGS: 1
+          LLAMA_CURL: 1
+        run: |
+          CC=gcc-8 make -j $(nproc)
+
  ubuntu-latest-cmake:
    runs-on: ubuntu-latest

@@ -76,40 +98,40 @@ jobs:
          cd build
          ctest -L main --verbose --timeout 900

-  ubuntu-latest-cmake-sanitizer:
-    runs-on: ubuntu-latest
-
-    continue-on-error: true
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
+#  ubuntu-latest-cmake-sanitizer:
+#    runs-on: ubuntu-latest
+#
+#    continue-on-error: true
+#
+#    strategy:
+#      matrix:
+#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+#        build_type: [Debug, Release]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v3
+#
+#      - name: Dependencies
+#        id: depends
+#        run: |
+#          sudo apt-get update
+#          sudo apt-get install build-essential
+#
+#      - name: Build
+#        id: cmake_build
+#        run: |
+#          mkdir build
+#          cd build
+#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+#
+#      - name: Test
+#        id: cmake_test
+#        run: |
+#          cd build
+#          ctest -L main --verbose --timeout 900

  ubuntu-latest-cmake-mpi:
    runs-on: ubuntu-latest
@@ -333,6 +355,7 @@ jobs:
          mkdir build
          cd build
          cmake -G Xcode .. \
+            -DLLAMA_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
@@ -361,6 +384,7 @@ jobs:
          mkdir build
          cd build
          cmake -G Xcode .. \
+            -DLLAMA_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_EXAMPLES=OFF \
            -DLLAMA_BUILD_TESTS=OFF \
            -DLLAMA_BUILD_SERVER=OFF \
@@ -0,0 +1,23 @@
+name: Close inactive issues
+on:
+  schedule:
+    - cron: "42 0 * * *"
+
+jobs:
+  close-issues:
+    runs-on: ubuntu-latest
+    permissions:
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v5
+        with:
+          exempt-issue-labels: "refactor,help wanted,good first issue,research"
+          days-before-issue-stale: 30
+          days-before-issue-close: 14
+          stale-issue-label: "stale"
+          close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+          operations-per-run: 1000
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
@@ -24,18 +24,15 @@ jobs:

    strategy:
      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        # TODO: temporary disabled due to linux kernel issues
+        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        sanitizer: [UNDEFINED]
+        build_type: [Debug]
        include:
          - build_type: Release
            sanitizer: ""
-        exclude:
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
-            sanitizer: THREAD
-          - build_type: Release
-            sanitizer: UNDEFINED
+            disabled_on_pr: true
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken

    container:
      image: ubuntu:latest
@@ -60,7 +57,8 @@ jobs:
            cmake \
            python3-pip \
            wget \
-            language-pack-en
+            language-pack-en \
+            libcurl4-openssl-dev

      - name: Build
        id: cmake_build
@@ -70,6 +68,7 @@ jobs:
          cmake .. \
              -DLLAMA_NATIVE=OFF \
              -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
          cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
@@ -81,13 +80,14 @@ jobs:

      - name: Tests
        id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
          cd examples/server/tests
          PORT=8888 ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' && matrix.build_type == 'Release' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
          cd examples/server/tests
          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
@@ -103,12 +103,21 @@ jobs:
        with:
          fetch-depth: 0

+      - name: libCURL
+        id: get_libcurl
+        env:
+          CURL_VERSION: 8.6.0_6
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+          mkdir $env:RUNNER_TEMP/libcurl
+          tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
+
      - name: Build
        id: cmake_build
        run: |
          mkdir build
          cd build
-          cmake ..  -DLLAMA_BUILD_SERVER=ON -DCMAKE_BUILD_TYPE=Release ;
+          cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server

      - name: Python setup
@@ -122,15 +131,21 @@ jobs:
        run: |
          pip install -r examples/server/tests/requirements.txt

+      - name: Copy Libcurl
+        id: prepare_libcurl
+        run: |
+          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
+
      - name: Tests
        id: server_integration_tests
+        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
          cd examples/server/tests
          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp

      - name: Slow tests
        id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' || github.event.inputs.slow_tests == 'true' }}
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
        run: |
          cd examples/server/tests
          behave.exe --stop --no-skipped --no-capture --tags slow
@@ -1,20 +0,0 @@
-name: clang-tidy review post comments
-
-on:
-  workflow_dispatch:
-    workflows: ["clang-tidy-review"]
-    types:
-      - completed
-
-jobs:
-  build:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: ZedThree/clang-tidy-review/post@v0.13.0
-        # lgtm_comment_body, max_comments, and annotations need to be set on the posting workflow in a split setup
-        with:
-          # adjust options as necessary
-          lgtm_comment_body: ''
-          annotations: false
-          max_comments: 25
@@ -1,23 +0,0 @@
-name: clang-tidy-review
-
-on:
-  pull_request:
-    branches:
-      - master
-
-jobs:
-  clang-tidy-review:
-    runs-on: ubuntu-latest
-
-    steps:
-    - uses: actions/checkout@v3
-
-    - uses: ZedThree/clang-tidy-review@v0.13.0
-      id: review
-      with:
-        lgtm_comment_body: ''
-        build_dir: build
-        cmake_command: cmake . -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=on
-        split_workflow: true
-
-    - uses: ZedThree/clang-tidy-review/upload@v0.13.0
@@ -25,6 +25,8 @@
 .vscode/
 .idea/

+ggml-metal-embed.metal
+
 lcov-report/
 gcovr-report/

@@ -99,6 +99,7 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "llama: max. batch size for using peer access")
+option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
@@ -118,6 +119,7 @@ option(LLAMA_SYCL                            "llama: use SYCL"
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
 option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
+set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")

 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
@@ -147,6 +149,8 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)

+add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
    add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -197,9 +201,6 @@ if (LLAMA_METAL)
        add_compile_definitions(GGML_METAL_NDEBUG)
    endif()

-    # get full path to the file
-    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
-
    # copy ggml-common.h and ggml-metal.metal to bin directory
    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
@@ -208,53 +209,62 @@ if (LLAMA_METAL)
        enable_language(ASM)
        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)

+        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+
        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+        # merge ggml-common.h and ggml-metal.metal into a single file
+        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")

        add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
+            OUTPUT ${METALLIB_EMBED_ASM}
+            COMMAND echo "Embedding Metal library"
+            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
+            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+            DEPENDS ggml-metal.metal ggml-common.h
            COMMENT "Generate assembly for embedded Metal library"
        )

-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
-    if (LLAMA_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
-        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        if (LLAMA_QKK_64)
-            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
+        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
+    else()
+        if (LLAMA_METAL_SHADER_DEBUG)
+            # custom command to do the following:
+            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+            #
+            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+            #       disabling fast math is needed in order to pass tests/test-backend-ops
+            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+            set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        else()
+            set(XC_FLAGS -O3)
        endif()

        add_custom_command(
            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DEPENDS ggml-metal.metal
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+            DEPENDS ggml-metal.metal ggml-common.h
            COMMENT "Compiling Metal kernels"
-        )
+            )

        add_custom_target(
            ggml-metal ALL
            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
-    endif()
+            )
+    endif() # LLAMA_METAL_EMBED_LIBRARY

    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
        ${FOUNDATION_LIBRARY}
@@ -167,6 +167,10 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif

+ifdef LLAMA_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+endif
+
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
@@ -549,19 +553,20 @@ endif
 endif # LLAMA_METAL

 ifdef LLAMA_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h
+ggml-metal.o: ggml-metal.m ggml-metal.h ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@

 ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal
+ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 	@echo "Embedding Metal library"
+	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
 	@$(AS) $(TEMP_ASSEMBLY) -o $@
 	@rm -f ${TEMP_ASSEMBLY}
 endif
@@ -590,6 +595,11 @@ include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif

+ifdef LLAMA_CURL
+override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
+endif
+
 #
 # Print build information
 #
@@ -10,12 +10,14 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)

 ### Recent API changes

- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
+- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
+- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849

 ### Hot topics

+- Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
 - Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
 - Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
@@ -110,6 +112,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
 - [x] [Gemma](https://ai.google.dev/gemma)
 - [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)

 **Multimodal models:**

@@ -131,6 +134,7 @@ Typically finetunes of the base models below are supported as well.
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
+- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
@@ -902,6 +906,9 @@ First, install the essential packages for termux:
 pkg install clang wget git cmake
 ```
 Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
+
+You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
+
 ```
 $ mkdir build-android
 $ cd build-android
@@ -910,7 +917,28 @@ $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROI
 $ make
 ```
 Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
-Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+```
+
+Here is a demo of an interactive session running on Pixel 5 phone:

 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4

@@ -68,6 +68,17 @@ if (BUILD_SHARED_LIBS)
    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

+set(LLAMA_COMMON_EXTRA_LIBS build_info)
+
+# Use curl to download model url
+if (LLAMA_CURL)
+    find_package(CURL REQUIRED)
+    add_definitions(-DLLAMA_USE_CURL)
+    include_directories(${CURL_INCLUDE_DIRS})
+    find_library(CURL_LIBRARY curl REQUIRED)
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+endif ()
+
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)
+target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)
@@ -37,10 +37,13 @@ extern char const *LLAMA_COMMIT;
 extern char const *LLAMA_COMPILER;
 extern char const *LLAMA_BUILD_TARGET;

+struct llama_control_vector_load_info;
+
+int32_t get_num_physical_cores();
+
 //
 // CLI argument parsing
 //
-int32_t get_num_physical_cores();

 struct gpt_params {
    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
@@ -51,7 +54,8 @@ struct gpt_params {
    int32_t n_threads_batch_draft = -1;
    int32_t n_predict             = -1;    // new tokens to predict
    int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
    int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
    int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
@@ -85,6 +89,7 @@ struct gpt_params {
    struct llama_sampling_params sparams;

    std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_url         = ""; // model url to download
    std::string model_draft       = "";                              // draft model for speculative decoding
    std::string model_alias       = "unknown"; // model alias
    std::string prompt            = "";
@@ -102,6 +107,11 @@ struct gpt_params {
    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
    std::string lora_base  = "";                              // base model path for the lora adapter

+    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
+
+    int32_t control_vector_layer_start = -1; // layer range for control vector
+    int32_t control_vector_layer_end   = -1; // layer range for control vector
+
    int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
    int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                    //                                       (which is more convenient to use for plotting)
@@ -182,6 +192,9 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);

+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model,
+                                                         struct llama_model_params     params);
+
 // Batch utils

 void llama_batch_clear(struct llama_batch & batch);
@@ -267,3 +280,25 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40

 void llama_embd_normalize(const float * inp, float * out, int n);

+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
+
+//
+// Control vector utils
+//
+
+struct llama_control_vector_data {
+    int n_embd;
+
+    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
+    std::vector<float> data;
+};
+
+struct llama_control_vector_load_info {
+    float strength;
+
+    std::string fname;
+};
+
+// Load control vectors, scale each by strength, and add them together.
+// On error, returns {-1, empty}
+llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
@@ -17,6 +17,13 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
            return nullptr;
        }

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

        result->grammar = llama_grammar_init(
@@ -32,13 +32,13 @@ typedef struct llama_sampling_params {
    float       dynatemp_range        = 0.00f;    // 0.0 = disabled
    float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
+    float       penalty_repeat        = 1.00f;    // 1.0 = disabled
    float       penalty_freq          = 0.00f;    // 0.0 = disabled
    float       penalty_present       = 0.00f;    // 0.0 = disabled
    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
    float       mirostat_tau          = 5.00f;    // target entropy
    float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = true;     // consider newlines as a repeatable token
+    bool        penalize_nl           = false;     // consider newlines as a repeatable token

    std::vector<llama_sampler_type> samplers_sequence = {
        llama_sampler_type::TOP_K,
@@ -1634,7 +1634,7 @@ in chat mode so that the conversation can end normally.")
                self.post_write_tensors(tensor_map, name, data_torch)


-@Model.register("BertModel")
+@Model.register("BertModel", "CamembertModel")
 class BertModel(Model):
    model_arch = gguf.MODEL_ARCH.BERT

@@ -1965,6 +1965,23 @@ class MambaModel(Model):
            self.gguf_writer.add_tensor(new_name, data)


+@Model.register("CohereForCausalLM")
+class CommandR2Model(Model):
+    model_arch = gguf.MODEL_ARCH.COMMAND_R
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # max_position_embeddings = 8192 in config.json but model was actually
+        # trained on 128k context length
+        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+
 ###### CONVERSION LOGIC ######


@@ -332,6 +332,9 @@ class Params:
 #

 class BpeVocab:
+    tokenizer_model = "gpt2"
+    name = "bpe"
+
    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
        if isinstance(self.bpe_tokenizer.get('model'), dict):
@@ -390,6 +393,9 @@ class BpeVocab:


 class SentencePieceVocab:
+    tokenizer_model = "llama"
+    name = "spm"
+
    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
        added_tokens: dict[str, int]
@@ -453,6 +459,9 @@ class SentencePieceVocab:


 class HfVocab:
+    tokenizer_model = "llama"
+    name = "hfft"
+
    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
        try:
            from transformers import AutoTokenizer
@@ -553,7 +562,15 @@ class HfVocab:
        return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"


-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
+class NoVocab:
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"


 #
@@ -935,8 +952,10 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
    # Handle special case where the model's vocab size is not set
    if params.n_vocab == -1:
        raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
+            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
        )
+    if isinstance(vocab, NoVocab):
+        return  # model has no vocab

    # Check for a vocab size mismatch
    if params.n_vocab == vocab.vocab_size:
@@ -977,6 +996,7 @@ class OutputFile:
            name = str(params.path_model.parent).split('/')[-1]

        self.gguf.add_name                (name)
+        self.gguf.add_vocab_size          (params.n_vocab)
        self.gguf.add_context_length      (params.n_ctx)
        self.gguf.add_embedding_length    (params.n_embd)
        self.gguf.add_block_count         (params.n_layer)
@@ -1013,21 +1033,9 @@ class OutputFile:
        if params.ftype is not None:
            self.gguf.add_file_type(params.ftype)

-    def handle_tokenizer_model(self, vocab: Vocab) -> str:
-        # Map the vocab types to the supported tokenizer models
-        tokenizer_model = {
-            SentencePieceVocab: "llama",
-            HfVocab: "llama",
-            BpeVocab: "gpt2",
-        }.get(type(vocab))
-
-        # Block if vocab type is not predefined
-        if tokenizer_model is None:
-            raise ValueError("Unknown vocab type: Not supported")
-
-        return tokenizer_model
-
    def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+        assert not isinstance(vocab, NoVocab)
+
        tokens = []
        scores = []
        toktypes = []
@@ -1043,11 +1051,8 @@ class OutputFile:
        return tokens, scores, toktypes

    def add_meta_vocab(self, vocab: Vocab) -> None:
-        # Handle the tokenizer model
-        tokenizer_model = self.handle_tokenizer_model(vocab)
-
        # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(tokenizer_model)
+        self.gguf.add_tokenizer_model(vocab.tokenizer_model)

        # Extract model vocabulary for model conversion
        tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
@@ -1074,6 +1079,26 @@ class OutputFile:
    def write_tensor_info(self) -> None:
        self.gguf.write_ti_data_to_file()

+    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(
+                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            self.gguf.write_tensor_data(ndarray)
+
    def close(self) -> None:
        self.gguf.close()

@@ -1082,7 +1107,7 @@ class OutputFile:
        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
    ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+        check_vocab_size(params, vocab, pad_vocab=pad_vocab)

        of = OutputFile(fname_out, endianess=endianess)

@@ -1120,8 +1145,11 @@ class OutputFile:

        # meta data
        of.add_meta_arch(params)
-        of.add_meta_vocab(vocab)
-        of.add_meta_special_vocab(svocab)
+        if isinstance(vocab, NoVocab):
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
+        else:
+            of.add_meta_vocab(vocab)
+            of.add_meta_special_vocab(svocab)

        # tensor info
        for name, lazy_tensor in model.items():
@@ -1131,24 +1159,7 @@ class OutputFile:
        of.write_tensor_info()

        # tensor data
-        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
-        if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
-        else:
-            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
-
-        start = time.time()
-        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            elapsed = time.time() - start
-            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
-            padi = len(str(len(model)))
-            print(
-                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
-            of.gguf.write_tensor_data(ndarray)
+        of.write_tensor_data(ftype, model, concurrency)

        of.close()

@@ -1156,9 +1167,9 @@ class OutputFile:
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type

-    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
+    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
        return GGMLFileType.AllF32
-    if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
+    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
        return GGMLFileType.MostlyF16
    if output_type_str == "q8_0":
        return GGMLFileType.MostlyQ8_0
@@ -1309,8 +1320,8 @@ class VocabFactory:
                return vtype, path
        raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")

-    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocabtype == "bpe"
+    def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
+        load_merges = vocab.name == "bpe"
        n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
        return gguf.SpecialVocab(
            model_parent_path,
@@ -1319,30 +1330,34 @@ class VocabFactory:
            n_vocab=n_vocab,
        )

-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
        vocab_type, path = self._select_file(vocab_types)
        print(f"Loading vocab file {path!r}, type {vocab_type!r}")

        added_tokens_path = path.parent / "added_tokens.json"
-        vocab: Vocab
        if vocab_type == "bpe":
-            vocab = BpeVocab(
+            return BpeVocab(
                path, added_tokens_path if added_tokens_path.exists() else None
            )
-        elif vocab_type == "spm":
-            vocab = SentencePieceVocab(
+        if vocab_type == "spm":
+            return SentencePieceVocab(
                path, added_tokens_path if added_tokens_path.exists() else None
            )
-        elif vocab_type == "hfft":
-            vocab = HfVocab(
+        if vocab_type == "hfft":
+            return HfVocab(
                path.parent, added_tokens_path if added_tokens_path.exists() else None
            )
+        raise ValueError(vocab_type)
+
+    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        vocab: Vocab
+        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+            vocab = NoVocab()
        else:
-            raise ValueError(vocab_type)
+            vocab = self._create_vocab_by_path(vocab_types)
        # FIXME: Respect --vocab-dir?
        special_vocab = self._create_special_vocab(
            vocab,
-            vocab_type,
            model_parent_path,
        )
        return vocab, special_vocab
@@ -1380,6 +1395,7 @@ def main(args_in: list[str] | None = None) -> None:
    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
+    parser.add_argument("--no-vocab",     action="store_true",    help="store model without the vocab")
    parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
    parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
    parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
@@ -1392,6 +1408,10 @@ def main(args_in: list[str] | None = None) -> None:
    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")

    args = parser.parse_args(args_in)
+    if args.no_vocab:
+        if args.vocab_only:
+            raise ValueError("no need to specify --vocab-only if using --no-vocab")
+        args.vocab_type = "no_vocab"

    if args.dump_single:
        model_plus = lazy_load_file(args.model)
@@ -1442,7 +1462,7 @@ def main(args_in: list[str] | None = None) -> None:
        print(f"Wrote {outfile}")
        return

-    if model_plus.vocab is not None and args.vocab_dir is None:
+    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
        vocab = model_plus.vocab

    print(f"Vocab info: {vocab}")
@@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;

    // ensure enough sequences are available
-    ctx_params.n_parallel = *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());

    llama_context * ctx = llama_new_context_with_model(model, ctx_params);

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

        return true;
@@ -80,7 +80,7 @@ int main(int argc, char ** argv) {
    ctx_params.seed  = 1234;
    ctx_params.n_ctx = n_kv_req;
    ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_parallel      = n_parallel;
+    ctx_params.n_seq_max       = n_parallel;
    ctx_params.n_threads       = params.n_threads;
    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;

@@ -107,18 +107,25 @@ int main(int argc, char ** argv) {

    // max batch size
    const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch == params.n_ctx);
+    GGML_ASSERT(params.n_batch >= params.n_ctx);

    // tokenize the prompts and trim
    std::vector<std::vector<int32_t>> inputs;
    for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true);
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
        if (inp.size() > n_batch) {
            inp.resize(n_batch);
        }
        inputs.push_back(inp);
    }

+    // add eos if not present
+    for (auto & inp : inputs) {
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+    }
+
    // tokenization stats
    if (params.verbose_prompt) {
        for (int i = 0; i < (int) inputs.size(); i++) {
@@ -167,15 +174,26 @@ int main(int argc, char ** argv) {
    float * out = emb + p * n_embd;
    batch_decode(ctx, batch, out, s, n_embd);

-    // print first 3 embeddings
-    for (int j = 0; j < std::min(3, n_prompts); j++) {
-        fprintf(stderr, "embedding %d: ", j);
-        for (int i = 0; i < n_embd; i++) {
-            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+    // print the first part of the embeddings
+    fprintf(stdout, "\n");
+    for (int j = 0; j < n_prompts; j++) {
+        fprintf(stdout, "embedding %d: ", j);
+        for (int i = 0; i < std::min(16, n_embd); i++) {
+            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
        }
-        fprintf(stderr, "\n\n");
+        fprintf(stdout, "\n");
+    }
+
+    // print cosine similarity matrix
+    fprintf(stdout, "\n");
+    printf("cosine similarity matrix:\n\n");
+    for (int i = 0; i < n_prompts; i++) {
+        for (int j = 0; j < n_prompts; j++) {
+            float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+            fprintf(stdout, "%6.2f ", sim);
+        }
+        fprintf(stdout, "\n");
    }
-    fprintf(stderr, "\n");

    // clean up
    llama_print_timings(ctx);
@@ -211,6 +211,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
                for (int j = 0; j < ggml_nelements(cur); ++j) {
                    if (data[j] != 100 + i) {
                        fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
+                        gguf_free(ctx);
                        return false;
                    }
                }
@@ -0,0 +1,62 @@
+## Generative Representational Instruction Tuning (GRIT) Example
+[gritlm] a model which can generate embeddings as well as "normal" text
+generation depending on the instructions in the prompt.
+
+* Paper: https://arxiv.org/pdf/2402.09906.pdf
+
+### Retrieval-Augmented Generation (RAG) use case
+One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
+that we take documents that we want to use as context, to ground the large
+language model (LLM), and we create token embeddings for them. We then store
+these token embeddings in a vector database.
+
+When we perform a query, prompt the LLM, we will first create token embeddings
+for the query and then search the vector database to retrieve the most
+similar vectors, and return those documents so they can be passed to the LLM as
+context. Then the query and the context will be passed to the LLM which will
+have to _again_ create token embeddings for the query. But because gritlm is used
+the first query can be cached and the second query tokenization generation does
+not have to be performed at all.
+
+### Running the example
+Download a Grit model:
+```console
+$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
+```
+
+Run the example using the downloaded model:
+```console
+$ ./gritlm -m gritlm-7b_q4_1.gguf
+
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
+Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
+Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
+Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+When shadows lurk and ghosts do roam,
+And darkness reigns, a fearsome sight.
+
+Thou didst set out, with heart aglow,
+To conquer this mountain, so high,
+And reach the summit, where the stars do glow,
+And the moon shines bright, up in the sky.
+
+Through the mist and fog, thou didst press on,
+With steadfast courage, and a steadfast will,
+Through the darkness, thou didst not be gone,
+But didst climb on, with a steadfast skill.
+
+At last, thou didst reach the summit's crest,
+And gazed upon the world below,
+And saw the beauty of the night's best,
+And felt the peace, that only nature knows.
+
+Oh, brave adventurer, who dared to climb
+The lofty peak of Mt. Fuji in the night,
+Thou art a hero, in the eyes of all,
+For thou didst conquer this mountain, so bright.
+```
+
+[gritlm]: https://github.com/ContextualAI/gritlm
@@ -6,22 +6,6 @@

 // #define GRIT_DEBUG

-static float dot_product(const std::vector<float> & v1, const std::vector<float> & v2) {
-    float dot = 0.0f;
-    for (uint64_t i = 0; i < v1.size(); ++i) {
-        dot += v1[i] * v2[i];
-    }
-    return dot;
-}
-
-static float norm(const std::vector<float> & v) {
-    return std::sqrt(dot_product(v, v));
-}
-
-static float cosine_similarity(const std::vector<float> & v1, const std::vector<float> & v2) {
-    return dot_product(v1, v2) / (norm(v1) * norm(v2));
-}
-
 static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
    std::vector<std::vector<float>> result;

@@ -203,10 +187,12 @@ int main(int argc, char * argv[]) {
        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));

-        const float cosine_sim_q0_d0 = cosine_similarity(q_rep[0], d_rep[0]);
-        const float cosine_sim_q0_d1 = cosine_similarity(q_rep[0], d_rep[1]);
-        const float cosine_sim_q1_d0 = cosine_similarity(q_rep[1], d_rep[0]);
-        const float cosine_sim_q1_d1 = cosine_similarity(q_rep[1], d_rep[1]);
+        const int n_embd = llama_n_embd(mdl);
+
+        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);

        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
@@ -56,13 +56,31 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
    const struct ggml_tensor * src0 = t->src[0];
    const struct ggml_tensor * src1 = t->src[1];

+    std::string wname;
+    {
+        // remove any prefix and suffixes from the name
+        // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+        const char * p = strchr(src0->name, '#');
+        if (p != NULL) {
+            p = p + 1;
+            const char * q = strchr(p, '#');
+            if (q != NULL) {
+                wname = std::string(p, q - p);
+            } else {
+                wname = p;
+            }
+        } else {
+            wname = src0->name;
+        }
+    }
+
    // when ask is true, the scheduler wants to know if we are interested in data from this tensor
    // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
    if (ask) {
        if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
        if (t->op != GGML_OP_MUL_MAT) return false;
        if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
+        if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
        return true;
    }

@@ -94,12 +112,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
        // this is necessary to guarantee equal number of "ncall" for each tensor
        for (int ex = 0; ex < n_as; ++ex) {
            src0 = t->src[2 + ex];
-            auto& e = m_stats[src0->name];
+            auto& e = m_stats[wname];
            if (e.values.empty()) {
                e.values.resize(src1->ne[0], 0);
            }
            else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
                exit(1); //GGML_ASSERT(false);
            }
            // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
@@ -107,7 +125,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
            //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
            ++e.ncall;
            if (m_params.verbosity > 1) {
-                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
            }
            for (int row = 0; row < (int)src1->ne[1]; ++row) {
                const int excur = m_ids[row*n_as + idx];
@@ -129,17 +147,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
            }
        }
    } else {
-        auto& e = m_stats[src0->name];
+        auto& e = m_stats[wname];
        if (e.values.empty()) {
            e.values.resize(src1->ne[0], 0);
        }
        else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
            exit(1); //GGML_ASSERT(false);
        }
        ++e.ncall;
        if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
        }
        for (int row = 0; row < (int)src1->ne[1]; ++row) {
            const float * x = data + row * src1->ne[0];
@@ -8,6 +8,7 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
@@ -103,6 +104,7 @@ static std::string get_cpu_info() {
                }
            }
        }
+        fclose(f);
    }
 #endif
    // TODO: other platforms
@@ -112,10 +114,10 @@ static std::string get_cpu_info() {
 static std::string get_gpu_info() {
    std::string id;
 #ifdef GGML_USE_CUBLAS
-    int count = ggml_cuda_get_device_count();
+    int count = ggml_backend_cuda_get_device_count();
    for (int i = 0; i < count; i++) {
        char buf[128];
-        ggml_cuda_get_device_description(i, buf, sizeof(buf));
+        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
        id += buf;
        if (i < count - 1) {
            id += "/";
@@ -164,6 +166,7 @@ struct cmd_params {
    std::vector<int> n_prompt;
    std::vector<int> n_gen;
    std::vector<int> n_batch;
+    std::vector<int> n_ubatch;
    std::vector<ggml_type> type_k;
    std::vector<ggml_type> type_v;
    std::vector<int> n_threads;
@@ -183,7 +186,8 @@ static const cmd_params cmd_params_defaults = {
    /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
    /* n_prompt      */ {512},
    /* n_gen         */ {128},
-    /* n_batch       */ {512},
+    /* n_batch       */ {2048},
+    /* n_ubatch      */ {512},
    /* type_k        */ {GGML_TYPE_F16},
    /* type_v        */ {GGML_TYPE_F16},
    /* n_threads     */ {get_num_physical_cores()},
@@ -208,6 +212,7 @@ static void print_usage(int /* argc */, char ** argv) {
    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
@@ -217,7 +222,7 @@ static void print_usage(int /* argc */, char ** argv) {
    printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
    printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
-    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
@@ -297,6 +302,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
            }
            auto p = split<int>(argv[i], split_delim);
            params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<int>(argv[i], split_delim);
+            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
        } else if (arg == "-ctk" || arg == "--cache-type-k") {
            if (++i >= argc) {
                invalid_param = true;
@@ -455,6 +467,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
    if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
    if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
    if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
+    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
    if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
@@ -474,6 +487,7 @@ struct cmd_params_instance {
    int n_prompt;
    int n_gen;
    int n_batch;
+    int n_ubatch;
    ggml_type type_k;
    ggml_type type_v;
    int n_threads;
@@ -511,6 +525,7 @@ struct cmd_params_instance {

        cparams.n_ctx = n_prompt + n_gen;
        cparams.n_batch = n_batch;
+        cparams.n_ubatch = n_ubatch;
        cparams.type_k = type_k;
        cparams.type_v = type_v;
        cparams.offload_kqv = !no_kv_offload;
@@ -532,6 +547,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
    for (const auto & mmp : params.use_mmap)
    for (const auto & embd : params.embeddings)
    for (const auto & nb : params.n_batch)
+    for (const auto & nub : params.n_ubatch)
    for (const auto & tk : params.type_k)
    for (const auto & tv : params.type_v)
    for (const auto & nkvo : params.no_kv_offload)
@@ -545,6 +561,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                /* .n_prompt     = */ n_prompt,
                /* .n_gen        = */ 0,
                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
                /* .type_k       = */ tk,
                /* .type_v       = */ tv,
                /* .n_threads    = */ nt,
@@ -568,6 +585,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                /* .n_prompt     = */ 0,
                /* .n_gen        = */ n_gen,
                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
                /* .type_k       = */ tk,
                /* .type_v       = */ tv,
                /* .n_threads    = */ nt,
@@ -604,6 +622,7 @@ struct test {
    uint64_t model_size;
    uint64_t model_n_params;
    int n_batch;
+    int n_ubatch;
    int n_threads;
    ggml_type type_k;
    ggml_type type_v;
@@ -627,6 +646,7 @@ struct test {
        model_size = llama_model_size(lmodel);
        model_n_params = llama_model_n_params(lmodel);
        n_batch = inst.n_batch;
+        n_ubatch = inst.n_ubatch;
        n_threads = inst.n_threads;
        type_k = inst.type_k;
        type_v = inst.type_v;
@@ -705,7 +725,8 @@ struct test {
            "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
            "cpu_info", "gpu_info",
            "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "type_k", "type_v",
+            "n_batch", "n_ubatch",
+            "n_threads", "type_k", "type_v",
            "n_gpu_layers", "split_mode",
            "main_gpu", "no_kv_offload",
            "tensor_split", "use_mmap", "embeddings",
@@ -719,7 +740,8 @@ struct test {
    enum field_type {STRING, BOOL, INT, FLOAT};

    static field_type get_field_type(const std::string & field) {
-        if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
+        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
+            field == "n_threads" ||
            field == "model_size" || field == "model_n_params" ||
            field == "n_gpu_layers" || field == "main_gpu" ||
            field == "n_prompt" || field == "n_gen" ||
@@ -759,7 +781,8 @@ struct test {
            std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
            cpu_info, gpu_info,
            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_batch), std::to_string(n_ubatch),
+            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
            std::to_string(n_gpu_layers), split_mode_str(split_mode),
            std::to_string(main_gpu), std::to_string(no_kv_offload),
            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
@@ -957,6 +980,9 @@ struct markdown_printer : public printer {
        if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
            fields.emplace_back("n_batch");
        }
+        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
+            fields.emplace_back("n_ubatch");
+        }
        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
            fields.emplace_back("type_k");
        }
@@ -1096,25 +1122,40 @@ struct sql_printer : public printer {
 };

 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
-    int n_processed = 0;
-
    llama_set_n_threads(ctx, n_threads, n_threads);

+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    std::vector<llama_token> tokens(n_batch);
+
+    int n_processed = 0;
+
    while (n_processed < n_prompt) {
        int n_tokens = std::min(n_prompt - n_processed, n_batch);
+        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        for (int i = 1; i < n_tokens; i++) {
+            tokens[i] = std::rand() % n_vocab;
+        }
        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
        n_processed += n_tokens;
    }
+
+    llama_synchronize(ctx);
 }

 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
-    llama_token token = llama_token_bos(llama_get_model(ctx));
-
    llama_set_n_threads(ctx, n_threads, n_threads);

+    const llama_model * model = llama_get_model(ctx);
+    const int32_t n_vocab = llama_n_vocab(model);
+
+    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+
    for (int i = 0; i < n_gen; i++) {
        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_synchronize(ctx);
+        token = std::rand() % n_vocab;
    }
 }

@@ -1203,7 +1244,8 @@ int main(int argc, char ** argv) {

        // warmup run
        if (t.n_prompt > 0) {
-            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
+            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
        }
        if (t.n_gen > 0) {
            test_gen(ctx, 1, 0, t.n_threads);
@@ -1219,6 +1261,7 @@ int main(int argc, char ** argv) {
            if (t.n_gen > 0) {
                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
            }
+
            uint64_t t_ns = get_time_ns() - t_start;
            t.samples_ns.push_back(t_ns);
        }
@@ -221,6 +221,7 @@ actor LlamaContext {
            if llama_decode(context, batch) != 0 {
                print("llama_decode() failed during prompt")
            }
+            llama_synchronize(context)

            let t_pp_end = ggml_time_us()

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

            let t_tg_end = ggml_time_us()
@@ -63,12 +63,20 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director
 ```console
 git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 ```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+
+2) Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
 python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+
+4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
 ```console
 mkdir vit
 cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
@@ -76,18 +84,18 @@ cp ../llava-v1.6-vicuna-7b/llava.projector vit/
 curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
 ```

-4) Create the visual gguf model:
+5) Create the visual gguf model:
 ```console
 python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP

-5) Then convert the model to gguf format:
+6) Then convert the model to gguf format:
 ```console
 python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
 ```

-6) And finally we can run the llava-cli using the 1.6 model version:
+7) And finally we can run the llava-cli using the 1.6 model version:
 ```console
 ./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
 ```
@@ -497,7 +497,6 @@ struct clip_ctx {

    // memory buffers to evaluate the model
    ggml_backend_buffer_t params_buffer  = NULL;
-    ggml_backend_buffer_t compute_buffer = NULL;

    ggml_backend_t backend       = NULL;
    ggml_gallocr_t compute_alloc = NULL;
@@ -995,6 +994,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
        if (!new_clip->ctx_data) {
            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
            clip_free(new_clip);
+            gguf_free(ctx);
            return nullptr;
        }

@@ -1002,6 +1002,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
        if (!fin) {
            printf("cannot open model file for loading tensors\n");
            clip_free(new_clip);
+            gguf_free(ctx);
            return nullptr;
        }

@@ -1023,6 +1024,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            if (!fin) {
                printf("%s: failed to seek for tensor %s\n", __func__, name);
                clip_free(new_clip);
+                gguf_free(ctx);
                return nullptr;
            }
            int num_bytes = ggml_nbytes(cur);
@@ -1232,16 +1234,16 @@ struct clip_image_f32 * clip_image_f32_init() {

 void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
 void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
-void clip_image_u8_batch_free(struct clip_image_u8_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
    }
 }
-void clip_image_f32_batch_free(struct clip_image_f32_batch  & batch) {
-    if (batch.size > 0) {
-        delete[] batch.data;
-        batch.size = 0;
+void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
+    if (batch->size > 0) {
+        delete[] batch->data;
+        batch->size = 0;
    }
 }

@@ -1494,7 +1496,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im

 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
-bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
    bool pad_to_square = true;
    if (!ctx->has_vision_encoder) {
        printf("This gguf file seems to have no vision encoder\n");
@@ -1506,11 +1508,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
        pad_to_square = false;
    }
    // free the previous res_imgs if any set
-    if (res_imgs.size > 0) {
+    if (res_imgs->size > 0) {
        clip_image_f32_batch_free(res_imgs);
    }
-    res_imgs.data = nullptr;
-    res_imgs.size = 0;
+    res_imgs->data = nullptr;
+    res_imgs->size = 0;

    // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
    // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
@@ -1565,11 +1567,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
            bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
            patches.insert(patches.begin(), image_original_resize);
            // clip_image_f32_batch_init(patches.size());
-            res_imgs.size = patches.size();
-            res_imgs.data = new clip_image_f32[res_imgs.size];
+            res_imgs->size = patches.size();
+            res_imgs->data = new clip_image_f32[res_imgs->size];
            int num=0;
            for (auto& patch : patches) {
-                normalize_image_u8_to_f32(patch, &res_imgs.data[num], ctx->image_mean, ctx->image_std);
+                normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std);
                num++;
            }

@@ -1657,9 +1659,9 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
    // }
    // res_imgs.push_back(res);

-    res_imgs.size = 1;
-    res_imgs.data = new clip_image_f32[res_imgs.size];
-    res_imgs.data[0] = *res;
+    res_imgs->size = 1;
+    res_imgs->data = new clip_image_f32[res_imgs->size];
+    res_imgs->data[0] = *res;
    clip_image_f32_free(res);

    return true;
@@ -1673,6 +1675,9 @@ void clip_free(clip_ctx * ctx) {
    ggml_free(ctx->ctx_data);
    gguf_free(ctx->ctx_gguf);

+    ggml_backend_buffer_free(ctx->params_buffer);
+    ggml_backend_free(ctx->backend);
+    ggml_gallocr_free(ctx->compute_alloc);
    delete ctx;
 }

@@ -1908,6 +1913,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                break;
            default:
                printf("Please use an input file in f32 or f16\n");
+                gguf_free(ctx_out);
                return false;
            }

@@ -60,8 +60,8 @@ CLIP_API struct clip_image_f32 * clip_image_f32_init();

 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
-CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  & batch);
-CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);

 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);

@@ -69,7 +69,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);

 /** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );

 CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);

@@ -223,7 +223,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
    clip_image_f32_batch img_res_v;
    img_res_v.size = 0;
    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
        fprintf(stderr, "%s: unable to preprocess image\n", __func__);
        delete[] img_res_v.data;
        return false;
@@ -29,9 +29,9 @@ struct llava_image_embed {
 };

 /** sanity check for clip <-> llava embed size match */
-LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
+LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);

-LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);

 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
@@ -67,6 +67,7 @@ main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt
 In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:

 -   `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
+-   `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
 -   `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
 -   `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
 -   `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
@@ -878,6 +878,7 @@ int main(int argc, char ** argv) {
                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
+
                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());

                    embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
@@ -589,9 +589,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
            }
        }

-        const auto t_end = std::chrono::high_resolution_clock::now();

        if (i == 0) {
+            llama_synchronize(ctx);
+            const auto t_end = std::chrono::high_resolution_clock::now();
            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
            int total_seconds = (int)(t_total*n_chunk/n_seq);
@@ -841,7 +842,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
    const int n_batch = params.n_batch;

    const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);

@@ -1118,7 +1119,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
    const int n_batch = params.n_batch;

    const int max_tasks_per_batch = 128;
-    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);

@@ -1470,7 +1471,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
    const int n_batch = params.n_batch;

    const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));

    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);

@@ -20,6 +20,7 @@ The project is under active development, and we are [looking for feedback and co
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
 - `--threads-http N`: number of threads in the http server pool to process requests (default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`)
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
+- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf).
 - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 - `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
 - `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
@@ -147,7 +147,7 @@ struct server_slot {
    int32_t n_decoded   = 0;
    int32_t n_remaining = -1;
    int32_t i_batch     = -1;
-    int32_t n_predict   = -1;
+    int32_t n_predict   = -1; // TODO: disambiguate from params.n_predict

    int32_t n_prompt_tokens           = 0;
    int32_t n_prompt_tokens_processed = 0;
@@ -739,7 +739,13 @@ struct server_context {
        default_generation_settings_for_props = get_formated_generation(slots.front());
        default_generation_settings_for_props["seed"] = -1;

-        batch = llama_batch_init(n_ctx, 0, params.n_parallel);
+        // the update_slots() logic will always submit a maximum of n_batch tokens
+        // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
+        {
+            const int32_t n_batch = llama_n_batch(ctx);
+
+            batch = llama_batch_init(n_batch, 0, params.n_parallel);
+        }

        metrics.init();
    }
@@ -1036,8 +1042,10 @@ struct server_context {
                llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
            }

-            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch) {
-                const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
+            const int32_t n_batch = llama_n_batch(ctx);
+
+            for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
+                const int32_t n_tokens = std::min(params.n_batch, batch.n_tokens - i);
                llama_batch batch_view = {
                    n_tokens,
                    batch.token    + i,
@@ -1226,7 +1234,7 @@ struct server_context {
            {"mirostat_eta",              slot.sparams.mirostat_eta},
            {"penalize_nl",               slot.sparams.penalize_nl},
            {"stop",                      slot.params.antiprompt},
-            {"n_predict",                 slot.params.n_predict},
+            {"n_predict",                 slot.params.n_predict}, // TODO: fix duplicate key n_predict
            {"n_keep",                    params.n_keep},
            {"ignore_eos",                ignore_eos},
            {"stream",                    slot.params.stream},
@@ -1738,7 +1746,8 @@ struct server_context {
        }

        // process in chunks of params.n_batch
-        int32_t n_batch = params.n_batch;
+        int32_t n_batch = llama_n_batch(ctx);
+        int32_t n_ubatch = llama_n_ubatch(ctx);

        // next, batch any pending prompts without exceeding n_batch
        if (params.cont_batching || batch.n_tokens == 0) {
@@ -1811,7 +1820,7 @@ struct server_context {

                        if (slot.embedding) {
                            // this prompt is too large to process - discard it
-                            if (slot.n_prompt_tokens > n_batch) {
+                            if (slot.n_prompt_tokens > n_ubatch) {
                                slot.state = SLOT_STATE_PROCESSING;
                                slot.command = SLOT_COMMAND_NONE;
                                slot.release();
@@ -2157,7 +2166,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
    printf("  --pooling {none,mean,cls} pooling type for embeddings, use model default if unspecified\n");
    printf("  -dt N, --defrag-thold N\n");
    printf("                            KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
-    printf("  -b N, --batch-size N      batch size for prompt processing (default: %d)\n", params.n_batch);
+    printf("  -b N, --batch-size N      logical maximum batch size (default: %d)\n", params.n_batch);
+    printf("  -ub N, --ubatch-size N    physical maximum batch size (default: %d)\n", params.n_ubatch);
    printf("  --memory-f32              use f32 instead of f16 for memory key+value (default: disabled)\n");
    printf("                            not recommended: doubles context memory required and no measurable increase in quality\n");
    if (llama_supports_mlock()) {
@@ -2185,6 +2195,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
    }
    printf("  -m FNAME, --model FNAME\n");
    printf("                            model path (default: %s)\n", params.model.c_str());
+    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
+    printf("                            model download url (default: %s)\n", params.model_url.c_str());
    printf("  -a ALIAS, --alias ALIAS\n");
    printf("                            set an alias for the model, will be added as `model` field in completion response\n");
    printf("  --lora FNAME              apply LoRA adapter (implies --no-mmap)\n");
@@ -2307,6 +2319,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                break;
            }
            params.model = argv[i];
+        } else if (arg == "-mu" || arg == "--model-url") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.model_url = argv[i];
        } else if (arg == "-a" || arg == "--alias") {
            if (++i >= argc) {
                invalid_param = true;
@@ -2424,6 +2442,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                break;
            }
            params.n_batch = std::stoi(argv[i]);
+        } else if (arg == "-ub" || arg == "--ubatch-size") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_ubatch = std::stoi(argv[i]);
        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
            if (++i >= argc) {
                invalid_param = true;
@@ -2763,6 +2787,7 @@ int main(int argc, char ** argv) {
        res.set_header("Access-Control-Allow-Credentials", "true");
        res.set_header("Access-Control-Allow-Methods",     "POST");
        res.set_header("Access-Control-Allow-Headers",     "*");
+        return res.set_content("", "application/json; charset=utf-8");
    });

    svr->set_logger(log_server_request);
@@ -3371,44 +3396,37 @@ int main(int argc, char ** argv) {
        const json body = json::parse(req.body);
        bool is_openai = false;

-        // an input prompt can string or a list of tokens (integer)
-        std::vector<json> prompts;
+        // an input prompt can be a string or a list of tokens (integer)
+        json prompt;
        if (body.count("input") != 0) {
            is_openai = true;
-            if (body["input"].is_array()) {
-                // support multiple prompts
-                for (const json & elem : body["input"]) {
-                    prompts.push_back(elem);
-                }
-            } else {
-                // single input prompt
-                prompts.push_back(body["input"]);
-            }
+            prompt = body["input"];
        } else if (body.count("content") != 0) {
-            // only support single prompt here
-            std::string content = body["content"];
-            prompts.push_back(content);
+            // with "content", we only support single prompt
+            prompt = std::vector<std::string>{body["content"]};
        } else {
            res_error(res, format_error_response("\"input\" or \"content\" must be provided", ERROR_TYPE_INVALID_REQUEST));
            return;
        }

-        // process all prompts
-        json responses = json::array();
-        for (auto & prompt : prompts) {
-            // TODO @ngxson : maybe support multitask for this endpoint?
-            // create and queue the task
+        // create and queue the task
+        json responses;
+        {
            const int id_task = ctx_server.queue_tasks.get_new_id();
-
            ctx_server.queue_results.add_waiting_task_id(id_task);
-            ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0}}, false, true);
+            ctx_server.request_completion(id_task, -1, {{"prompt", prompt}}, false, true);

            // get the result
            server_task_result result = ctx_server.queue_results.recv(id_task);
            ctx_server.queue_results.remove_waiting_task_id(id_task);
            if (!result.error) {
-                // append to the responses
-                responses.push_back(result.data);
+                if (result.data.count("results")) {
+                    // result for multi-task
+                    responses = result.data["results"];
+                } else {
+                    // result for single task
+                    responses = std::vector<json>{result.data};
+                }
            } else {
                // error received, ignore everything else
                res_error(res, result.data);
@@ -3417,24 +3435,19 @@ int main(int argc, char ** argv) {
        }

        // write JSON response
-        json root;
-        if (is_openai) {
-            json res_oai = json::array();
-            int i = 0;
-            for (auto & elem : responses) {
-                res_oai.push_back(json{
-                    {"embedding", json_value(elem, "embedding", json::array())},
-                    {"index",     i++},
-                    {"object",    "embedding"}
-                });
-            }
-            root = format_embeddings_response_oaicompat(body, res_oai);
-        } else {
-            root = responses[0];
-        }
+        json root = is_openai
+            ? format_embeddings_response_oaicompat(body, responses)
+            : responses[0];
        return res.set_content(root.dump(), "application/json; charset=utf-8");
    };

+    auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
+        return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
+            res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
+            return false;
+        };
+    };
+
    //
    // Router
    //
@@ -3446,17 +3459,6 @@ int main(int argc, char ** argv) {
    }

    // using embedded static files
-    auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
-        return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
-            res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
-            return false;
-        };
-    };
-
-    svr->Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) {
-        // TODO @ngxson : I have no idea what it is... maybe this is redundant?
-        return res.set_content("", "application/json; charset=utf-8");
-    });
    svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
    svr->Get("/index.js", handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
    svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
@@ -57,7 +57,7 @@ Feature or Scenario must be annotated with `@llama.cpp` to be included in the de
 To run a scenario annotated with `@bug`, start:

 ```shell
-DEBUG=ON ./tests.sh --no-skipped --tags bug
+DEBUG=ON ./tests.sh --no-skipped --tags bug --stop
 ```

 After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.
@@ -4,11 +4,13 @@ Feature: llama.cpp server

  Background: Server startup
    Given a server listening on localhost:8080
-    And   a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
+    And   a model url https://huggingface.co/ggml-org/models/resolve/main/bert-bge-small/ggml-model-f16.gguf
+    And   a model file ggml-model-f16.gguf
    And   a model alias bert-bge-small
    And   42 as server seed
    And   2 slots
    And   1024 as batch size
+    And   1024 as ubatch size
    And   2048 KV cache size
    And   embeddings extraction
    Then  the server is starting
@@ -1,10 +1,12 @@
-import errno
 import os
-import socket
-import subprocess
-import time
-from contextlib import closing
 import signal
+import socket
+import sys
+import time
+import traceback
+from contextlib import closing
+
+import psutil


 def before_scenario(context, scenario):
@@ -20,33 +22,40 @@ def before_scenario(context, scenario):


 def after_scenario(context, scenario):
-    if context.server_process is None:
-        return
-    if scenario.status == "failed":
-        if 'GITHUB_ACTIONS' in os.environ:
-            print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
-            if os.path.isfile('llama.log'):
-                with closing(open('llama.log', 'r')) as f:
-                    for line in f:
-                        print(line)
-        if not is_server_listening(context.server_fqdn, context.server_port):
-            print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
+    try:
+        if 'server_process' not in context or context.server_process is None:
+            return
+        if scenario.status == "failed":
+            if 'GITHUB_ACTIONS' in os.environ:
+                print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
+                if os.path.isfile('llama.log'):
+                    with closing(open('llama.log', 'r')) as f:
+                        for line in f:
+                            print(line)
+            if not is_server_listening(context.server_fqdn, context.server_port):
+                print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")

-    if not pid_exists(context.server_process.pid):
-        assert False, f"Server not running pid={context.server_process.pid} ..."
+        if not pid_exists(context.server_process.pid):
+            assert False, f"Server not running pid={context.server_process.pid} ..."

-    server_graceful_shutdown(context)
+        server_graceful_shutdown(context)

-    # Wait few for socket to free up
-    time.sleep(0.05)
+        # Wait few for socket to free up
+        time.sleep(0.05)

-    attempts = 0
-    while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
-        server_kill(context)
-        time.sleep(0.1)
-        attempts += 1
-        if attempts > 5:
-            server_kill_hard(context)
+        attempts = 0
+        while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
+            server_kill(context)
+            time.sleep(0.1)
+            attempts += 1
+            if attempts > 5:
+                server_kill_hard(context)
+    except:
+        exc = sys.exception()
+        print("error in after scenario: \n")
+        print(exc)
+        print("*** print_tb: \n")
+        traceback.print_tb(exc.__traceback__, file=sys.stdout)


 def server_graceful_shutdown(context):
@@ -67,11 +76,11 @@ def server_kill_hard(context):
    path = context.server_path

    print(f"Server dangling exits, hard killing force {pid}={path}...\n")
-    if os.name == 'nt':
-        process = subprocess.check_output(['taskkill', '/F', '/pid', str(pid)]).decode()
-        print(process)
-    else:
-        os.kill(-pid, signal.SIGKILL)
+    try:
+        psutil.Process(pid).kill()
+    except psutil.NoSuchProcess:
+        return False
+    return True


 def is_server_listening(server_fqdn, server_port):
@@ -84,17 +93,9 @@ def is_server_listening(server_fqdn, server_port):


 def pid_exists(pid):
-    """Check whether pid exists in the current process table."""
-    if pid < 0:
+    try:
+        psutil.Process(pid)
+    except psutil.NoSuchProcess:
        return False
-    if os.name == 'nt':
-        output = subprocess.check_output(['TASKLIST', '/FI', f'pid eq {pid}']).decode()
-        print(output)
-        return "No tasks are running" not in output
-    else:
-        try:
-            os.kill(pid, 0)
-        except OSError as e:
-            return e.errno == errno.EPERM
-        else:
-            return True
+    return True
+
@@ -4,7 +4,8 @@ Feature: llama.cpp server

  Background: Server startup
    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   a model url https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf
+    And   a model file stories260K.gguf
    And   a model alias tinyllama-2
    And   42 as server seed
      # KV Cache corresponds to the total amount of tokens
@@ -5,6 +5,8 @@ import os
 import re
 import socket
 import subprocess
+import sys
+import threading
 import time
 from contextlib import closing
 from re import RegexFlag
@@ -32,7 +34,10 @@ def step_server_config(context, server_fqdn, server_port):
    context.base_url = f'http://{context.server_fqdn}:{context.server_port}'

    context.model_alias = None
+    context.model_file = None
+    context.model_url = None
    context.n_batch = None
+    context.n_ubatch = None
    context.n_ctx = None
    context.n_ga = None
    context.n_ga_w = None
@@ -64,6 +69,16 @@ def step_download_hf_model(context, hf_file, hf_repo):
        print(f"model file: {context.model_file}\n")


+@step('a model file {model_file}')
+def step_model_file(context, model_file):
+    context.model_file = model_file
+
+
+@step('a model url {model_url}')
+def step_model_url(context, model_url):
+    context.model_url = model_url
+
+
@step('a model alias {model_alias}')
 def step_model_alias(context, model_alias):
    context.model_alias = model_alias
@@ -118,6 +133,10 @@ def step_server_metrics(context):
 def step_start_server(context):
    start_server_background(context)
    attempts = 0
+    max_attempts = 20
+    if 'GITHUB_ACTIONS' in os.environ:
+        max_attempts *= 2
+
    while True:
        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
            result = sock.connect_ex((context.server_fqdn, context.server_port))
@@ -125,7 +144,7 @@ def step_start_server(context):
                print("\x1b[33;46mserver started!\x1b[0m")
                return
            attempts += 1
-            if attempts > 20:
+            if attempts > max_attempts:
                assert False, "server not started"
            print(f"waiting for server to start, connect error code = {result}...")
            time.sleep(0.1)
@@ -136,7 +155,8 @@ def step_start_server(context):
 async def step_wait_for_the_server_to_be_started(context, expecting_status):
    match expecting_status:
        case 'healthy':
-            await wait_for_health_status(context, context.base_url, 200, 'ok')
+            await wait_for_health_status(context, context.base_url, 200, 'ok',
+                                         timeout=30)

        case 'ready' | 'idle':
            await wait_for_health_status(context, context.base_url, 200, 'ok',
@@ -278,6 +298,11 @@ def step_n_batch(context, n_batch):
    context.n_batch = n_batch


+@step('{n_ubatch:d} as ubatch size')
+def step_n_ubatch(context, n_ubatch):
+    context.n_ubatch = n_ubatch
+
+
@step('{seed:d} as seed')
 def step_seed(context, seed):
    context.seed = seed
@@ -937,6 +962,9 @@ async def wait_for_health_status(context,
        print(f"Starting checking for health for expected_health_status={expected_health_status}\n")
    interval = 0.5
    counter = 0
+    if 'GITHUB_ACTIONS' in os.environ:
+        timeout *= 2
+
    async with aiohttp.ClientSession() as session:
        while True:
            async with await session.get(f'{base_url}/health', params=params) as health_response:
@@ -1025,10 +1053,15 @@ def start_server_background(context):
    server_args = [
        '--host', server_listen_addr,
        '--port', context.server_port,
-        '--model', context.model_file
    ]
+    if context.model_file:
+        server_args.extend(['--model', context.model_file])
+    if context.model_url:
+        server_args.extend(['--model-url', context.model_url])
    if context.n_batch:
        server_args.extend(['--batch-size', context.n_batch])
+    if context.n_ubatch:
+        server_args.extend(['--ubatch-size', context.n_ubatch])
    if context.n_gpu_layer:
        server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
    if context.server_continuous_batching:
@@ -1064,8 +1097,23 @@ def start_server_background(context):

    pkwargs = {
        'creationflags': flags,
+        'stdout': subprocess.PIPE,
+        'stderr': subprocess.PIPE
    }
    context.server_process = subprocess.Popen(
        [str(arg) for arg in [context.server_path, *server_args]],
        **pkwargs)
+
+    def log_stdout(process):
+        for line in iter(process.stdout.readline, b''):
+            print(line.decode('utf-8'), end='')
+    thread_stdout = threading.Thread(target=log_stdout, args=(context.server_process,))
+    thread_stdout.start()
+
+    def log_stderr(process):
+        for line in iter(process.stderr.readline, b''):
+            print(line.decode('utf-8'), end='', file=sys.stderr)
+    thread_stderr = threading.Thread(target=log_stderr, args=(context.server_process,))
+    thread_stderr.start()
+
    print(f"server pid={context.server_process.pid}, behave pid={os.getpid()}")
@@ -3,4 +3,5 @@ behave~=1.2.6
 huggingface_hub~=0.20.3
 numpy~=1.24.4
 openai~=0.25.0
+psutil~=5.9.8
 prometheus-client~=0.20.0
@@ -529,6 +529,16 @@ static std::vector<json> format_partial_response_oaicompat(json result, const st
 }

 static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
+    json data = json::array();
+    int i = 0;
+    for (auto & elem : embeddings) {
+        data.push_back(json{
+            {"embedding", json_value(elem, "embedding", json::array())},
+            {"index",     i++},
+            {"object",    "embedding"}
+        });
+    }
+
    json res = json {
        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
        {"object", "list"},
@@ -536,7 +546,7 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
            {"prompt_tokens", 0},
            {"total_tokens", 0}
        }},
-        {"data", embeddings}
+        {"data", data}
    };

    return res;
@@ -13,8 +13,11 @@ source /opt/intel/oneapi/setvars.sh
 #for FP32
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx

-#build example/main only
+#build example/main
 #cmake --build . --config Release --target main

+#build example/llama-bench
+#cmake --build . --config Release --target llama-bench
+
 #build all binary
 cmake --build . --config Release -v
@@ -9,18 +9,28 @@ source /opt/intel/oneapi/setvars.sh

 if [ $# -gt 0 ]; then
    GGML_SYCL_DEVICE=$1
+    GGML_SYCL_SINGLE_GPU=1
 else
    GGML_SYCL_DEVICE=0
 fi
-echo "use $GGML_SYCL_DEVICE as main GPU"
+
 #export GGML_SYCL_DEBUG=1


 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.

-#use all GPUs with same max compute units
-ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
+    echo "use $GGML_SYCL_DEVICE as main GPU"
+    #use signle GPU only
+    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+else
+    #use multiple GPUs with same max compute units
+    ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+fi

 #use main GPU only
 #ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none

+#use multiple GPUs with same max compute units
+#ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+
@@ -711,6 +711,7 @@ static bool load_checkpoint_file(const char * filename, struct my_llama_model *

    load_checkpoint_gguf(fctx, f_ggml_ctx, model, train);

+    gguf_free(fctx);
    return true;
 }

@@ -20,11 +20,11 @@
    },
    "nixpkgs": {
      "locked": {
-        "lastModified": 1709703039,
-        "narHash": "sha256-6hqgQ8OK6gsMu1VtcGKBxKQInRLHtzulDo9Z5jxHEFY=",
+        "lastModified": 1710451336,
+        "narHash": "sha256-pP86Pcfu3BrAvRO7R64x7hs+GaQrjFes+mEPowCfkxY=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "9df3e30ce24fd28c7b3e2de0d986769db5d6225d",
+        "rev": "d691274a972b3165335d261cc4671335f5c67de9",
        "type": "github"
      },
      "original": {
@@ -61,7 +61,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
    }
 }

-// TODO: GGML_PAD ?
 static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
    assert(alignment && !(alignment & (alignment - 1))); // power of 2
    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
@@ -69,25 +68,14 @@ static size_t aligned_offset(const void * buffer, size_t offset, size_t alignmen
 }

 // tallocr
-struct ggml_tallocr {
-    ggml_backend_buffer_t buffer;
-    void * base;
-    size_t alignment;
-    size_t offset;
-};
-
-ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
-    ggml_tallocr_t talloc = malloc(sizeof(struct ggml_tallocr));
-    if (talloc == NULL) {
-        return NULL;
-    }

+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
    void * base = ggml_backend_buffer_get_base(buffer);
    size_t align = ggml_backend_buffer_get_alignment(buffer);

    assert(align && !(align & (align - 1))); // power of 2

-    *talloc = (struct ggml_tallocr) {
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
        /*.buffer    = */ buffer,
        /*.base      = */ base,
        /*.alignment = */ align,
@@ -96,11 +84,7 @@ ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
    return talloc;
 }

-void ggml_tallocr_free(ggml_tallocr_t talloc) {
-    free(talloc);
-}
-
-void ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor) {
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
    size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
    size = GGML_PAD(size, talloc->alignment);

@@ -354,12 +338,16 @@ struct hash_node {
    bool allocated;
 };

-//
 struct tensor_alloc {
    size_t offset;
    size_t size_max; // 0 = pre-allocated, unused, or view
 };

+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
 struct node_alloc {
    int buffer_id;
    struct tensor_alloc dst;
@@ -378,7 +366,7 @@ struct ggml_gallocr {
    struct node_alloc * node_allocs; // [n_nodes]
    int n_nodes;

-    struct tensor_alloc * leaf_allocs; // [n_leafs]
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
    int n_leafs;
 };

@@ -543,17 +531,28 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
    return node_buffer_ids ? node_buffer_ids[i] : 0;
 }

-static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
    // clear hash tables
    memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
    memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));

+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
    // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
+    // allocate other graph inputs and leafs first to avoid overwriting them
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];

-        if (ggml_is_view(node)) {
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
            struct ggml_tensor * view_src = node->view_src;
            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
        }
@@ -570,26 +569,13 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr

            ggml_gallocr_hash_get(galloc, src)->n_children += 1;

-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
            }
        }
    }

-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
-
    // allocate tensors
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
@@ -652,7 +638,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
    }
 }

-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
    size_t hash_size = graph->visited_hash_table.size;

    // initialize hash table
@@ -676,7 +662,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
    }

    // allocate in hash table
-    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids);
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);

    // set the node_allocs from the hash table
    if (galloc->n_nodes < graph->n_nodes) {
@@ -711,15 +697,16 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
    }
    if (galloc->n_leafs < graph->n_leafs) {
        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
+        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
        GGML_ASSERT(galloc->leaf_allocs != NULL);
    }
    galloc->n_leafs = graph->n_leafs;
    for (int i = 0; i < graph->n_leafs; i++) {
        struct ggml_tensor * leaf = graph->leafs[i];
        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        galloc->leaf_allocs[i].leaf.offset = hn->offset;
+        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
    }

    // reallocate buffers if needed
@@ -727,7 +714,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);

-        if (new_size > cur_size) {
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
@@ -744,30 +732,30 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 }

 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-    return ggml_gallocr_reserve_n(galloc, graph, NULL);
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }

-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);

-    if (node->view_src != NULL) {
-        if (node->buffer == NULL) {
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
            assert(tensor_alloc->offset == SIZE_MAX);
-            if (node->view_src->buffer == NULL) {
+            if (tensor->view_src->buffer == NULL) {
                // this tensor was allocated without ggml-backend
                return;
            }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
        }
    } else {
-        if (node->data == NULL) {
+        if (tensor->data == NULL) {
            assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
            void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
        } else {
-            if (node->buffer == NULL) {
+            if (tensor->buffer == NULL) {
                // this tensor was allocated without ggml-backend
                return;
            }
@@ -843,13 +831,18 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)

    // reset buffers
    for (int i = 0; i < galloc->n_buffers; i++) {
-        // zero size buffers are not allocated
        if (galloc->buffers[i] != NULL) {
            ggml_backend_buffer_reset(galloc->buffers[i]);
        }
    }

    // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
+    }
    // nodes
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
@@ -863,12 +856,6 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
        }
        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
    }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
-    }

    return true;
 }
@@ -900,12 +887,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
        return false;
    }

-    struct ggml_tallocr * tallocr = ggml_tallocr_new(buffer);
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);

    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
        if (t->data == NULL) {
            if (t->view_src == NULL) {
-                ggml_tallocr_alloc(tallocr, t);
+                ggml_tallocr_alloc(&tallocr, t);
            } else if (t->buffer == NULL) {
                ggml_backend_view_init(buffer, t);
            }
@@ -917,8 +904,6 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
        }
    }

-    ggml_tallocr_free(tallocr);
-
    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
    (*buffers)[(*n_buffers)++] = buffer;

@@ -11,11 +11,15 @@ typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
 typedef struct ggml_backend * ggml_backend_t;

 // Tensor allocator
-typedef struct ggml_tallocr * ggml_tallocr_t;
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};

-GGML_API ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer);
-GGML_API void           ggml_tallocr_free(ggml_tallocr_t talloc);
-GGML_API void           ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor);
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);

 // Graph allocator
 /*
@@ -50,7 +54,11 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-GGML_API bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);

 // automatic reallocation if the topology changes when using a single buffer
 // returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
@@ -86,12 +86,12 @@ extern "C" {
        // (optional) asynchronous tensor data access
        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);

        // (optional) complete all pending operations
        void (*GGML_CALL synchronize)(ggml_backend_t backend);

-        // create a plan for ggml_cgraph and free it
+        // compute graph with a plan (not used currently)
        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);

@@ -102,16 +102,32 @@ extern "C" {

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

    struct ggml_backend {
        ggml_guid_t guid;

        struct ggml_backend_i iface;
-
        ggml_backend_context_t context;
    };

+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
+
    //
    // Backend registry
    //
@@ -9,6 +9,7 @@ extern "C" {

    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
    typedef struct ggml_backend * ggml_backend_t;
    typedef void * ggml_backend_graph_plan_t;

@@ -69,14 +70,27 @@ extern "C" {
    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);

-    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // events
+    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
+    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event

    //
    // CPU backend
@@ -123,27 +137,31 @@ extern "C" {
    /*
      Example usage:

-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
-        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be asigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);

-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);

-        // in build_graph:
-        build_graph(...) {
-            // manually assign nodes to a backend (optional, should not be needed in most cases)
-            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
-        }
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);

-        // allocate backend buffers from measure graph
-        ggml_backend_sched_init_measure(sched, measure_graph);
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);

-        // the scheduler is now ready to compute graphs
+        ggml_backend_sched_reserve(sched, reserve_graph);

        // compute
        graph = build_graph(sched);
        ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
    */

    struct ggml_backend_sched;
@@ -158,20 +176,26 @@ extern "C" {
    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);

    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+
    // Initialize backend buffers from a measure graph
    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
    // Get the number of splits of the last graph
    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);

    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);

-    GGML_API void                 ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);

    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);

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

 #if defined(GGML_COMMON_IMPL_C)
 #include <stdint.h>
@@ -14,7 +418,7 @@
 #define GGML_TABLE_END() };

 #define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA)
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
 #include <cstdint>

 #define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
@@ -645,6 +1049,7 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
 GGML_TABLE_END()

 #define NGRID_IQ1S 2048
+#define IQ1S_DELTA 0.125f
 #if defined(GGML_COMMON_IMPL_C)
 GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
@@ -1422,3 +1827,4 @@ GGML_TABLE_END()
 #endif

 #endif // GGML_COMMON_IMPL
+#endif // GGML_COMMON_IMPL
@@ -17,29 +17,17 @@ extern "C" {

 #define GGML_CUDA_MAX_DEVICES       16

-// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
-GGML_API GGML_CALL void   ggml_init_cublas(void);
-
-// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
-GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
-
-GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
-GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);
-
-GGML_API GGML_CALL bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API GGML_CALL bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-GGML_API GGML_CALL int    ggml_cuda_get_device_count(void);
-GGML_API GGML_CALL void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
-
 // backend API
 GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);

 GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);

+// device buffer
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

@@ -47,6 +35,9 @@ GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
 GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);

+GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
 #ifdef  __cplusplus
 }
 #endif
@@ -1951,6 +1951,12 @@ static struct ggml_backend_i kompute_backend_i = {
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
    /* .supports_op             = */ ggml_backend_kompute_supports_op,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_kompute_guid() {
@@ -280,6 +280,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
    id<MTLLibrary> metal_library;

    // load library
+    //
+    // - first check if the library is embedded
+    // - then check if the library is in the bundle
+    // - if not found, load the source and compile it
+    // - if that fails, return NULL
    {
        NSBundle * bundle = nil;
 #ifdef SWIFT_PACKAGE
@@ -287,12 +292,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #else
        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
+
        NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (libPath != nil) {
+
+#if GGML_METAL_EMBED_LIBRARY
+        const bool try_metallib = false;
+#else
+        const bool try_metallib = true;
+#endif
+
+        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (try_metallib && path_lib != nil) {
            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:libPath];
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
+            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
            metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
            if (error) {
                GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
@@ -305,38 +319,41 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
            extern const char ggml_metallib_start[];
            extern const char ggml_metallib_end[];

-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
+            NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
 #else
            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);

-            NSString * sourcePath;
-            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+            NSString * path_source;
+            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];

-            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil");
+            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");

-            if (ggmlMetalPathResources) {
-                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
+            if (path_resource) {
+                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
            } else {
-                sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
            }
-            if (sourcePath == nil) {
+
+            if (path_source == nil) {
                GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                sourcePath = @"ggml-metal.metal";
+                path_source = @"ggml-metal.metal";
            }
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
-            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
+
+            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+            NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
            if (error) {
                GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                return NULL;
            }
-#endif
+#endif // GGML_METAL_EMBED_LIBRARY

            @autoreleasepool {
                // dictionary of preprocessor macros
                NSMutableDictionary * prep = [NSMutableDictionary dictionary];

 #ifdef GGML_QKK_64
-                prep[@"QK_K"] = @(64);
+                prep[@"GGML_QKK_64"] = @(1);
 #endif

                MTLCompileOptions* options = [MTLCompileOptions new];
@@ -2820,6 +2837,12 @@ static struct ggml_backend_i ggml_backend_metal_i = {
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_metal_graph_compute,
    /* .supports_op             = */ ggml_backend_metal_supports_op,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
@@ -1,49 +1,15 @@
-#include <metal_stdlib>
-
+#define GGML_COMMON_DECL_METAL
 #define GGML_COMMON_IMPL_METAL
 #include "ggml-common.h"

+#include <metal_stdlib>
+
 using namespace metal;

 #define MAX(x, y) ((x) > (y) ? (x) : (y))
 #define MIN(x, y) ((x) < (y) ? (x) : (y))
 #define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }

-#define QK4_0 32
-#define QR4_0 2
-typedef struct {
-    half    d;             // delta
-    uint8_t qs[QK4_0 / 2]; // nibbles / quants
-} block_q4_0;
-
-#define QK4_1 32
-typedef struct {
-    half d;                 // delta
-    half m;                 // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-
-#define QK5_0 32
-typedef struct {
-    half d;                // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-
-#define QK5_1 32
-typedef struct {
-    half d;                 // delta
-    half m;                 // min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-
-#define QK8_0 32
-typedef struct {
-    half    d;         // delta
-    int8_t  qs[QK8_0]; // quants
-} block_q8_0;
-
 #define N_SIMDWIDTH 32 // assuming SIMD group size is 32

 enum ggml_sort_order {
@@ -2478,147 +2444,6 @@ kernel void kernel_concat(
    }
 }

-//============================================ k-quants ======================================================
-
-#ifndef QK_K
-#define QK_K 256
-#else
-static_assert(QK_K == 256 || QK_K == 64, "QK_K must be 256 or 64");
-#endif
-
-#if QK_K == 256
-#define K_SCALE_SIZE 12
-#else
-#define K_SCALE_SIZE 4
-#endif
-
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    half d;           // super-block scale for quantized scales
-    half dmin;        // super-block scale for quantized mins
-} block_q2_K;
-// 84 bytes / block
-
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#if QK_K == 64
-    uint8_t scales[2];
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    half d;             // super-block scale
-} block_q3_K;
-
-#if QK_K == 64
-typedef struct {
-    half    d[2];          // super-block scales/mins
-    uint8_t scales[2];
-    uint8_t qs[QK_K/2];    // 4-bit quants
-} block_q4_K;
-#else
-typedef struct {
-    half d;             // super-block scale for quantized scales
-    half dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-#endif
-
-#if QK_K == 64
-typedef struct {
-    half  d;                     // super-block scales/mins
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-#else
-typedef struct {
-    half d;                      // super-block scale for quantized scales
-    half dmin;                   // super-block scale for quantized mins
-    uint8_t scales[3*QK_K/64];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-// 176 bytes / block
-#endif
-
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    half d;                  // super-block scale
-} block_q6_K;
-// 210 bytes / block
-
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-// 66 bytes / block for QK_K = 256, so 2.0625 bpw
-
-typedef struct {
-    half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-// 74 bytes / block for QK_K = 256, so 2.3125 bpw
-
-// 2.5625 bpw quants
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-
-typedef struct {
-    half d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-// 98 bytes / block for QK_K = 256, so 3.0625 bpw
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-
-typedef struct {
-    half d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    half    d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
-typedef struct {
-    half d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-#endif
-
-//====================================== dot products =========================
-
 void kernel_mul_mv_q2_K_f32_impl(
        device const  void * src0,
        device const float * src1,
@@ -4377,7 +4202,7 @@ void kernel_mul_mv_iq1_s_f32_impl(
                     + yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
                     + yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
            }
-            sumf[row] += (float)dh[0] * (sum - sumy) * (2*(qh[0] >> 12) + 1);
+            sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);

            dh += nb*sizeof(block_iq1_s)/2;
            qs += nb*sizeof(block_iq1_s);
@@ -5076,14 +4901,16 @@ void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 &
    const float d = xb->d;
    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
    device const uint16_t * qh = xb->qh;
-    const float dl = d * (2*(qh[ib32] >> 12) + 1);
-    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | (((qh[ib32] >> (6*il+0)) & 7) << 8)));
-    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | (((qh[ib32] >> (6*il+3)) & 7) << 8)));
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
    for (int i = 0; i < 4; ++i) {
-        reg[0][i] = dl * (grid1[i] & 0xf) - dl;
-        reg[1][i] = dl * (grid1[i] >>  4) - dl;
-        reg[2][i] = dl * (grid2[i] & 0xf) - dl;
-        reg[3][i] = dl * (grid2[i] >>  4) - dl;
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
    }
 }

@@ -1,3 +1,6 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
 #include "ggml-quants.h"
 #include "ggml-impl.h"

@@ -951,7 +954,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);

        int sum = 0;

@@ -966,7 +969,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
            sum += y[i].qs[QK8_1/2 + j];
        }

-        y[i].s = sum*d;
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
    }
 }

@@ -994,7 +997,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);

        int32x4_t accv = vdupq_n_s32(0);

@@ -1010,7 +1013,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
            accv = vaddq_s32(accv, vi);
        }

-        y[i].s = d * vaddvq_s32(accv);
+        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
    }
 #elif defined(__wasm_simd128__)
    for (int i = 0; i < nb; i++) {
@@ -1033,7 +1036,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
        const float d = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);

        v128_t accv = wasm_i32x4_splat(0);

@@ -1049,10 +1052,11 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
            accv = wasm_i32x4_add(accv, vi);
        }

-        y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
-                      wasm_i32x4_extract_lane(accv, 1) +
-                      wasm_i32x4_extract_lane(accv, 2) +
-                      wasm_i32x4_extract_lane(accv, 3));
+        y[i].s = GGML_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
    }
 #elif defined(__AVX2__) || defined(__AVX__)
    for (int i = 0; i < nb; i++) {
@@ -1077,7 +1081,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {

        // Quantize these floats
        const float d = maxScalar / 127.f;
-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);
        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
        const __m256 mul = _mm256_set1_ps( id );

@@ -1101,7 +1105,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {

 #if defined(__AVX2__)
        // Compute the sum of the quants and set y[i].s
-        y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));

        // Convert int32 to int16
        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -1131,7 +1135,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
        // Compute the sum of the quants and set y[i].s
        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));

        // Convert int32 to int16
        ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -1162,7 +1166,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
        const float d  = amax / ((1 << 7) - 1);
        const float id = d ? 1.0f/d : 0.0f;

-        y[i].d = d;
+        y[i].d = GGML_FP32_TO_FP16(d);

        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);

@@ -1179,7 +1183,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {

        // set y[i].s
        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = sum*d;
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
    }
 #else
    GGML_UNUSED(nb);
@@ -3456,11 +3460,12 @@ void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, in
        const uint16_t * qh = x[i].qh;

        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const float dl = d * (2*(qh[ib] >> 12) + 1);
+            const float dl = d * (2*((qh[ib] >> 12) & 7) + 1);
+            const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA;
            for (int l = 0; l < 4; ++l) {
                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
                for (int j = 0; j < 8; ++j) {
-                    y[j] = dl * grid[j];
+                    y[j] = dl * (grid[j] + delta);
                }
                y += 8;
            }
@@ -4018,10 +4023,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
            const block_q8_1 * restrict b_y0 = &vy0[i];
            const block_q8_1 * restrict b_y1 = &vy1[i];

-            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
-                                   GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
-                                   GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
-                                   GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
+            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                   GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                                   GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
            summs0 += summs_t;

            const uint8x16_t m4b = vdupq_n_u8(0x0F);
@@ -4086,7 +4091,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
        const block_q8_1 * restrict y0 = &y[i + 0];
        const block_q8_1 * restrict y1 = &y[i + 1];

-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

@@ -4109,8 +4114,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);

-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@@ -4123,9 +4128,9 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
    // Main loop
    for (int i = 0; i < nb; ++i) {
        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-        const float d1 = y[i].d;
+        const float d1 = GGML_FP16_TO_FP32(y[i].d);

-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);

        const __m256 d0v = _mm256_set1_ps( d0 );
        const __m256 d1v = _mm256_set1_ps( d1 );
@@ -4177,7 +4182,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r

        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);

-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
    }

    *s = sumf;
@@ -4195,7 +4200,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
        }

-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
    }

    *s = sumf;
@@ -4531,8 +4536,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r

        const uint8x16_t m4b = vdupq_n_u8(0x0F);

-        summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s;
-        summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s;
+        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);

        // extract the 5th bit via lookup table ((b) << 4)
        memcpy(&qh0, x0->qh, sizeof(qh0));
@@ -4576,10 +4581,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r

        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
    }

    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@@ -4596,7 +4601,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
        const block_q5_1 * restrict x0 = &x[i];
        const block_q8_1 * restrict y0 = &y[i];

-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s;
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);

        const v128_t m4b = wasm_i8x16_splat(0x0F);

@@ -4643,7 +4648,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
    }

    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -4658,14 +4663,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
    for (int i = 0; i < nb; i++) {
        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));

-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);

        __m256i qx = bytes_from_nibbles_32(x[i].qs);
        __m256i bxhi = bytes_from_bits_32(x[i].qh);
        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
        qx = _mm256_or_si256(qx, bxhi);

-        const __m256 dy = _mm256_set1_ps(y[i].d);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);

        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -4685,7 +4690,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
    for (int i = 0; i < nb; i++) {
        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));

-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
+        summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);

        __m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
        const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@@ -4699,7 +4704,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
        bxh = _mm_or_si128(bxh, bxhih);
        bx_0 = MM256_SET_M128I(bxh, bxl);

-        const __m256 dy = _mm256_set1_ps(y[i].d);
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);

        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@@ -4766,7 +4771,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r

        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);

-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
    }

    *s = sumf;
@@ -4790,7 +4795,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
            sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
        }

-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
+        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
    }

    *s = sumf;
@@ -9024,7 +9029,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
            qs += 8;

-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
            vs.val[0] = vceqq_u8(vs.val[0], mask2);
@@ -9033,7 +9038,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);

-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
            vs.val[0] = vceqq_u8(vs.val[0], mask2);
@@ -9104,12 +9109,12 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void *
                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
            qs += 8;

-            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);

-            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
@@ -9385,7 +9390,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);


-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
@@ -9394,7 +9399,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));

-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
@@ -9582,7 +9587,7 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
        const uint8_t  * qs = x[i].qs;
        const uint16_t * qh = x[i].qh;

-        int sumi1 = 0, sumi2 = 0;
+        int sumi1 = 0, sumi2 = 0, sumi3 = 0;

        for (int ib = 0; ib < QK_K/32; ib += 2) {

@@ -9601,12 +9606,16 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);

-            sumi1 += vaddvq_s32(p1) * (2*(qh[ib+0] >> 12) + 1);
-            sumi2 += vaddvq_s32(p2) * (2*(qh[ib+1] >> 12) + 1);
+            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            sumi1 += vaddvq_s32(p1) * ls1;
+            sumi2 += vaddvq_s32(p2) * ls2;
+            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);

        }

-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2);
+        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
    }

    *s = sumf;
@@ -9614,6 +9623,7 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
 #elif defined __AVX2__

    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
    for (int i = 0; i < nb; ++i) {

        const int8_t   * q8 = y[i].qs;
@@ -9621,6 +9631,7 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
        const uint16_t * qh = x[i].qh;

        __m256i sumi = _mm256_setzero_si256();
+        int sumi1 = 0;
        for (int ib = 0; ib < QK_K/32; ib += 2) {
            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
                                                    iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
@@ -9632,17 +9643,23 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void

            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*(qh[ib+0] >> 12) + 1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*(qh[ib+1] >> 12) + 1));
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));

            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
        }

-        accum = _mm256_fmadd_ps(_mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)), _mm256_cvtepi32_ps(sumi), accum);
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
+        accum1 += d * sumi1;

    }

-    *s = hsum_float_8(accum);
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;

 #else

@@ -9653,9 +9670,10 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
        const uint8_t  * qs = x[i].qs;
        const uint16_t * qh = x[i].qh;

-        int sumi = 0;
+        int sumi = 0, sumi1 = 0;
        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*(qh[ib] >> 12) + 1;
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
            int lsum = 0;
            for (int l = 0; l < 4; ++l) {
                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
@@ -9664,11 +9682,12 @@ void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void
                }
                q8 += 8;
            }
-            sumi += ls * lsum;
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
            qs += 4;
        }

-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * sumi;
+        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
    }

    *s = sumf;
@@ -11438,7 +11457,7 @@ static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const u
 }

 static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
-        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L, int ngrid) {
+        const float * restrict xval, const float * restrict weight, float scale, const float * restrict xg, int8_t * restrict L, int ngrid) {
    int num_neighbors = neighbours[0];
    GGML_ASSERT(num_neighbors > 0);
    float best_score = FLT_MAX;
@@ -11447,7 +11466,7 @@ static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const
        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
        float d2 = 0;
        for (int i = 0; i < 8; ++i) {
-            float q = (pg[i] - 3)/2;
+            float q = xg[(pg[i] - 1)/2];
            float w = weight[i];
            float diff = scale*q - xval[i];
            d2 += w*diff*diff;
@@ -11463,7 +11482,7 @@ static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const
            float d2 = 0;
            for (int j = 0; j < 8; ++j) {
                float w = weight[j];
-                float q = (grid_i[j] - 3)/2;
+                float q = xg[(grid_i[j] - 1)/2];
                float diff = scale*q - xval[i];
                d2 += w*diff*diff;
            }
@@ -11480,7 +11499,7 @@ static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const
            const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
            float sumqx = 0, sumq2 = 0;
            for (int i = 0; i < 8; ++i) {
-                float q = (pg[i] - 3)/2;
+                float q = xg[(pg[i] - 1)/2];
                float w = weight[i];
                sumqx += w*q*xval[i];
                sumq2 += w*q*q;
@@ -11519,6 +11538,9 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy

    block_iq1_s * y = vy;

+    const float x_p[3] = {-1 + IQ1S_DELTA,  IQ1S_DELTA, 1 + IQ1S_DELTA};
+    const float x_m[3] = {-1 - IQ1S_DELTA, -IQ1S_DELTA, 1 - IQ1S_DELTA};
+
    float  scales[QK_K/IQ1S_BLOCK_SIZE];
    float  weight[IQ1S_BLOCK_SIZE];
    int8_t L[IQ1S_BLOCK_SIZE];
@@ -11527,6 +11549,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
    float  pairs[2*IQ1S_BLOCK_SIZE];
    int * idx = (int *)(pairs + 1);
    uint16_t index[IQ1S_BLOCK_SIZE/8];
+    int8_t shifts[QK_K/IQ1S_BLOCK_SIZE];

    for (int ibl = 0; ibl < nbl; ++ibl) {

@@ -11572,25 +11595,33 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                }
            }
            float best_score = 0, scale = max;
-            int besti1 = 0, besti2 = 0;
+            int besti1 = -1, besti2 = -1, best_shift = 0;
            for (int i1 = 0; i1 <= IQ1S_BLOCK_SIZE; ++i1) {
                for (int i2 = i1; i2 <= IQ1S_BLOCK_SIZE; ++i2) {
-                    float sumqx = -(sumx[i1] - sumx[0]) + (sumx[IQ1S_BLOCK_SIZE] - sumx[i2]);
-                    float sumq2 =  (sumw[i1] - sumw[0]) + (sumw[IQ1S_BLOCK_SIZE] - sumw[i2]);
+                    float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[IQ1S_BLOCK_SIZE] - sumx[i2])*x_p[2];
+                    float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[IQ1S_BLOCK_SIZE] - sumw[i2])*x_p[2]*x_p[2];
                    if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
                        scale = sumqx/sumq2; best_score = scale*sumqx;
-                        besti1 = i1; besti2 = i2;
+                        besti1 = i1; besti2 = i2; best_shift = 1;
+                    }
+                    sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[IQ1S_BLOCK_SIZE] - sumx[i2])*x_m[2];
+                    sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[IQ1S_BLOCK_SIZE] - sumw[i2])*x_m[2]*x_m[2];
+                    if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+                        scale = sumqx/sumq2; best_score = scale*sumqx;
+                        besti1 = i1; besti2 = i2; best_shift = -1;
                    }
                }
            }
+            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_shift != 0);
            for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
            for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
            for (int j = besti2; j < IQ1S_BLOCK_SIZE; ++j) L[idx[2*j]] = 2;
            if (scale < 0) {
                for (int j = 0; j < IQ1S_BLOCK_SIZE; ++j) L[j] = 2 - L[j];
-                scale = -scale;
+                scale = -scale; best_shift = -best_shift;
            }
            bool all_on_grid = true;
+            const float * xx = best_shift == 1 ? x_p : x_m;
            for (int k = 0; k < IQ1S_BLOCK_SIZE/8; ++k) {
                uint16_t u = 0;
                for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
@@ -11598,7 +11629,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                if (grid_index < 0) {
                    all_on_grid = false;
                    const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
-                    grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, L + 8*k, NGRID_IQ1S);
+                    grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
                    GGML_ASSERT(grid_index >= 0);
                }
                index[k] = grid_index;
@@ -11609,7 +11640,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
                    for (int j = 0; j < 8; ++j) {
                        float w = weight[8*k + j];
-                        float q = (pg[j] - 3)/2;
+                        float q = xx[(pg[j] - 1)/2];
                        sumqx += w*q*xb[8*k+j];
                        sumq2 += w*q*q;
                    }
@@ -11624,6 +11655,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
            y[ibl].qh[ib] = h;
            GGML_ASSERT(scale >= 0);
            scales[ib] = scale;
+            shifts[ib] = best_shift;
            max_scale = MAX(max_scale, scale);
        }

@@ -11632,12 +11664,13 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
            continue;
        }

-        float d = max_scale/31;
+        float d = max_scale/15;
        y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.085f is another fudge factor. Don't ask me why it is needed.
        float id = 1/d;
        for (int ib = 0; ib < QK_K/IQ1S_BLOCK_SIZE; ++ib) {
            int l = nearest_int(0.5f*(id*scales[ib]-1));
-            l = MAX(0, MIN(15, l));
+            l = MAX(0, MIN(7, l));
+            if (shifts[ib] == -1) l |= 8;
            y[ibl].qh[ib] |= (l << 12);
        }
    }
@@ -1,248 +1,12 @@
 #pragma once

+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
 // GGML internal header

-#include "ggml-impl.h"
-
-#include <stdint.h>
-#include <stddef.h>
-
-#define QK4_0 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    ggml_fp16_t m;          // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // min
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    int8_t  qs[QK8_0];     // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    float d;               // delta
-    float s;               // d * sum(qs[i])
-    int8_t  qs[QK8_1];     // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    ggml_fp16_t d;           // super-block scale for quantized scales
-    ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;             // super-block scale for quantized scales
-    ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;               // super-block scale for quantized scales
-    ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-// (Almost) "true" 2-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 2.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-// 2.3125 bpw quants
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-// 2.5625 bpw quants
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
-// (Almost) "true" 3-bit quantization.
-// Due to the need to use blocks as per ggml design, it ends up using
-// 3.0625 bpw because of the 16-bit scale for each block of 256.
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[3*QK_K/8];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-// 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t  qs[QK_K/8];
-    uint16_t qh[QK_K/32];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-//typedef struct block_iq4_nl block_iq4_xs;
-#else
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -16,6 +16,7 @@
 #include <cinttypes>
 #include <cstddef>
 #include <cstdint>
+#include <cstdlib>
 #include <float.h>
 #include <limits>
 #include <stdint.h>
@@ -24,10 +25,9 @@
 #include <cmath>
 #include <iostream>
 #include <fstream>
-
 #include <stdio.h>
 #include <stdlib.h>
-
+#include <regex>

 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
@@ -82,6 +82,30 @@ Following definition copied from DPCT head files, which are used by ggml-sycl.cp
 #define __dpct_noinline__ __attribute__((noinline))
 #endif

+
+std::string get_device_type_name(const sycl::device &Device) {
+    auto DeviceType = Device.get_info<sycl::info::device::device_type>();
+    switch (DeviceType) {
+    case sycl::info::device_type::cpu:
+        return "cpu";
+    case sycl::info::device_type::gpu:
+        return "gpu";
+    case sycl::info::device_type::host:
+        return "host";
+    case sycl::info::device_type::accelerator:
+        return "acc";
+    default:
+        return "unknown";
+    }
+}
+
+std::string get_device_backend_and_type(const sycl::device &device) {
+    std::stringstream device_type;
+    sycl::backend backend = device.get_backend();
+    device_type <<  backend << ":" << get_device_type_name(device);
+    return device_type.str();
+}
+
 namespace dpct
 {
    typedef sycl::queue *queue_ptr;
@@ -202,24 +226,29 @@ namespace dpct
            // Version string has the following format:
            // a. OpenCL<space><major.minor><space><vendor-specific-information>
            // b. <major.minor>
+            // c. <AmdGcnArchName> e.g gfx1030
            std::string ver;
            ver = dev.get_info<sycl::info::device::version>();
            std::string::size_type i = 0;
-            while (i < ver.size())
-            {
-                if (isdigit(ver[i]))
-                    break;
-                i++;
+            while (i < ver.size()) {
+              if (isdigit(ver[i]))
+                break;
+              i++;
            }
            major = std::stoi(&(ver[i]));
-            while (i < ver.size())
-            {
-                if (ver[i] == '.')
-                    break;
-                i++;
+            while (i < ver.size()) {
+              if (ver[i] == '.')
+                break;
+              i++;
+            }
+            if (i < ver.size()) {
+              // a. and b.
+              i++;
+              minor = std::stoi(&(ver[i]));
+            } else {
+              // c.
+              minor = 0;
            }
-            i++;
-            minor = std::stoi(&(ver[i]));
        }

        template <typename tag, typename T>
@@ -937,17 +966,65 @@ namespace dpct

    private:
        mutable std::recursive_mutex m_mutex;
+        static bool compare_dev(sycl::device &device1, sycl::device &device2)
+        {
+            dpct::device_info prop1;
+            dpct::get_device_info(prop1, device1);
+            dpct::device_info prop2;
+            dpct::get_device_info(prop2, device2);
+            return prop1.get_max_compute_units() > prop2.get_max_compute_units();
+        }
+        static int convert_backend_index(std::string & backend) {
+            if (backend == "ext_oneapi_level_zero:gpu") return 0;
+            if (backend == "opencl:gpu") return 1;
+            if (backend == "opencl:cpu") return 2;
+            if (backend == "opencl:acc") return 3;
+            printf("convert_backend_index: can't handle backend=%s\n", backend.c_str());
+            GGML_ASSERT(false);
+        }
+        static bool compare_backend(std::string &backend1, std::string &backend2) {
+            return convert_backend_index(backend1) < convert_backend_index(backend2);
+        }
        dev_mgr()
        {
            sycl::device default_device =
                sycl::device(sycl::default_selector_v);
            _devs.push_back(std::make_shared<device_ext>(default_device));

-            std::vector<sycl::device> sycl_all_devs =
-                sycl::device::get_devices(sycl::info::device_type::all);
+            std::vector<sycl::device> sycl_all_devs;
            // Collect other devices except for the default device.
            if (default_device.is_cpu())
                _cpu_device = 0;
+
+            auto Platforms = sycl::platform::get_platforms();
+            // Keep track of the number of devices per backend
+            std::map<sycl::backend, size_t> DeviceNums;
+            std::map<std::string, std::vector<sycl::device>> backend_devices;
+
+            while (!Platforms.empty()) {
+                auto Platform = Platforms.back();
+                Platforms.pop_back();
+                auto devices = Platform.get_devices();
+                std::string backend_type = get_device_backend_and_type(devices[0]);
+                for (const auto &device : devices) {
+                    backend_devices[backend_type].push_back(device);
+                }
+            }
+
+            std::vector<std::string> keys;
+            for(auto it = backend_devices.begin(); it != backend_devices.end(); ++it) {
+                keys.push_back(it->first);
+            }
+            std::sort(keys.begin(), keys.end(), compare_backend);
+
+            for (auto &key : keys) {
+                std::vector<sycl::device> devs = backend_devices[key];
+                std::sort(devs.begin(), devs.end(), compare_dev);
+                for (const auto &dev : devs) {
+                    sycl_all_devs.push_back(dev);
+                }
+            }
+
            for (auto &dev : sycl_all_devs)
            {
                if (dev == default_device)
@@ -3144,6 +3221,7 @@ namespace dpct

 } // COPY from DPCT head files

+#define GGML_COMMON_DECL_SYCL
 #define GGML_COMMON_IMPL_SYCL
 #include "ggml-common.h"

@@ -3196,6 +3274,11 @@ static int g_work_group_size = 0;
 #define GGML_SYCL_MMV_Y 1
 #endif

+enum ggml_sycl_backend_gpu_mode {
+    SYCL_UNSET_GPU_MODE = -1,
+    SYCL_SINGLE_GPU_MODE = 0,
+    SYCL_MUL_GPU_MODE
+};

 static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");

@@ -3312,66 +3395,6 @@ typedef void (*ggml_sycl_op_flatten_t)(const ggml_tensor *src0,
                                       const float *src1_dd, float *dst_dd,
                                       const dpct::queue_ptr &main_stream);

-// QK = number of values after dequantization
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QK4_0 32
-#define QR4_0 2
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-typedef struct dpct_type_block_q4_0 {
-    sycl::half d;           // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-#define QR4_1 2
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-typedef struct dpct_type_block_q4_1 {
-    sycl::half2 dm;         // dm.x = delta, dm.y = min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-#define QR5_0 2
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-typedef struct dpct_type_block_q5_0 {
-    sycl::half d;           // delta
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2];  // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-#define QR5_1 2
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-typedef struct dpct_type_block_q5_1 {
-    sycl::half2 dm;         // dm.x = delta, dm.y = min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-#define QR8_0 1
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-typedef struct dpct_type_block_q8_0 {
-    sycl::half d;           // delta
-    int8_t  qs[QK8_0];      // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-#define QR8_1 1
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-typedef struct dpct_type_block_q8_1 {
-    sycl::half2 ds;         // ds.x = delta, ds.y = sum
-    int8_t  qs[QK8_0];      // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
-
 typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_sycl_t)(int **x_ql, sycl::half2 **x_dm,
                                      int **x_qh, int **x_sc);
@@ -3388,137 +3411,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ms,
    const int &i, const int &j, const int &k);

-//================================= k-quants
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-#define QR2_K 4
-#define QI2_K (QK_K / (4*QR2_K))
-typedef struct dpct_type_block_q2_K {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    sycl::half2 dm;          // super-block scale for quantized scales/mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-#define QR3_K 4
-#define QI3_K (QK_K / (4*QR3_K))
-typedef struct dpct_type_block_q3_K {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#ifdef GGML_QKK_64
-    uint8_t scales[2]; // scales, quantized with 8 bits
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    sycl::half d; // super-block scale
-} block_q3_K;
-//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
-
-#define QR4_K 2
-#define QI4_K (QK_K / (4*QR4_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    sycl::half dm[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == sizeof(sycl::half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct dpct_type_block_q4_K {
-    sycl::half2 dm;            // super-block scale for quantized scales/mins
-    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-#define QR5_K 2
-#define QI5_K (QK_K / (4*QR5_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    sycl::half d;                  // super-block scale
-    int8_t scales[QK_K/16];  // block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct dpct_type_block_q5_K {
-    sycl::half2 dm;               // super-block scale for quantized scales/mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-#define QR6_K 2
-#define QI6_K (QK_K / (4*QR6_K))
-typedef struct dpct_type_block_q6_K {
-    uint8_t ql[QK_K/2];   // quants, lower 4 bits
-    uint8_t qh[QK_K/4];   // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales
-    sycl::half d;            // delta
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
-
-#define QR2_XXS 8
-#define QI2_XXS (QK_K / (4*QR2_XXS))
-typedef struct dpct_type_block_iq2_xxs {
-    sycl::half d;
-    uint16_t qs[QK_K/8];
-} block_iq2_xxs;
-static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
-
-#define QR2_XS 8
-#define QI2_XS (QK_K / (4*QR2_XS))
-typedef struct dpct_type_block_iq2_xs {
-    sycl::half d;
-    uint16_t qs[QK_K/8];
-    uint8_t  scales[QK_K/32];
-} block_iq2_xs;
-static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
-
-#define QR3_XXS 8
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-typedef struct dpct_type_block_iq3_xxs {
-    sycl::half d;
-    uint8_t qs[3*(QK_K/8)];
-} block_iq3_xxs;
-static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
-
-#define QR3_XS 8
-#define QI3_XS (QK_K / (4*QR3_XS))
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
-#define IQ3S_N_SCALE QK_K/64
-#endif
-typedef struct {
-    sycl::half d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
-#define QR1_S 8
-#define QI1_S (QK_K / (4*QR1_S))
-typedef struct {
-    sycl::half d;
-    uint8_t qs[QK_K/8];
-    uint8_t scales[QK_K/16];
-} block_iq1_s;
-static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses

@@ -3586,12 +3478,31 @@ class sycl_gpu_mgr {
        int work_group_size = 0;
        std::string gpus_list = "";

+        /*
+        Use all GPUs with same top max compute units
+        */
        sycl_gpu_mgr() {
            detect_sycl_gpu_list_with_max_cu();
            get_allow_gpus();
            create_context_with_gpus();
        }

+        /*
+        Only use the assigned GPU
+        */
+        sycl_gpu_mgr(int main_gpu_id) {
+            sycl::device device = dpct::dev_mgr::instance().get_device(main_gpu_id);
+            dpct::device_info prop;
+            dpct::get_device_info(prop, device);
+            gpus.push_back(main_gpu_id);
+            devices.push_back(device);
+            work_group_size = prop.get_max_work_group_size();
+            max_compute_units = prop.get_max_compute_units();
+
+            get_allow_gpus();
+            create_context_with_gpus();
+        }
+
        void create_context_with_gpus() {
            sycl::context ctx = sycl::context(devices);
            assert(gpus.size() > 0);
@@ -3607,7 +3518,7 @@ class sycl_gpu_mgr {
                gpus_list += std::to_string(gpus[i]);
                gpus_list += ",";
            }
-            if (gpus_list.length() > 2) {
+            if (gpus_list.length() > 1) {
                gpus_list.pop_back();
            }
        }
@@ -3636,7 +3547,7 @@ class sycl_gpu_mgr {
                dpct::device_info prop;
                dpct::get_device_info(prop, device);
                if (max_compute_units == prop.get_max_compute_units() &&
-                    prop.get_major_version() == 1) {
+                    is_ext_oneapi_device(device)) {
                    gpus.push_back(id);
                    devices.push_back(device);
                    work_group_size = prop.get_max_work_group_size();
@@ -3656,8 +3567,8 @@ class sycl_gpu_mgr {
                if (gpus[i] == id)
                    return i;
            }
-            assert(false);
-            return -1;
+            printf("miss to get device index by id=%d\n", id);
+            GGML_ASSERT(false);
        }

        int get_next_index(int id) {
@@ -3666,8 +3577,16 @@ class sycl_gpu_mgr {
                if (gpus[i] == id)
                    return i;
            }
-            assert(false);
-            return -1;
+            GGML_ASSERT(false);
+        }
+
+        bool is_ext_oneapi_device(const sycl::device &dev) {
+            sycl::backend dev_backend = dev.get_backend();
+            if (dev_backend == sycl::backend::ext_oneapi_level_zero ||
+                dev_backend == sycl::backend::ext_oneapi_cuda ||
+                dev_backend == sycl::backend::ext_oneapi_hip)
+                return true;
+            return false;
        }
 };

@@ -3676,11 +3595,14 @@ static int g_device_count = -1;
 static int g_all_sycl_device_count = -1;
 static int g_main_device = -1;
 static int g_main_device_id = -1;
+static bool g_ggml_backend_sycl_buffer_type_initialized = false;

 static std::array<float, GGML_SYCL_MAX_DEVICES> g_default_tensor_split = {};

 static float g_tensor_split[GGML_SYCL_MAX_DEVICES] = {0};

+static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode = SYCL_UNSET_GPU_MODE;
+
 struct sycl_device_capabilities {
    int     cc;                 // compute capability
    bool    vmm;                // virtual memory support
@@ -4891,10 +4813,9 @@ static void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restr
 template<typename dst_t>
 static void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                     const sycl::nd_item<3> &item_ct1,
-                                     const uint64_t *iq1s_grid,
+                                     const uint32_t *iq1s_grid,
                                     const uint8_t *ksigns_iq2xs,
                                     const uint8_t *kmask_iq2xs) {
-
    const int i = item_ct1.get_group(2);
    const block_iq1_s * x = (const block_iq1_s  *) vx;

@@ -4903,11 +4824,15 @@ static void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restr
    const int il = tid/8; // 0...3
    const int ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const int i8 = 4*ib+il;
-    uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
-    const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
-    const float d = (float)x[i].d * (2*(h & 7) + 1);
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
+    const uint8_t  * qs = x[i].qs + 8*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq1s_grid + qs[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq1s_grid + qs[2*il+1]);
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 0xf) + 1);
+    const uint8_t signs = ksigns_iq2xs[(x[i].qh[ib] >> 3*il) & 7];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
 #else
    assert(false);
 #endif
@@ -7803,28 +7728,27 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
 static __dpct_inline__ float
 vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
                     const block_q8_1 *__restrict__ bq8_1, const int &iqs,
-                     const uint64_t *iq1s_grid, const uint64_t *ksigns64) {
+                     const uint32_t *iq1s_grid, const uint64_t *ksigns64) {
 #if QK_K == 256
    const block_iq1_s * bq1 = (const block_iq1_s *) vbq;

    const int ib32 = iqs;
-    int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
-    const uint8_t h1 = bq1->scales[2*ib32+0];
-    const uint8_t h2 = bq1->scales[2*ib32+1];
-    const int * q8 = (const int *)bq8_1[ib32].qs;
-    const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
-    const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
-    const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
-    const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
-    for (int j = 0; j < 2; ++j) {
-        sumi1 = dpct::dp4a(q8[j+0], grid1[j], sumi1);
-        sumi2 = dpct::dp4a(q8[j+2], grid2[j], sumi2);
-        sumi3 = dpct::dp4a(q8[j+4], grid3[j], sumi3);
-        sumi4 = dpct::dp4a(q8[j+6], grid4[j], sumi4);
+    const uint8_t  * qs = bq1->qs + 4*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq1s_grid + qs[l]);
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (qs[l] >> 8));
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs[0], signs[0], std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs[1], signs[1], std::minus<>());
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        q8 += 8;
    }
-    const float d = (float)bq1->d  * bq8_1[ib32].ds[0];
-    return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
-                sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
+    const float d = (float)bq1->d * bq8_1[ib32].ds[0] * 0.25f;
+    return d * sumi;
 #else
    assert(false);
    return 0.f;
@@ -8644,7 +8568,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void * __restrict__ vx, const void *
 template <int qk, int qi, typename block_q_t, int vdr>
 static void mul_mat_vec_q_iq1_s_q8_1(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
                          const sycl::nd_item<3> &item_ct1,
-                          const uint64_t *iq1s_grid_ptr, const uint64_t *ksigns64_ptr ) {
+                          const uint32_t *iq1s_grid_ptr, const uint64_t *ksigns64_ptr ) {
    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
                    item_ct1.get_local_id(1);

@@ -10406,7 +10330,7 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
                                        dpct::queue_ptr stream) {
    const int nb = k / QK_K;
    {
-        iq1s_grid.init(*stream);
+        iq1s_grid_gpu.init(*stream);
        ksigns_iq2xs.init(*stream);
        kmask_iq2xs.init(*stream);

@@ -10414,7 +10338,7 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k,
                                     {sycl::aspect::fp16});

        stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid.get_ptr();
+            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
            auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr();
            auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr();

@@ -11154,11 +11078,11 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
    const sycl::range<3> block_nums(1, 1, block_num_y);
    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
    {
-        iq1s_grid.init(*stream);
+        iq1s_grid_gpu.init(*stream);
        ksigns64.init(*stream);

        stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = iq1s_grid.get_ptr();
+            auto iq1s_grid_ptr_ct1 = iq1s_grid_gpu.get_ptr();
            auto ksigns64_ptr_ct1 = ksigns64.get_ptr();

            cgh.parallel_for(
@@ -13182,17 +13106,20 @@ bool ggml_sycl_loaded(void) {
    return g_sycl_loaded;
 }

-void print_device_detail(int id) {
+void print_device_detail(int id, sycl::device &device, std::string device_type) {
+
    dpct::device_info prop;
    SYCL_CHECK(CHECK_TRY_ERROR(
-        dpct::get_device_info(prop, dpct::dev_mgr::instance().get_device(id))));
-    sycl::device cur_device = dpct::dev_mgr::instance().get_device(id);
+        dpct::get_device_info(prop, device)));
+
    std::string version;
    version += std::to_string(prop.get_major_version());
    version += ".";
    version += std::to_string(prop.get_minor_version());

-    fprintf(stderr, "|%2d|%45s|%18s|%17d|%14d|%13d|%15lu|\n", id,
+    device_type = std::regex_replace(device_type, std::regex("ext_oneapi_"), "");
+
+    fprintf(stderr, "|%2d|%18s|%45s|%10s|%11d|%8d|%7d|%15lu|\n", id, device_type.c_str(),
            prop.get_name(), version.c_str(), prop.get_max_compute_units(),
            prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
            prop.get_global_mem_size());
@@ -13200,19 +13127,35 @@ void print_device_detail(int id) {

 void ggml_backend_sycl_print_sycl_devices() {
    int device_count = dpct::dev_mgr::instance().device_count();
+    std::map<std::string, size_t> DeviceNums;
    fprintf(stderr, "found %d SYCL devices:\n", device_count);
-    fprintf(stderr, "|ID| Name                                        |compute capability|Max compute units|Max work group|Max sub group|Global mem size|\n");
-    fprintf(stderr, "|--|---------------------------------------------|------------------|-----------------|--------------|-------------|---------------|\n");
+    fprintf(stderr, "|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |\n");
+    fprintf(stderr, "|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|\n");
+    fprintf(stderr, "|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|\n");
    for (int id = 0; id < device_count; ++id) {
-        print_device_detail(id);
+        sycl::device device = dpct::dev_mgr::instance().get_device(id);
+        sycl::backend backend = device.get_backend();
+        std::string backend_type = get_device_backend_and_type(device);
+        int type_id=DeviceNums[backend_type]++;
+        std::stringstream device_type;
+        device_type << "[" <<  backend_type << ":" << std::to_string(type_id) << "]";
+        print_device_detail(id, device, device_type.str());
    }
 }

 void print_gpu_device_list() {
-    fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n",
-            g_sycl_gpu_mgr->get_gpu_count(),
-            g_sycl_gpu_mgr->gpus_list.c_str(),
-            g_sycl_gpu_mgr->max_compute_units);
+    GGML_ASSERT(g_sycl_gpu_mgr);
+
+    char* hint=NULL;
+    if (g_ggml_sycl_backend_gpu_mode == SYCL_SINGLE_GPU_MODE) {
+        hint = "use %d SYCL GPUs: [%s] with Max compute units:%d\n";
+    } else {
+        hint = "detect %d SYCL GPUs: [%s] with top Max compute units:%d\n";
+    }
+    fprintf(stderr, hint,
+        g_sycl_gpu_mgr->get_gpu_count(),
+        g_sycl_gpu_mgr->gpus_list.c_str(),
+        g_sycl_gpu_mgr->max_compute_units);
 }

 int get_sycl_env(const char *env_name, int default_val) {
@@ -13248,23 +13191,6 @@ void ggml_init_sycl() try {
 #else
        fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
 #endif
-        if (CHECK_TRY_ERROR(g_all_sycl_device_count =
-                            dpct::dev_mgr::instance().device_count()) != 0) {
-            initialized = true;
-            g_sycl_loaded = false;
-            return;
-        }
-        GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
-        ggml_backend_sycl_print_sycl_devices();
-
-        if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
-
-        g_device_count = g_sycl_gpu_mgr->get_gpu_count();
-        g_work_group_size = g_sycl_gpu_mgr->work_group_size;
-
-        print_gpu_device_list();
-
-        int64_t total_vram = 0;

 /* NOT REMOVE, keep it for next optimize for XMX.
 #if defined(SYCL_USE_XMX)
@@ -13273,49 +13199,15 @@ void ggml_init_sycl() try {
        fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
 #endif
 */
-        for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) {
-            g_device_caps[id].vmm = 0;
-            g_device_caps[id].device_id = -1;
-            g_device_caps[id].cc = 0;
-            g_tensor_split[id] = 0;
-            g_default_tensor_split[id] = 0;
+
+        if (CHECK_TRY_ERROR(g_all_sycl_device_count =
+                            dpct::dev_mgr::instance().device_count()) != 0) {
+            initialized = true;
+            g_sycl_loaded = false;
+            return;
        }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            int device_id = g_sycl_gpu_mgr->gpus[i];
-            g_device_caps[i].vmm = 0;
-
-            dpct::device_info prop;
-            SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
-                prop, dpct::dev_mgr::instance().get_device(device_id))));
-
-            g_default_tensor_split[i] = total_vram;
-            total_vram += prop.get_global_mem_size();
-
-            g_device_caps[i].cc =
-                100 * prop.get_major_version() + 10 * prop.get_minor_version();
-        }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            g_default_tensor_split[i] /= total_vram;
-        }
-
-        for (int i = 0; i < g_device_count; ++i) {
-            SYCL_CHECK(ggml_sycl_set_device(i));
-
-            // create sycl streams
-            for (int is = 0; is < MAX_STREAMS; ++is) {
-                SYCL_CHECK(CHECK_TRY_ERROR(
-                    g_syclStreams[i][is] =
-                        dpct::get_current_device().create_queue(
-                            g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device())));
-            }
-
-            const dpct::queue_ptr stream = g_syclStreams[i][0];
-            // create sycl handle
-            SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream));
-        }
-
+        GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
+        ggml_backend_sycl_print_sycl_devices();
        initialized = true;
        g_sycl_loaded = true;
    }
@@ -13326,6 +13218,63 @@ catch (sycl::exception const &exc) {
  std::exit(1);
 }

+void ggml_init_by_gpus(int device_count) try {
+    g_device_count = device_count;
+    g_work_group_size = g_sycl_gpu_mgr->work_group_size;
+
+    int64_t total_vram = 0;
+
+    print_gpu_device_list();
+
+    for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) {
+        g_device_caps[id].vmm = 0;
+        g_device_caps[id].device_id = -1;
+        g_device_caps[id].cc = 0;
+        g_tensor_split[id] = 0;
+        g_default_tensor_split[id] = 0;
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        int device_id = g_sycl_gpu_mgr->gpus[i];
+        g_device_caps[i].vmm = 0;
+
+        dpct::device_info prop;
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
+            prop, dpct::dev_mgr::instance().get_device(device_id))));
+
+        g_default_tensor_split[i] = total_vram;
+        total_vram += prop.get_global_mem_size();
+
+        g_device_caps[i].cc =
+            100 * prop.get_major_version() + 10 * prop.get_minor_version();
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        g_default_tensor_split[i] /= total_vram;
+    }
+
+    for (int i = 0; i < g_device_count; ++i) {
+        SYCL_CHECK(ggml_sycl_set_device(i));
+
+        // create sycl streams
+        for (int is = 0; is < MAX_STREAMS; ++is) {
+            SYCL_CHECK(CHECK_TRY_ERROR(
+                g_syclStreams[i][is] =
+                    dpct::get_current_device().create_queue(
+                        g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device())));
+        }
+
+        const dpct::queue_ptr stream = g_syclStreams[i][0];
+        // create sycl handle
+        SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream));
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
 void *ggml_sycl_host_malloc(size_t size) try {
    if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
        return nullptr;
@@ -16725,22 +16674,24 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
    /* .is_host          = */ nullptr,
 };

-ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
+ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) {
+    if (device_index>=g_device_count or device_index<0) {
+        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+            device_index, g_device_count-1);
+        GGML_ASSERT(device_index<g_device_count);
+    }
    static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];

-    static bool ggml_backend_sycl_buffer_type_initialized = false;
-
-    if (!ggml_backend_sycl_buffer_type_initialized) {
+    if (!g_ggml_backend_sycl_buffer_type_initialized) {
        for (int i = 0; i < g_device_count; i++) {
            ggml_backend_sycl_buffer_types[i] = {
                /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(g_sycl_gpu_mgr->gpus[i])},
            };
        }
-        ggml_backend_sycl_buffer_type_initialized = true;
+        g_ggml_backend_sycl_buffer_type_initialized = true;
    }
-
-    return &ggml_backend_sycl_buffer_types[device];
+    return &ggml_backend_sycl_buffer_types[device_index];
 }

 // sycl split buffer type
@@ -17432,13 +17383,19 @@ static ggml_backend_i ggml_backend_sycl_interface = {
    /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type,
    /* .set_tensor_async        = */ ggml_backend_sycl_set_tensor_async,
    /* .get_tensor_async        = */ ggml_backend_sycl_get_tensor_async,
-    /* .cpy_tensor_async        = */ ggml_backend_sycl_cpy_tensor_async,
+    /* .cpy_tensor_async        = */ NULL, //ggml_backend_sycl_cpy_tensor_async, // TODO: update for the new interface
    /* .synchronize             = */ ggml_backend_sycl_synchronize,
    /* .graph_plan_create       = */ NULL,
    /* .graph_plan_free         = */ NULL,
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
    /* .supports_op             = */ ggml_backend_sycl_supports_op,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -17488,11 +17445,42 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) {
    return g_sycl_gpu_mgr->get_index(device_id);
 }

+GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index) {
+    return g_sycl_gpu_mgr->gpus[device_index];
+}
+
+GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id) {
+    GGML_ASSERT(main_gpu_id<g_all_sycl_device_count);
+    fprintf(stderr, "ggml_backend_sycl_set_single_device: use single device: [%d]\n", main_gpu_id);
+    if (g_sycl_gpu_mgr) {
+        delete g_sycl_gpu_mgr;
+    }
+    g_sycl_gpu_mgr = new sycl_gpu_mgr(main_gpu_id);
+    g_ggml_sycl_backend_gpu_mode = SYCL_SINGLE_GPU_MODE;
+    ggml_init_by_gpus(g_sycl_gpu_mgr->get_gpu_count());
+    g_ggml_backend_sycl_buffer_type_initialized = false;
+}
+
+GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode() {
+    if (g_ggml_sycl_backend_gpu_mode == SYCL_MUL_GPU_MODE) {
+        return;
+    }
+
+    fprintf(stderr, "ggml_backend_sycl_set_mul_device_mode: true\n");
+
+    if (g_sycl_gpu_mgr) {
+        delete g_sycl_gpu_mgr;
+    }
+    g_sycl_gpu_mgr = new sycl_gpu_mgr();
+    g_ggml_sycl_backend_gpu_mode = SYCL_MUL_GPU_MODE;
+    ggml_init_by_gpus(g_sycl_gpu_mgr->get_gpu_count());
+    g_ggml_backend_sycl_buffer_type_initialized = false;
+}
+
 extern "C" int ggml_backend_sycl_reg_devices();

 int ggml_backend_sycl_reg_devices() {
-    if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr();
-    g_device_count = g_sycl_gpu_mgr->get_gpu_count();
+    ggml_backend_sycl_set_mul_device_mode();
    assert(g_device_count>0);
    for (int i = 0; i < g_device_count; i++) {
        int id = g_sycl_gpu_mgr->gpus[i];
@@ -29,6 +29,11 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_typ
 GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);

+// TODO: these are temporary
+//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
+GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
+GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
+GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
 #ifdef  __cplusplus
 }
 #endif
@@ -710,6 +710,12 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope
        }
    }

+    // All commands that are allowed on a queue that supports transfer operations are also allowed on a queue that supports either graphics or compute operations.
+    // Thus, if the capabilities of a queue family include VK_QUEUE_GRAPHICS_BIT or VK_QUEUE_COMPUTE_BIT, then reporting the VK_QUEUE_TRANSFER_BIT capability separately for that queue family is optional.
+    if (compute_index >= 0) {
+        return compute_index;
+    }
+
    std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;

    for(auto &q_family : queue_family_props) {
@@ -5693,6 +5699,12 @@ static ggml_backend_i ggml_backend_vk_interface = {
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_vk_graph_compute,
    /* .supports_op             = */ ggml_backend_vk_supports_op,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_vk_guid() {
@@ -282,8 +282,6 @@ inline static void * ggml_calloc(size_t num, size_t size) {
 #else
 #include <cblas.h>
 #endif
-#elif defined(GGML_USE_CUBLAS)
-#include "ggml-cuda.h"
 #elif defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
 #elif defined(GGML_USE_VULKAN)
@@ -470,6 +468,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
        .type_size                = sizeof(int32_t),
        .is_quantized             = false,
    },
+    [GGML_TYPE_I64] = {
+        .type_name                = "i64",
+        .blck_size                = 1,
+        .type_size                = sizeof(int64_t),
+        .is_quantized             = false,
+    },
+    [GGML_TYPE_F64] = {
+        .type_name                = "f64",
+        .blck_size                = 1,
+        .type_size                = sizeof(double),
+        .is_quantized             = false,
+        .nrows                    = 1,
+    },
    [GGML_TYPE_F32] = {
        .type_name                = "f32",
        .blck_size                = 1,
@@ -918,6 +929,101 @@ inline static float vaddvq_f32(float32x4_t v) {
    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif

+#elif defined(__AVX512F__)
+
+#define GGML_SIMD
+
+// F32 AVX512
+
+#define GGML_F32_STEP 64
+#define GGML_F32_EPR  16
+
+#define GGML_F32x16         __m512
+#define GGML_F32x16_ZERO    _mm512_setzero_ps()
+#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
+#define GGML_F32x16_LOAD    _mm512_loadu_ps
+#define GGML_F32x16_STORE   _mm512_storeu_ps
+// _mm512_fmadd_ps is defined in AVX512F so no guard is required
+#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32x16_ADD     _mm512_add_ps
+#define GGML_F32x16_MUL     _mm512_mul_ps
+#define GGML_F32x16_REDUCE(res, x)                                    \
+do {                                                                  \
+    int offset = GGML_F32_ARR >> 1;                                   \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    res = _mm512_reduce_add_ps(x[0]);                                 \
+} while (0)
+
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC        GGML_F32x16
+#define GGML_F32_VEC_ZERO   GGML_F32x16_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x16_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x16_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x16_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x16_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x16_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x16_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
+
+// F16 AVX512
+
+// F16 AVX
+
+#define GGML_F16_STEP 64
+#define GGML_F16_EPR  16
+
+// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead
+
+#define GGML_F32Cx16             __m512
+#define GGML_F32Cx16_ZERO        _mm512_setzero_ps()
+#define GGML_F32Cx16_SET1(x)     _mm512_set1_ps(x)
+
+// unlike  _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
+// so F16C guard isn't required
+#define GGML_F32Cx16_LOAD(x)     _mm512_cvtph_ps(_mm256_loadu_si256((__m256i *)(x)))
+#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
+
+#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32Cx16_ADD         _mm512_add_ps
+#define GGML_F32Cx16_MUL         _mm512_mul_ps
+#define GGML_F32Cx16_REDUCE(res, x)                               \
+do {                                                              \
+    int offset = GGML_F32_ARR >> 1;                               \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    res = _mm512_reduce_add_ps(x[0]);                             \
+} while (0)
+
+#define GGML_F16_VEC                GGML_F32Cx16
+#define GGML_F16_VEC_ZERO           GGML_F32Cx16_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32Cx16_SET1
+#define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx16_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32Cx16_FMA
+#define GGML_F16_VEC_ADD            GGML_F32Cx16_ADD
+#define GGML_F16_VEC_MUL            GGML_F32Cx16_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
+
 #elif defined(__AVX__)

 #define GGML_SIMD
@@ -2532,9 +2638,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
            GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
        }

-#if defined(GGML_USE_CUBLAS)
-        ggml_init_cublas();
-#elif defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_CLBLAST)
        ggml_cl_init();
 #elif defined(GGML_USE_VULKAN)
        ggml_vk_init_cpu_assist();
@@ -10997,7 +11101,6 @@ static void ggml_compute_forward_out_prod_f32(
    // nb01 >= nb00 - src0 is not transposed
    //   compute by src0 rows

-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
    // TODO: #if defined(GGML_USE_CLBLAST)

 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
@@ -11197,7 +11300,6 @@ static void ggml_compute_forward_out_prod_q_f32(
    // nb01 >= nb00 - src0 is not transposed
    //   compute by src0 rows

-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
    // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)

    if (params->type == GGML_TASK_TYPE_INIT) {
@@ -11560,8 +11662,6 @@ static void ggml_compute_forward_get_rows_q(
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

-    assert(params->ith == 0);
-
    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
        return;
    }
@@ -11569,7 +11669,7 @@ static void ggml_compute_forward_get_rows_q(
    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);

    const enum ggml_type type = src0->type;
    ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
@@ -11579,17 +11679,25 @@ static void ggml_compute_forward_get_rows_q(
    assert(nb00 == ggml_type_size(type));
    assert(ggml_nrows(dst) == nr);

-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;

-                dequantize_row_q(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        dequantize_row_q(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
    }
 }

@@ -11600,8 +11708,6 @@ static void ggml_compute_forward_get_rows_f16(
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

-    assert(params->ith == 0);
-
    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
        return;
    }
@@ -11609,24 +11715,32 @@ static void ggml_compute_forward_get_rows_f16(
    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);

    assert(ne0  == nc);
    assert(ne02 == ne11);
    assert(nb00 == sizeof(ggml_fp16_t));
    assert(ggml_nrows(dst) == nr);

-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;

-                ggml_fp16_to_fp32_row(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        ggml_fp16_to_fp32_row(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
    }
 }

@@ -11637,8 +11751,6 @@ static void ggml_compute_forward_get_rows_f32(
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

-    assert(params->ith == 0);
-
    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
        return;
    }
@@ -11646,24 +11758,32 @@ static void ggml_compute_forward_get_rows_f32(
    GGML_TENSOR_BINARY_OP_LOCALS

    const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);

    assert(ne0  == nc);
    assert(ne02 == ne11);
    assert(nb00 == sizeof(float));
    assert(ggml_nrows(dst) == nr);

-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;

-                ggml_vec_cpy_f32(nc,
-                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
-                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        ggml_vec_cpy_f32(nc,
+                (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
    }
 }

@@ -12400,6 +12520,8 @@ static void ggml_compute_forward_alibi(
        case GGML_TYPE_I8:
        case GGML_TYPE_I16:
        case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+        case GGML_TYPE_F64:
        case GGML_TYPE_COUNT:
            {
                GGML_ASSERT(false);
@@ -12486,6 +12608,8 @@ static void ggml_compute_forward_clamp(
        case GGML_TYPE_I8:
        case GGML_TYPE_I16:
        case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+        case GGML_TYPE_F64:
        case GGML_TYPE_COUNT:
            {
                GGML_ASSERT(false);
@@ -15921,14 +16045,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
        return;
    }

-#ifdef GGML_USE_CUBLAS
-    bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
-    if (skip_cpu) {
-        return;
-    }
-    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU);
-    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
-#elif defined(GGML_USE_VULKAN)
+#if defined(GGML_USE_VULKAN)
    const bool skip_cpu = ggml_vk_compute_forward_cpu_assist(params, tensor);
 #ifdef GGML_VULKAN_CHECK_RESULTS
    if (skip_cpu) {
@@ -15940,7 +16057,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
    }
    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU);
    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
-#endif // GGML_USE_CUBLAS
+#endif // GGML_USE_VULKAN

 #ifdef GGML_USE_SYCL
    bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
@@ -17796,7 +17913,7 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const
    node->perf_time_us += time_us_cur;
 }

-static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_threads) {
    int n_tasks = 0;

    switch (node->op) {
@@ -17877,6 +17994,12 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
            {
                n_tasks = n_threads;
            } break;
+        case GGML_OP_GET_ROWS:
+            {
+                // FIXME: the cost of launching additional threads decreases performance with GPU offloading
+                //n_tasks = MIN(n_threads, ggml_nelements(node->src[1]));
+                n_tasks = MIN(n_cur_threads, ggml_nelements(node->src[1]));
+            } break;
        case GGML_OP_SCALE:
        case GGML_OP_SET:
        case GGML_OP_CONT:
@@ -17884,7 +18007,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
        case GGML_OP_VIEW:
        case GGML_OP_PERMUTE:
        case GGML_OP_TRANSPOSE:
-        case GGML_OP_GET_ROWS:
        case GGML_OP_GET_ROWS_BACK:
        case GGML_OP_DIAG:
            {
@@ -18102,7 +18224,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
                /* FINALIZE */
                struct ggml_tensor * node = cgraph->nodes[node_n];
                if (GGML_OP_HAS_FINALIZE[node->op]) {
-                    params.nth = ggml_get_n_tasks(node, n_threads);
+                    params.nth = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);
                    ggml_compute_forward(&params, node);
                }
                ggml_graph_compute_perf_stats_node(node, state->shared);
@@ -18112,7 +18234,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
            while (++node_n < cgraph->n_nodes) {
                GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
                struct ggml_tensor * node = cgraph->nodes[node_n];
-                const int n_tasks = ggml_get_n_tasks(node, n_threads);
+                const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);

                state->shared->perf_node_start_cycles  = ggml_perf_cycles();
                state->shared->perf_node_start_time_us = ggml_perf_time_us();
@@ -18160,7 +18282,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {

        /* INIT & COMPUTE */
        struct ggml_tensor * node = cgraph->nodes[node_n];
-        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+        const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads);

        struct ggml_compute_params params = {
            /*.type  =*/ GGML_TASK_TYPE_INIT,
@@ -18225,7 +18347,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * node = cgraph->nodes[i];

-        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+        const int n_tasks = ggml_get_n_tasks(node, n_threads, 1);

        max_tasks = MAX(max_tasks, n_tasks);

@@ -337,24 +337,24 @@ extern "C" {
    struct ggml_object;
    struct ggml_context;

+    // NOTE: always add types at the end of the enum to keep backward compatibility
    enum ggml_type {
-        GGML_TYPE_F32  = 0,
-        GGML_TYPE_F16  = 1,
-        GGML_TYPE_Q4_0 = 2,
-        GGML_TYPE_Q4_1 = 3,
+        GGML_TYPE_F32     = 0,
+        GGML_TYPE_F16     = 1,
+        GGML_TYPE_Q4_0    = 2,
+        GGML_TYPE_Q4_1    = 3,
        // GGML_TYPE_Q4_2 = 4, support has been removed
-        // GGML_TYPE_Q4_3 (5) support has been removed
-        GGML_TYPE_Q5_0 = 6,
-        GGML_TYPE_Q5_1 = 7,
-        GGML_TYPE_Q8_0 = 8,
-        GGML_TYPE_Q8_1 = 9,
-        // k-quantizations
-        GGML_TYPE_Q2_K = 10,
-        GGML_TYPE_Q3_K = 11,
-        GGML_TYPE_Q4_K = 12,
-        GGML_TYPE_Q5_K = 13,
-        GGML_TYPE_Q6_K = 14,
-        GGML_TYPE_Q8_K = 15,
+        // GGML_TYPE_Q4_3 = 5, support has been removed
+        GGML_TYPE_Q5_0    = 6,
+        GGML_TYPE_Q5_1    = 7,
+        GGML_TYPE_Q8_0    = 8,
+        GGML_TYPE_Q8_1    = 9,
+        GGML_TYPE_Q2_K    = 10,
+        GGML_TYPE_Q3_K    = 11,
+        GGML_TYPE_Q4_K    = 12,
+        GGML_TYPE_Q5_K    = 13,
+        GGML_TYPE_Q6_K    = 14,
+        GGML_TYPE_Q8_K    = 15,
        GGML_TYPE_IQ2_XXS = 16,
        GGML_TYPE_IQ2_XS  = 17,
        GGML_TYPE_IQ3_XXS = 18,
@@ -363,9 +363,11 @@ extern "C" {
        GGML_TYPE_IQ3_S   = 21,
        GGML_TYPE_IQ2_S   = 22,
        GGML_TYPE_IQ4_XS  = 23,
-        GGML_TYPE_I8,
-        GGML_TYPE_I16,
-        GGML_TYPE_I32,
+        GGML_TYPE_I8      = 24,
+        GGML_TYPE_I16     = 25,
+        GGML_TYPE_I32     = 26,
+        GGML_TYPE_I64     = 27,
+        GGML_TYPE_F64     = 28,
        GGML_TYPE_COUNT,
    };

@@ -383,20 +385,20 @@ extern "C" {

    // model file types
    enum ggml_ftype {
-        GGML_FTYPE_UNKNOWN     = -1,
-        GGML_FTYPE_ALL_F32     = 0,
-        GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        GGML_FTYPE_UNKNOWN        = -1,
+        GGML_FTYPE_ALL_F32        = 0,
+        GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
        GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
        GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
        GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
        GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
@@ -32,6 +32,7 @@ class Keys:
        FILE_TYPE            = "general.file_type"

    class LLM:
+        VOCAB_SIZE            = "{arch}.vocab_size"
        CONTEXT_LENGTH        = "{arch}.context_length"
        EMBEDDING_LENGTH      = "{arch}.embedding_length"
        BLOCK_COUNT           = "{arch}.block_count"
@@ -41,6 +42,7 @@ class Keys:
        EXPERT_COUNT          = "{arch}.expert_count"
        EXPERT_USED_COUNT     = "{arch}.expert_used_count"
        POOLING_TYPE          = "{arch}.pooling_type"
+        LOGIT_SCALE           = "{arch}.logit_scale"

    class Attention:
        HEAD_COUNT        = "{arch}.attention.head_count"
@@ -120,6 +122,7 @@ class MODEL_ARCH(IntEnum):
    GEMMA      = auto()
    STARCODER2 = auto()
    MAMBA      = auto()
+    COMMAND_R  = auto()


 class MODEL_TENSOR(IntEnum):
@@ -186,6 +189,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.GEMMA:          "gemma",
    MODEL_ARCH.STARCODER2:     "starcoder2",
    MODEL_ARCH.MAMBA:          "mamba",
+    MODEL_ARCH.COMMAND_R:      "command-r",
 }

 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -578,6 +582,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.SSM_D,
        MODEL_TENSOR.SSM_OUT,
    ],
+    MODEL_ARCH.COMMAND_R: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    # TODO
 }

@@ -661,6 +677,11 @@ class GGMLQuantizationType(IntEnum):
    IQ3_S   = 21
    IQ2_S   = 22
    IQ4_XS  = 23
+    I8      = 24
+    I16     = 25
+    I32     = 26
+    I64     = 27
+    F64     = 28


 class GGUFEndian(IntEnum):
@@ -727,6 +748,11 @@ GGML_QUANT_SIZES = {
    GGMLQuantizationType.IQ3_S:   (256, 2 + QK_K // 4 + QK_K // 8 + QK_K // 32 + 4),
    GGMLQuantizationType.IQ2_S:   (256, 2 + QK_K // 4 + QK_K // 16),
    GGMLQuantizationType.IQ4_XS:  (256, 2 + 2 + QK_K // 2 + QK_K // 64),
+    GGMLQuantizationType.I8:      (1, 1),
+    GGMLQuantizationType.I16:     (1, 2),
+    GGMLQuantizationType.I32:     (1, 4),
+    GGMLQuantizationType.I64:     (1, 8),
+    GGMLQuantizationType.F64:     (1, 8),
 }


@@ -746,6 +772,7 @@ KEY_GENERAL_SOURCE_HF_REPO       = Keys.General.SOURCE_HF_REPO
 KEY_GENERAL_FILE_TYPE            = Keys.General.FILE_TYPE

 # LLM
+KEY_VOCAB_SIZE            = Keys.LLM.VOCAB_SIZE
 KEY_CONTEXT_LENGTH        = Keys.LLM.CONTEXT_LENGTH
 KEY_EMBEDDING_LENGTH      = Keys.LLM.EMBEDDING_LENGTH
 KEY_BLOCK_COUNT           = Keys.LLM.BLOCK_COUNT
@@ -242,12 +242,27 @@ class GGUFReader:
            n_bytes = n_elems * type_size // block_size
            data_offs = int(start_offs + offset_tensor[0])
            item_type: npt.DTypeLike
-            if ggml_type == GGMLQuantizationType.F32:
-                item_count = n_elems
-                item_type = np.float32
-            elif ggml_type == GGMLQuantizationType.F16:
+            if ggml_type == GGMLQuantizationType.F16:
                item_count = n_elems
                item_type = np.float16
+            elif ggml_type == GGMLQuantizationType.F32:
+                item_count = n_elems
+                item_type = np.float32
+            elif ggml_type == GGMLQuantizationType.F64:
+                item_count = n_elems
+                item_type = np.float64
+            elif ggml_type == GGMLQuantizationType.I8:
+                item_count = n_elems
+                item_type = np.int8
+            elif ggml_type == GGMLQuantizationType.I16:
+                item_count = n_elems
+                item_type = np.int16
+            elif ggml_type == GGMLQuantizationType.I32:
+                item_count = n_elems
+                item_type = np.int32
+            elif ggml_type == GGMLQuantizationType.I64:
+                item_count = n_elems
+                item_type = np.int64
            else:
                item_count = n_bytes
                item_type = np.uint8
@@ -196,9 +196,6 @@ class GGUFWriter:
        if self.state is not WriterState.EMPTY:
            raise ValueError(f'Expected output file to be empty, got {self.state}')

-        if raw_dtype is None and tensor_dtype not in (np.float32, np.float16):
-            raise ValueError("Only F32 and F16 tensors are supported for now")
-
        encoded_name = name.encode("utf8")
        self.ti_data += self._pack("Q", len(encoded_name))
        self.ti_data += encoded_name
@@ -207,7 +204,22 @@ class GGUFWriter:
        for i in range(n_dims):
            self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
        if raw_dtype is None:
-            dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
+            if tensor_dtype == np.float16:
+                dtype = GGMLQuantizationType.F16
+            elif tensor_dtype == np.float32:
+                dtype = GGMLQuantizationType.F32
+            elif tensor_dtype == np.float64:
+                dtype = GGMLQuantizationType.F64
+            elif tensor_dtype == np.int8:
+                dtype = GGMLQuantizationType.I8
+            elif tensor_dtype == np.int16:
+                dtype = GGMLQuantizationType.I16
+            elif tensor_dtype == np.int32:
+                dtype = GGMLQuantizationType.I32
+            elif tensor_dtype == np.int64:
+                dtype = GGMLQuantizationType.I64
+            else:
+                raise ValueError("Only F16, F32, F64, I8, I16, I32, I64 tensors are supported for now")
        else:
            dtype = raw_dtype
        self.ti_data += self._pack("I", dtype)
@@ -313,6 +325,9 @@ class GGUFWriter:
        self.data_alignment = alignment
        self.add_uint32(Keys.General.ALIGNMENT, alignment)

+    def add_vocab_size(self, size: int) -> None:
+        self.add_uint32(Keys.LLM.VOCAB_SIZE.format(arch=self.arch), size)
+
    def add_context_length(self, length: int) -> None:
        self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length)

@@ -346,6 +361,9 @@ class GGUFWriter:
    def add_clamp_kqv(self, value: float) -> None:
        self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)

+    def add_logit_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
+
    def add_expert_count(self, count: int) -> None:
        self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.7.0"
+version = "0.8.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
@@ -59,9 +59,10 @@ extern "C" {
    typedef int32_t llama_seq_id;

    enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
-        LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
-        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
+        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
    };

    // note: these values should be synchronized with ggml_rope
@@ -234,8 +235,9 @@ extern "C" {
    struct llama_context_params {
        uint32_t seed;              // RNG seed, -1 for random
        uint32_t n_ctx;             // text context, 0 = from model
-        uint32_t n_batch;           // prompt processing maximum batch size
-        uint32_t n_parallel;        // number of parallel sequences (i.e. distinct states for recurrent models)
+        uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
+        uint32_t n_ubatch;          // physical maximum batch size
+        uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
        uint32_t n_threads;         // number of threads to use for generation
        uint32_t n_threads_batch;   // number of threads to use for batch processing

@@ -377,7 +379,8 @@ extern "C" {

    LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
    LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
-    LLAMA_API uint32_t llama_n_max_seq  (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
+    LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);

    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);
@@ -385,6 +388,7 @@ extern "C" {
    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);

    // Get the model's RoPE frequency scaling factor
    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
@@ -432,10 +436,24 @@ extern "C" {
    // Returns 0 on success
    LLAMA_API int32_t llama_model_apply_lora_from_file(
            const struct llama_model * model,
-                      const char * path_lora,
-                           float   scale,
-                      const char * path_base_model,
-                         int32_t   n_threads);
+                          const char * path_lora,
+                               float   scale,
+                          const char * path_base_model,
+                             int32_t   n_threads);
+
+    // Apply a loaded control vector to a llama_context, or if data is NULL, clear
+    // the currently loaded vector.
+    // n_embd should be the size of a single layer's control, and data should point
+    // to an n_embd x n_layers buffer starting from layer 1.
+    // il_start and il_end are the layer range the vector should apply to (both inclusive)
+    // See llama_control_vector_load in common to load a control vector.
+    LLAMA_API int32_t llama_control_vector_apply(
+            struct llama_context * lctx,
+                     const float * data,
+                          size_t   len,
+                         int32_t   n_embd,
+                         int32_t   il_start,
+                         int32_t   il_end);

    //
    // KV cache
@@ -456,7 +474,7 @@ extern "C" {
        // Maximum number of sequences that can exist in a cell. It's not an error
        // if there are more sequences in a cell than this value, however they will
        // not be visible in the view cells_sequences.
-        int32_t n_max_seq;
+        int32_t n_seq_max;

        // Number of tokens in the cache. For example, if there are two populated
        // cells, the first with 1 sequence id in it and the second with 2 sequence
@@ -476,12 +494,12 @@ extern "C" {
        // Information for an individual cell.
        struct llama_kv_cache_view_cell * cells;

-        // The sequences for each cell. There will be n_max_seq items per cell.
+        // The sequences for each cell. There will be n_seq_max items per cell.
        llama_seq_id * cells_sequences;
    };

    // Create an empty KV cache view. (use only for debugging purposes)
-    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq);
+    LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max);

    // Free a KV cache view. (use only for debugging purposes)
    LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view);
@@ -650,6 +668,11 @@ extern "C" {
    // Set abort callback
    LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);

+    // Wait until all computations are finished
+    // This is automatically done when using one of the functions below to obtain the computation results
+    // and is not necessary to call it explicitly in most cases
+    LLAMA_API void llama_synchronize(struct llama_context * ctx);
+
    // Token logits obtained from the last call to llama_decode()
    // The logits for the last token are stored in the last row
    // Logits for which llama_batch.logits[i] == 0 are undefined
@@ -708,7 +731,7 @@ extern "C" {

    /// @details Convert the provided text into tokens.
    /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
-    /// @return Returns the number of tokens on success, no more than n_max_tokens
+    /// @return Returns the number of tokens on success, no more than n_tokens_max
    /// @return Returns a negative number on failure - the number of tokens that would have been returned
    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
    ///                Does not insert a leading space.
@@ -717,7 +740,7 @@ extern "C" {
                      const char * text,
                         int32_t   text_len,
                     llama_token * tokens,
-                         int32_t   n_max_tokens,
+                         int32_t   n_tokens_max,
                            bool   add_bos,
                            bool   special);

@@ -2222,8 +2222,8 @@ static void usage(char ** argv) {

 int main(int argc, char ** argv) {
    test_mode mode = MODE_TEST;
-    const char * op_name = NULL;
-    const char * backend = NULL;
+    const char * op_name_filter = NULL;
+    const char * backend_filter = NULL;

    for (int i = 1; i < argc; i++) {
        if (strcmp(argv[i], "test") == 0) {
@@ -2232,14 +2232,14 @@ int main(int argc, char ** argv) {
            mode = MODE_PERF;
        } else if (strcmp(argv[i], "-o") == 0) {
            if (i + 1 < argc) {
-                op_name = argv[++i];
+                op_name_filter = argv[++i];
            } else {
                usage(argv);
                return 1;
            }
        } else if (strcmp(argv[i], "-b") == 0) {
            if (i + 1 < argc) {
-                backend = argv[++i];
+                backend_filter = argv[++i];
            } else {
                usage(argv);
                return 1;
@@ -2258,7 +2258,7 @@ int main(int argc, char ** argv) {
    for (size_t i = 0; i < ggml_backend_reg_get_count(); i++) {
        printf("Backend %zu/%zu (%s)\n", i + 1, ggml_backend_reg_get_count(), ggml_backend_reg_get_name(i));

-        if (backend != NULL && strcmp(backend, ggml_backend_reg_get_name(i)) != 0) {
+        if (backend_filter != NULL && strcmp(backend_filter, ggml_backend_reg_get_name(i)) != 0) {
            printf("  Skipping\n");
            n_ok++;
            continue;
@@ -2266,9 +2266,17 @@ int main(int argc, char ** argv) {

        ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL);
        GGML_ASSERT(backend != NULL);
+
+        if (backend_filter == NULL && ggml_backend_is_cpu(backend)) {
+            printf("  Skipping CPU backend\n");
+            ggml_backend_free(backend);
+            n_ok++;
+            continue;
+        }
+
        printf("  Backend name: %s\n", ggml_backend_name(backend));

-        bool ok = test_backend(backend, mode, op_name);
+        bool ok = test_backend(backend, mode, op_name_filter);

        printf("  Backend %s: ", ggml_backend_name(backend));
        if (ok) {
@@ -31,6 +31,8 @@ int main(void) {
        "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}",
        // google/gemma-7b-it
        "{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\\n' + message['content'] | trim + '<end_of_turn>\\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\\n'}}{% endif %}",
+        // OrionStarAI/Orion-14B-Chat
+        "{% for message in messages %}{% if loop.first %}{{ bos_token }}{% endif %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'] + '\\n\\nAssistant: ' + eos_token }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token }}{% endif %}{% endfor %}",
    };
    std::vector<std::string> expected_output = {
        // teknium/OpenHermes-2.5-Mistral-7B
@@ -45,6 +47,8 @@ int main(void) {
        "system\nYou are a helpful assistant</s>\n<s>user\nHello</s>\n<s>assistant\nHi there</s>\n<s>user\nWho are you</s>\n<s>assistant\n   I am an assistant   </s>\n<s>user\nAnother question</s>\n<s>assistant\n",
        // google/gemma-7b-it
        "<start_of_turn>user\nYou are a helpful assistant\n\nHello<end_of_turn>\n<start_of_turn>model\nHi there<end_of_turn>\n<start_of_turn>user\nWho are you<end_of_turn>\n<start_of_turn>model\nI am an assistant<end_of_turn>\n<start_of_turn>user\nAnother question<end_of_turn>\n<start_of_turn>model\n",
+        // OrionStarAI/Orion-14B-Chat
+        "Human: You are a helpful assistant\n\nHello\n\nAssistant: </s>Hi there</s>Human: Who are you\n\nAssistant: </s>   I am an assistant   </s>Human: Another question\n\nAssistant: </s>",
    };
    std::vector<char> formatted_chat(1024);
    int32_t res;
Author	SHA1	Message	Date
Georgi GerganovandGitHub	b80cf3b2d1	common : disable repeat penalties by default (#6127 )	2024-03-19 10:21:54 +02:00
slarenandGitHub	970a48060a	ci : exempt some labels from being tagged as stale (#6140 )	2024-03-19 10:06:54 +02:00
DAN™andGitHub	4c28b82529	common : print usage on '-h' and '--help' (#6145 )	2024-03-19 07:59:36 +02:00
github-actions[bot]andSomeone	2d15886bb0	flake.lock: Update Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/9df3e30ce24fd28c7b3e2de0d986769db5d6225d' (2024-03-06) → 'github:NixOS/nixpkgs/d691274a972b3165335d261cc4671335f5c67de9' (2024-03-14)	2024-03-18 18:51:30 +00:00
Jared Van BortelandGitHub	d199ca79f2	mpt : implement backwards compatiblity with duped output tensor (#6139 )	2024-03-18 12:49:02 -04:00
FelixandGitHub	104f5e0fc1	clip : fix memory leak (#6138 )	2024-03-18 17:40:22 +02:00
slarenandGitHub	5e1b7f94a0	backend : set max split inputs to GGML_MAX_SRC (#6137 )	2024-03-18 16:33:44 +01:00
Georgi GerganovandGitHub	ac9ee6a4ad	ci : disable stale issue messages (#6126 )	2024-03-18 13:45:38 +02:00
Georgi GerganovandGitHub	4f6d1337ca	ci : temporary disable sanitizer builds (#6128 )	2024-03-18 13:45:27 +02:00
slaren GitHub Johannes Gäßler	2bf8d0f7c4	backend : offload large batches to GPU (#6083 ) * backend : offload large batches to GPU * fix hip * code cleanup * fix CUDA split buffers * Update ggml-backend-impl.h Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix memset without set_device * imatrix : remove sched affix from weight names * sched : add a new split if the current one has too many inputs reduce max inputs per split more cleanup * update backends ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2024-03-18 11:03:04 +01:00
DAN™GitHub Georgi Gerganov	496bc79bc2	common : tidy-up argument parsing (#6105 ) * Tidy-up argument parsing. * Missing ref. * common : minor * common : add static classifier --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-18 10:27:44 +02:00
ThérenceandGitHub	9b03719ad7	convert : add support for CamembertModel architecture (#6119 ) Adding support for CamembertModel architecture used by : https://huggingface.co/dangvantuan/sentence-camembert-large	2024-03-18 10:17:00 +02:00
Romain DandGitHub	3a6efdd03c	convert : use f32 outtype for bf16 tensors (#6106 ) The old behaviour is to use f16, but bf16 to f16 is not a lossless conversion. Change the outtype to f32 to default to a lossless conversion.	2024-03-18 10:04:41 +02:00
Pierrick Hymbert GitHub Georgi Gerganov	d01b3c4c32	common: llama_load_model_from_url using --model-url (#6098 ) * common: llama_load_model_from_url with libcurl dependency Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-17 19:12:37 +01:00
Georgi GerganovandGitHub	cd776c37c9	ci : close all stale issues at once (#6115 )	2024-03-17 18:51:57 +01:00
GainLee GitHub GainLee	dc0f612548	ggml:fix finding transfer queue family index error (#6094 ) Co-authored-by: GainLee <ligen@meizu.com>	2024-03-17 18:12:22 +01:00
AmirAli MirianandGitHub	c47cf414ef	ggml : add AVX512F SIMD (#6088 )	2024-03-16 17:52:02 +02:00
Daniel BeveniusandGitHub	b5f4ae09c3	gritlm : add initial README.md (#6086 ) * gritlm: add initial README.md to examples/gritlm This commit adds a suggestion for an initial README.md for the gritlm example. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> * squash! gritlm: add initial README.md to examples/gritlm Use the `scripts/hf.sh` script to download the model file. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> * squash! gritlm: add initial README.md to examples/gritlm Fix editorconfig-checker error in examples/gritlm/README.md. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com> --------- Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>	2024-03-16 17:46:29 +02:00
Xuan Son NguyenandGitHub	dfbfdd60f9	readme : add wllama as a wasm binding (#6100 )	2024-03-16 17:42:08 +02:00
DAN™andGitHub	15961ec04d	common : refactor nested if causing error C1061 on MSVC (#6101 ) * Refactor nested if causing error C1061 on MSVC. * Revert back and remove else's. * Add flag to track found arguments.	2024-03-16 17:39:15 +02:00
Pierrick HymbertandGitHub	a56d09a440	ci : close inactive issue with workflow (#6053 ) * issues: ci - close inactive issue with workflow * ci: close issue, change workflow schedule time	2024-03-16 14:20:53 +02:00
slarenandGitHub	d84c48505f	llama : fix Baichuan2 13B (#6092 )	2024-03-15 23:14:16 +02:00
Theia Vogel GitHub Georgi Gerganov	877b4d0c62	llama : add support for control vectors (#5970 ) * control vector api and implementation * control-vectors : minor code style updates * disable control vector when data == nullptr use -1 for disabled range (also on init) in case we ever support controlling layer 0 (embeddings) --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-15 22:43:02 +02:00
Andrew CanisandGitHub	12247f4c69	llama : add Command-R support (#6033 ) Information about the Command-R 35B model (128k context) can be found at: https://huggingface.co/CohereForAI/c4ai-command-r-v01 Based on the llama2 model with a few changes: 1) New hyper parameter to scale output logits (logit_scale) 2) Uses LayerNorm instead of RMSNorm 3) Transfomer layers have a single shared LayerNorm that feeds into both the self-attention and FFN layers in parallel. There is no post-attention LayerNorm. 4) No support for Rotary Position Embeddings (RoPE) scaling 5) No biases used Find GGUF files here: https://huggingface.co/andrewcanis/c4ai-command-r-v01-GGUF To convert model to GGUF format yourself: 1) Download Command-R Hugging Face safetensors: git lfs install git clone https://huggingface.co/CohereForAI/c4ai-command-r-v01 2) Run: python3 convert-hf-to-gguf.py --outtype f16 ./c4ai-command-r-v01	2024-03-15 22:41:22 +02:00
Ting Lou GitHub Lou Ting	4e9a7f7f7f	llava : change API to pure C style for Rust FFI bindgen (#6079 ) Co-authored-by: Lou Ting <louting.t@alibaba-inc.com>	2024-03-15 16:31:05 +02:00
slarenandGitHub	3020327f6c	cuda : disable unused cudaLaunchHostFunc code (#6078 )	2024-03-15 14:24:03 +02:00
Neo Zhang JianyuandGitHub	46acb36767	fix set main gpu error (#6073 )	2024-03-15 18:53:53 +08:00
Georgi Gerganov	131b058409	make : ggml-metal.o depends on ggml.h	2024-03-15 11:38:40 +02:00
AidanBeltonS GitHub Neo Zhang Jianyu Abhilash Majumder	753e36f650	[SYCL] Fix non-intel device selection (#6042 ) * Fix non-intel device selection * Update ggml-sycl.cpp Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com> * Update ggml-sycl.cpp Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com> --------- Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> Co-authored-by: Neo Zhang Jianyu <jianyu.zhang@intel.com>	2024-03-15 14:56:20 +05:30
Ondřej ČertíkandGitHub	7ce2c77f88	gguf : add support for I64 and F64 arrays (#6062 ) * gguf : add support for I64 and F64 arrays GGML currently does not support I64 or F64 arrays and they are not often used in machine learning, however if in the future the need arises, it would be nice to add them now, so that the types are next to the other types I8, I16, I32 in the enums, and it also reserves their type number. Furthermore, with this addition the GGUF format becomes very usable for most computational applications of NumPy (being compatible with the most common NumPy dtypes: i8, i16, i32, i64, f32, f64), providing a faster, and more versatile alternative to the `npz` format, and a simpler alternative to the `hdf5` format. The change in this PR seems small, not significantly increasing the maintenance burden. I tested this from Python using GGUFWriter/Reader and `gguf-dump`, as well as from C, everything seems to work. * Fix compiler warnings	2024-03-15 10:46:51 +02:00
Xuan Son NguyenandGitHub	aab606a11f	llama : add Orion chat template (#6066 )	2024-03-15 10:44:57 +02:00
slarenandGitHub	b0bc9f4a9d	llama-bench : use random tokens to improve accuracy with mixtral (#6069 )	2024-03-15 10:22:24 +02:00
Georgi GerganovandGitHub	4755afd1cb	llama : fix integer overflow during quantization (#6063 )	2024-03-14 22:58:41 +02:00
Steve GrubbandGitHub	6e0438da3c	gguf : fix resource leaks (#6061 ) There several places where a gguf context is allocated. A call to gguf_free is missing in some error paths. Also on linux, llama-bench was missing a fclose.	2024-03-14 20:29:32 +02:00
Ondřej ČertíkandGitHub	727107707a	gguf-py : bump version to 0.8.0 (#6060 )	2024-03-14 19:57:31 +02:00
Michael PodvitskiyandGitHub	69ff61397d	llama : support models without vocabulary (#5798 ) * additional methods to read model and ctx parameters * vocab size as a part of a model metadata * models without vocabulary, convert.py part * models without vocabulary, llama.cpp part * PR clean up * converter scrypt fixes * llama_vocab_type update (renamed the new key) * pr review fixes * revert function renaming * one more NoVocab assert	2024-03-14 18:21:56 +02:00
Georgi Gerganov	044ec4b2a5	embedding : add EOS token if not present (#899 )	2024-03-14 15:14:14 +02:00
Georgi Gerganov	77178eedc8	gguf-py : fix dtype check (#6045 )	2024-03-14 13:32:14 +02:00
Jian Liao GitHub Jian Liao	15a333260a	readme : improve readme for Llava-1.6 example (#6044 ) Co-authored-by: Jian Liao <jianliao@adobe.com>	2024-03-14 13:18:23 +02:00
Pierrick HymbertandGitHub	43241adf22	server: disable debug release type sanitizer, simplify trigger (#6047 ) - increase time out for server - do not fail fast	2024-03-14 13:15:39 +02:00
Georgi Gerganov	a44bc969e4	llama : fix typo	2024-03-14 13:13:06 +02:00
Michael Podvitskiy GitHub Georgi Gerganov	2c4fb69246	llama : optimize defrag moves + fix fragmentation calculation (#6037 ) * attempt to reduce the impact of a worst-case scenario * fragmentation calculation fix * Update llama.cpp --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-14 12:56:48 +02:00
Ondřej ČertíkandGitHub	3ca23481dd	gguf-py : add support for I8, I16 and I32 (#6045 ) * Refactor dtype handling to be extensible This code is equivalent as before, but now it is prepared to easily add more NumPy dtypes. * Add support for I8, I16 and I32 These types are allowed in the GGUF specification. * Add support for I8, I16 and I32 to gguf_writer * Add support for I8, I16, I32 to gguf_reader	2024-03-14 12:40:14 +02:00
Georgi GerganovandGitHub	3fe8d7a17f	ggml : designate enum vals for integer types (#6050 )	2024-03-14 12:38:37 +02:00
Georgi Gerganov	68265ebfc6	embedding : print all resulting embeddings (#899 )	2024-03-14 12:37:20 +02:00
Georgi GerganovandGitHub	381da2d9f0	metal : build metallib + fix embed path (#6015 ) * metal : build metallib + fix embed path ggml-ci * metal : fix embed build + update library load logic ggml-ci * metal : fix embeded library build ggml-ci * ci : fix iOS builds to use embedded library	2024-03-14 11:55:23 +02:00
Georgi Gerganov	0fd6c1f015	embedding : print cosine similarity (#899 )	2024-03-14 10:12:29 +02:00
Linwei WangandGitHub	19885d205e	readme : update details about running llama in Termux on Android (#6039 )	2024-03-13 20:34:40 +02:00
Georgi GerganovandGitHub	76a936c893	readme : update API changes and hot topics	2024-03-13 20:33:56 +02:00
Clint HerronandGitHub	463628372d	grammar : handle missing "root" node (#6004 )	2024-03-13 20:10:40 +02:00
slaren GitHub Francis Couture-Harpin Georgi Gerganov	f30ea47a87	llama : add pipeline parallelism support (#6017 ) * llama : add pipeline parallelism support for batch processing with multiple CUDA GPUs ggml-ci * server : add -ub, --ubatch-size parameter * fix server embedding test * llama : fix Mamba inference for pipeline parallelism Tested to work correctly with both `main` and `parallel` examples. * llama : limit max batch size to n_batch * add LLAMA_SCHED_MAX_COPIES to configure the number of input copies for pipeline parallelism default increase to 4 (from 2) changing this value may improve performance for some systems, but increases memory usage * fix hip build * fix sycl build (disable cpy_tensor_async) * fix hip build * llama : limit n_batch and n_ubatch to n_ctx during context creation * llama : fix norm backend * batched-bench : sync after decode * swiftui : sync after decode * ggml : allow ggml_get_rows to use multiple threads if they are available * check n_ubatch >= n_tokens with non-casual attention * llama : do not limit n_batch to n_ctx with non-casual attn * server : construct batch with size of llama_n_batch * ggml_backend_cpu_graph_compute : fix return value when alloc fails * llama : better n_batch and n_ubatch comment * fix merge * small fix * reduce default n_batch to 2048 --------- Co-authored-by: Francis Couture-Harpin <git@compilade.net> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-03-13 18:54:21 +01:00
slarenandGitHub	d8fd0ccf6a	test-backend-ops : skip CPU backend by default (#6028 )	2024-03-13 15:58:30 +02:00
AidanBeltonSandGitHub	b3d978600f	Update get version (#6025 )	2024-03-13 18:47:54 +05:30
Xuan Son NguyenandGitHub	99b71c068f	Server: Use multi-task for embeddings endpoint (#6001 ) * use multitask for embd endpoint * specify types * remove redundant {"n_predict", 0}	2024-03-13 11:39:11 +01:00
slarenandGitHub	306d34be7a	ci : remove tidy-review (#6021 )	2024-03-12 17:55:19 +02:00
Georgi GerganovandGitHub	8030da7afe	ggml : reuse quantum structs across backends (#5943 ) * ggml : reuse quant blocks across backends ggml-ci * ggml : define helper constants only for CUDA and SYCL ggml-ci * ggml : define helper quantum constants for SYCL ggml-ci	2024-03-12 14:27:20 +02:00
Georgi GerganovandGitHub	184215e783	ggml : fix UB in IQ2_S and IQ3_S (#6012 )	2024-03-12 13:49:55 +02:00
Georgi GerganovandGitHub	48358b2e5b	sycl : update IQ1_S kernels (WIP - not working!) (#5995 ) * sycl : try to fix after IQ1_S changes * sycl : iq1s_grid -> iq1s_grid_gpu * sycl : fix grid type	2024-03-12 11:15:05 +02:00
glipticandGitHub	5cdb371731	grammar : fix unnecessarily retained pointer to rules (#6003 )	2024-03-11 21:59:03 +02:00
Kawrakow GitHub Iwan Kawrakow	44ca159faf	1.5 bit: we can do even better (#5999 ) * iq1_s: we can do even better Spent one of the 4 scale bits on a signs of a 0.125 shift. I.e., quants are now -1 + delta, delta, 1 + delta, where delta is +/- 0.125. CUDA works, same performance as before. PPL(LLaMA-v2-7B) is now 11.85! * iq1_s: make scalar and AVX2 work with the new version * iq1_s: make Neon work with new version. ~10% drop in performance, so will need some more work. * iq1_s: make Metal work with new version * iq1_s: very slightly faster dequantize on Metal * iq1_s: fix dequantize on the CPU --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>	2024-03-11 17:53:15 +02:00
Georgi GerganovandGitHub	05b06210c9	llama : more consistent names of count variables (#5994 ) * llama : more consistent names of count variables ggml-ci * llama : n_parallel -> n_seq_max * common : fix param name * examples : fix param name	2024-03-11 17:49:47 +02:00