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
embedding : add EOS token if not present (#899 )
2026-08-03 10:08:03 +02:00 · 2024-03-14 18:21:56 +02:00 · 2024-03-14 15:14:14 +02:00 · 2024-03-14 13:32:14 +02:00 · 2024-03-14 13:18:23 +02:00 · 2024-03-14 13:15:39 +02:00
47 changed files with 2572 additions and 2239 deletions
@@ -333,6 +333,7 @@ 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 +362,7 @@ 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 \
@@ -25,17 +25,14 @@ jobs:
    strategy:
      matrix:
        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        build_type: [Debug]
        include:
          - build_type: Release
            sanitizer: ""
-        exclude:
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
+          - build_type: Debug
            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
@@ -81,13 +78,14 @@ 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
@@ -124,13 +122,14 @@ jobs:

      - 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/

@@ -118,6 +118,7 @@ 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 +148,8 @@ 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 +200,6 @@ 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 +208,62 @@ 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
@@ -553,15 +557,16 @@ ggml-metal.o: ggml-metal.m ggml-metal.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
@@ -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
@@ -902,6 +904,9 @@ 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 +915,28 @@ $ 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

@@ -483,6 +483,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                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 == "--keep") {
            if (++i >= argc) {
                invalid_param = true;
@@ -977,7 +983,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    printf("                        binary file containing multiple choice tasks.\n");
    printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
    printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
-    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\n");
+    printf("                        physical maximum batch size (default: %d)\n", params.n_ubatch);
    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
    printf("                        (default: %s)\n", sampler_type_names.c_str());
    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
@@ -1287,8 +1295,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
    auto cparams = llama_context_default_params();

    cparams.n_ctx             = params.n_ctx;
+    cparams.n_seq_max         = params.n_parallel;
    cparams.n_batch           = params.n_batch;
-    cparams.n_parallel        = params.n_parallel;
+    cparams.n_ubatch          = params.n_ubatch;
    cparams.n_threads         = params.n_threads;
    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
    cparams.seed              = params.seed;
@@ -1379,6 +1388,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
        std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
        llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
        llama_kv_cache_clear(lctx);
+        llama_synchronize(lctx);
        llama_reset_timings(lctx);
    }

@@ -1786,17 +1796,17 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";

    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);

    llama_kv_cache_view_cell * c_curr = view.cells;
    llama_seq_id * cs_curr = view.cells_sequences;

-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
        if (i % row_size == 0) {
            printf("\n%5d: ", i);
        }
        int seq_count = 0;
-        for (int j = 0; j < view.n_max_seq; j++) {
+        for (int j = 0; j < view.n_seq_max; j++) {
            if (cs_curr[j] >= 0) { seq_count++; }
        }
        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
@@ -1809,14 +1819,14 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";

    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+        view.n_cells, view.n_seq_max, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);

    std::unordered_map<llama_seq_id, size_t> seqs;
    llama_kv_cache_view_cell * c_curr = view.cells;
    llama_seq_id * cs_curr = view.cells_sequences;

-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
-        for (int j = 0; j < view.n_max_seq; j++) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
+        for (int j = 0; j < view.n_seq_max; j++) {
            if (cs_curr[j] < 0) { continue; }
            if (seqs.find(cs_curr[j]) == seqs.end()) {
                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
@@ -1835,11 +1845,11 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {

    c_curr = view.cells;
    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_seq_max) {
        if (i % row_size == 0) {
            printf("\n%5d: ", i);
        }
-        for (int j = 0; j < view.n_max_seq; j++) {
+        for (int j = 0; j < view.n_seq_max; j++) {
            if (cs_curr[j] >= 0) {
                const auto & it = seqs.find(cs_curr[j]);
                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
@@ -1867,3 +1877,16 @@ void llama_embd_normalize(const float * inp, float * out, int n) {
    }
 }

+float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n){
+    double sum  = 0.0;
+    double sum1 = 0.0;
+    double sum2 = 0.0;
+
+    for (int i = 0; i < n; i++) {
+        sum  += embd1[i] * embd2[i];
+        sum1 += embd1[i] * embd1[i];
+        sum2 += embd2[i] * embd2[i];
+    }
+
+    return sum / (sqrt(sum1) * sqrt(sum2));
+}
@@ -51,7 +51,8 @@ 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)
@@ -267,3 +268,4 @@ 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);
@@ -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(
@@ -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()

@@ -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);
@@ -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);
@@ -164,6 +164,7 @@ 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 +184,8 @@ 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 +210,7 @@ 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 +220,7 @@ 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 +300,13 @@ 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 +465,7 @@ 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 +485,7 @@ 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 +523,7 @@ 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 +545,7 @@ 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 +559,7 @@ 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 +583,7 @@ 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 +620,7 @@ 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 +644,7 @@ 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 +723,8 @@ 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 +738,8 @@ 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 +779,8 @@ 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 +978,9 @@ 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 +1120,32 @@ struct sql_printer : public printer {
 };

 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
+    //std::vector<llama_token> tokens(n_prompt, llama_token_bos(llama_get_model(ctx)));
+    //llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt, n_past, 0));
+    //GGML_UNUSED(n_batch);
+
    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);
-
    while (n_processed < n_prompt) {
        int n_tokens = std::min(n_prompt - n_processed, n_batch);
        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);

+    llama_token token = llama_token_bos(llama_get_model(ctx));
+
    for (int i = 0; i < n_gen; i++) {
        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_synchronize(ctx);
    }
 }

@@ -1203,7 +1234,8 @@ 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 +1251,7 @@ 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
 ```
@@ -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);

@@ -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()) {
@@ -2424,6 +2434,12 @@ 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 +2779,7 @@ 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 +3388,37 @@ 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 +3427,19 @@ 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 +3451,6 @@ 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"));
@@ -9,6 +9,7 @@ Feature: llama.cpp server
    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
@@ -33,6 +33,7 @@ def step_server_config(context, server_fqdn, server_port):

    context.model_alias = None
    context.n_batch = None
+    context.n_ubatch = None
    context.n_ctx = None
    context.n_ga = None
    context.n_ga_w = None
@@ -118,6 +119,10 @@ 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 +130,7 @@ 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)
@@ -278,6 +283,11 @@ 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 +947,9 @@ 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:
@@ -1029,6 +1042,8 @@ def start_server_background(context):
    ]
    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:
@@ -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;
@@ -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;
 };

@@ -543,13 +531,20 @@ 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];

@@ -577,19 +572,6 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
        }
    }

-    // 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 +634,7 @@ 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 +658,7 @@ 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 +693,16 @@ 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 +710,8 @@ 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 +728,30 @@ 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 +827,18 @@ 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 +852,6 @@ 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 +883,12 @@ 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 +900,6 @@ 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 +102,27 @@ extern "C" {

        // check if the backend supports an operation
        bool (*GGML_CALL supports_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
    //
@@ -221,29 +221,29 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");

    if (!size) {
        return;
    }

-    tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
+    buf->iface.set_tensor(buf, tensor, data, offset, size);
 }

 GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");

    if (!size) {
        return;
    }

-    tensor->buffer->iface.get_tensor(buf, tensor, data, offset, size);
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
 }

 void ggml_backend_synchronize(ggml_backend_t backend) {
@@ -255,18 +255,30 @@ void ggml_backend_synchronize(ggml_backend_t backend) {
 }

 ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend->iface.graph_plan_create != NULL);
+
    return backend->iface.graph_plan_create(backend, cgraph);
 }

 void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_free != NULL);
+
    backend->iface.graph_plan_free(backend, plan);
 }

 enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
+
    return backend->iface.graph_plan_compute(backend, plan);
 }

 enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
+    ggml_backend_synchronize(backend);
+    return err;
+}
+
+bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    return backend->iface.graph_compute(backend, cgraph);
 }

@@ -314,34 +326,68 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
    }
 }

-void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");

    if (src == dst) {
        return;
    }

-    if (ggml_backend_buft_supports_backend(src->buffer->buft, backend) && ggml_backend_buft_supports_backend(dst->buffer->buft, backend)) {
-        if (backend->iface.cpy_tensor_async != NULL) {
-            if (backend->iface.cpy_tensor_async(backend, src, dst)) {
-                return;
-            }
+    if (backend_dst->iface.cpy_tensor_async != NULL) {
+        if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
+            return;
        }
    }

-    size_t nbytes = ggml_nbytes(src);
+    // an async copy would normally happen after all the queued operations on both backends are completed
+    // sync src, set_async dst
    if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_tensor_set_async(backend, dst, src->data, 0, nbytes);
-    }
-    else {
+        ggml_backend_synchronize(backend_src);
+        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
+    } else {
+        ggml_backend_synchronize(backend_src);
        ggml_backend_tensor_copy(src, dst);
+        ggml_backend_synchronize(backend_dst);
    }
 }

+// events
+
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
+    if (backend->iface.event_new == NULL) {
+        return NULL;
+    }
+    return backend->iface.event_new(backend);
+}
+
+void ggml_backend_event_free(ggml_backend_event_t event) {
+    if (event == NULL) {
+        return;
+    }
+    event->backend->iface.event_free(event);
+}
+
+void ggml_backend_event_record(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_record != NULL);
+
+    event->backend->iface.event_record(event);
+}
+
+void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
+
+    event->backend->iface.event_synchronize(event);
+}
+
+void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend->iface.event_wait != NULL);
+
+    backend->iface.event_wait(backend, event);
+}

 // backend registry

-#define GGML_MAX_BACKENDS_REG 16
+#define GGML_REG_MAX_BACKENDS 16

 struct ggml_backend_reg {
    char name[128];
@@ -350,7 +396,7 @@ struct ggml_backend_reg {
    void * user_data;
 };

-static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
+static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
 static size_t ggml_backend_registry_count = 0;

 GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
@@ -395,7 +441,7 @@ GGML_CALL static void ggml_backend_registry_init(void) {
 }

 GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
-    GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
+    GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);

    size_t id = ggml_backend_registry_count;

@@ -746,8 +792,12 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);

    if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        free(cpu_ctx->work_data);
+        cpu_ctx->work_data = malloc(cplan.work_size);
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
        cpu_ctx->work_size = cplan.work_size;
    }
    cplan.work_data = cpu_ctx->work_data;
@@ -784,6 +834,11 @@ static struct ggml_backend_i cpu_backend_i = {
    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
    /* .supports_op             = */ ggml_backend_cpu_supports_op,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_cpu_guid(void) {
@@ -939,15 +994,27 @@ static bool ggml_is_view_op(enum ggml_op op) {

 // scheduler

-#define GGML_MAX_BACKENDS 16
-#define GGML_MAX_SPLITS 256
-#define GGML_MAX_SPLIT_INPUTS 16
+#ifndef GGML_SCHED_MAX_BACKENDS
+#define GGML_SCHED_MAX_BACKENDS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLITS
+#define GGML_SCHED_MAX_SPLITS 256
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
+#define GGML_SCHED_MAX_SPLIT_INPUTS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_COPIES
+#define GGML_SCHED_MAX_COPIES 4
+#endif

 struct ggml_backend_sched_split {
    int backend_id;
    int i_start;
    int i_end;
-    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
+    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
    int n_inputs;
    // graph view of this split
    struct ggml_cgraph graph;
@@ -955,45 +1022,53 @@ struct ggml_backend_sched_split {

 struct ggml_backend_sched {
    bool is_reset; // true if the scheduler has been reset since the last graph split
+    bool is_alloc;

    int n_backends;
-    ggml_backend_t backends[GGML_MAX_BACKENDS];
-    ggml_backend_buffer_type_t bufts[GGML_MAX_BACKENDS];

+    ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
+    ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
    ggml_gallocr_t galloc;

    // hash keys of the nodes in the graph
    struct ggml_hash_set    hash_set;
    // hash values
    int * tensor_backend_id;
-    struct ggml_tensor * (* tensor_copies)[GGML_MAX_BACKENDS];
+    struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];

-    int * node_backend_ids; // [n_nodes]
-    int n_nodes;
+    int * node_backend_ids; // [graph_size]
+    int * leaf_backend_ids; // [graph_size]

    // copy of the graph with modified inputs
    struct ggml_cgraph * graph;

-    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
+    // graph splits
+    struct ggml_backend_sched_split splits[GGML_SCHED_MAX_SPLITS];
    int n_splits;

+    // pipeline parallelism support
+    int n_copies;
+    int cur_copy;
+    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_graph_inputs;
+
    struct ggml_context * ctx;

    ggml_backend_sched_eval_callback callback_eval;
    void * callback_eval_user_data;

    // align context_buffer to GGML_MEM_ALIGN
-    #ifdef _MSC_VER
+#ifdef _MSC_VER
    __declspec(align(GGML_MEM_ALIGN))
-    #else
+#else
    __attribute__((aligned(GGML_MEM_ALIGN)))
-    #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
+#endif
+    char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
 };

-#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
-#define tensor_backend_id(node) sched->tensor_backend_id[hash_id(node)]
-#define tensor_backend(node) (tensor_backend_id(node) == -1 ? NULL : sched->backends[tensor_backend_id(node)])
+#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)]

 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
@@ -1005,7 +1080,8 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
    return -1;
 }

-static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) {
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor) {
+    ggml_backend_buffer_t buffer = tensor->buffer;
    if (buffer == NULL) {
        return -1;
    }
@@ -1016,12 +1092,16 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, gg
            return i;
        }
    }
-    GGML_ASSERT(false && "tensor buffer type not supported by any backend");
-    return -1; // silence warning
+
+    fprintf(stderr, "%s: error: no backend supports buffer type %s used in tensor %s\n",
+        __func__, ggml_backend_buffer_name(buffer), tensor->name);
+    GGML_ASSERT(false);
+
+    return -1;
 }

 #if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug only
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1035,19 +1115,28 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st

    // assign pre-allocated nodes to their backend
    // dst
-    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->buffer);
+    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor);
    if (cur_backend != -1) {
-        SET_CAUSE(node, "1.dst");
+        SET_CAUSE(tensor, "1.dst");
        return cur_backend;
    }
+
    // view_src
    if (tensor->view_src != NULL) {
-        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src->buffer);
+        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
        if (cur_backend != -1) {
-            SET_CAUSE(node, "1.vsrc");
+            SET_CAUSE(tensor, "1.vsrc");
            return cur_backend;
        }
    }
+
+    // input
+    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
+        cur_backend = sched->n_backends - 1; // last backend (assumed CPU)
+        SET_CAUSE(tensor, "1.inp");
+        return cur_backend;
+    }
+
    // assign nodes that use weights to the backend of the weights
    for (int i = 0; i < GGML_MAX_SRC; i++) {
        const struct ggml_tensor * src = tensor->src[i];
@@ -1055,9 +1144,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
            continue;
        }
        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src->buffer);
+            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src);
            // operations with weights are always run on the same backend as the weights
-            SET_CAUSE(node, "1.wgt%d", i);
+            SET_CAUSE(tensor, "1.wgt%d", i);
            return src_backend;
        }
    }
@@ -1093,7 +1182,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
        if (ggml_is_view_op(node->op)) {
            continue;
        }
-        ggml_backend_t tensor_backend = tensor_backend(node);
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
        for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -1101,7 +1190,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
            if (src == NULL) {
                continue;
            }
-            ggml_backend_t src_backend = tensor_backend(src);
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
            fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
        }
@@ -1118,6 +1207,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
 static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
    // reset splits
    sched->n_splits = 0;
+    sched->n_graph_inputs = 0;
    sched->is_reset = false;

    struct ggml_init_params params = {
@@ -1163,7 +1253,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
        }
    }
 #ifdef DEBUG_PASS1
-    fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif

    // pass 2: expand current backend assignments
@@ -1171,28 +1261,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
    // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
    // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops

-    // pass 2.1 expand gpu up
-    {
-        int cur_backend_id = -1;
-        for (int i = graph->n_nodes - 1; i >= 0; i--) {
-            struct ggml_tensor * node = graph->nodes[i];
-            if (ggml_is_view_op(node->op)) {
-                continue;
-            }
-            int tensor_backend_id = tensor_backend_id(node);
-            if (tensor_backend_id != -1) {
-                if (tensor_backend_id == sched->n_backends - 1) {
-                    // skip cpu (lowest prio backend)
-                    cur_backend_id = -1;
-                } else {
-                    cur_backend_id = tensor_backend_id;
-                }
-            } else {
-                tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.1");
-            }
-        }
-    }

    // pass 2.2 expand gpu down
    {
@@ -1217,7 +1285,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
        }
    }

-    // pass 2.3 expand rest up
+    // pass 2.1 expand gpu up
    {
        int cur_backend_id = -1;
        for (int i = graph->n_nodes - 1; i >= 0; i--) {
@@ -1227,14 +1295,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            }
            int tensor_backend_id = tensor_backend_id(node);
            if (tensor_backend_id != -1) {
-                cur_backend_id = tensor_backend_id;
+                if (tensor_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = tensor_backend_id;
+                }
            } else {
                tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.3");
+                SET_CAUSE(node, "2.1");
            }
        }
    }

+
    // pass 2.4 expand rest down
    {
        int cur_backend_id = -1;
@@ -1252,8 +1326,26 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            }
        }
    }
+        // pass 2.3 expand rest up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int tensor_backend_id = tensor_backend_id(node);
+            if (tensor_backend_id != -1) {
+                cur_backend_id = tensor_backend_id;
+            } else {
+                tensor_backend_id(node) = cur_backend_id;
+                SET_CAUSE(node, "2.3");
+            }
+        }
+    }
+
 #ifdef DEBUG_PASS2
-    fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif

    // pass 3: assign backends to remaining src from dst and view_src
@@ -1283,7 +1375,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
        }
    }
 #ifdef DEBUG_PASS3
-    fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif

    // pass 4: split graph, find tensors that need to be copied
@@ -1315,7 +1407,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            if (tensor_backend_id != cur_backend_id) {
                sched->splits[cur_split].i_end = i;
                cur_split++;
-                GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
+                GGML_ASSERT(cur_split < GGML_SCHED_MAX_SPLITS);
                sched->splits[cur_split].backend_id = tensor_backend_id;
                sched->splits[cur_split].i_start = i;
                sched->splits[cur_split].n_inputs = 0;
@@ -1328,25 +1420,57 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                if (src == NULL) {
                    continue;
                }
+
                int src_backend_id = tensor_backend_id(src);
                assert(src_backend_id != -1); // all inputs should be assigned by now
+
+                if (src->flags & GGML_TENSOR_FLAG_INPUT)  {
+                    size_t id = hash_id(src);
+                    if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
+                        ggml_backend_t backend = sched->backends[src_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy;
+                            if (c == sched->cur_copy) {
+                                tensor_copy = src; // use the original tensor as the current copy
+                            } else {
+                                tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                                ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            }
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = src_backend_id;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_graph_inputs = sched->n_graph_inputs++;
+                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        sched->graph_inputs[n_graph_inputs] = src;
+                    }
+                }
+
                if (src_backend_id != tensor_backend_id) {
                    // create a copy of the input in the split's backend
                    size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][cur_backend_id] == NULL) {
+                    if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
                        ggml_backend_t backend = sched->backends[cur_backend_id];
-                        struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                        ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-
-                        sched->tensor_copies[id][cur_backend_id] = tensor_copy;
-                        tensor_backend_id(tensor_copy) = cur_backend_id;
-                        SET_CAUSE(tensor_copy, "4.cpy");
-
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                            ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = cur_backend_id;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
                        int n_inputs = sched->splits[cur_split].n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
                        sched->splits[cur_split].inputs[n_inputs] = src;
                    }
-                    node->src[j] = sched->tensor_copies[id][cur_backend_id];
+                    node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
                }
            }
        }
@@ -1354,37 +1478,39 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
        sched->n_splits = cur_split + 1;
    }
 #ifdef DEBUG_PASS4
-    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif

 #ifndef NDEBUG
    // sanity check: all sources should have the same backend as the node
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
-        ggml_backend_t tensor_backend = tensor_backend(node);
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
        if (tensor_backend == NULL) {
            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
        }
-        if (node->view_src != NULL && tensor_backend != tensor_backend(node->view_src)) {
+        if (node->view_src != NULL && tensor_backend != ggml_backend_sched_get_tensor_backend(sched, node->view_src)) {
            fprintf(stderr, "!!!!!!! %s has backend %s, view_src %s has backend %s\n",
                node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
-                node->view_src->name, tensor_backend(node->view_src) ? ggml_backend_name(tensor_backend(node->view_src)) : "NULL");
+                node->view_src->name, ggml_backend_sched_get_tensor_backend(sched, node->view_src) ?
+                    ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, node->view_src)) : "NULL");
        }
        for (int j = 0; j < GGML_MAX_SRC; j++) {
            struct ggml_tensor * src = node->src[j];
            if (src == NULL) {
                continue;
            }
-            ggml_backend_t src_backend = tensor_backend(src);
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
            if (src_backend != tensor_backend /* && src_backend != NULL */) {
                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
                    node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
                    j, src->name, src_backend ? ggml_backend_name(src_backend) : "NULL");
            }
-            if (src->view_src != NULL && src_backend != tensor_backend(src->view_src)) {
+            if (src->view_src != NULL && src_backend != ggml_backend_sched_get_tensor_backend(sched, src->view_src)) {
                fprintf(stderr, "!!!!!!! [src] %s has backend %s, view_src %s has backend %s\n",
                    src->name, src_backend ? ggml_backend_name(src_backend) : "NULL",
-                    src->view_src->name, tensor_backend(src->view_src) ? ggml_backend_name(tensor_backend(src->view_src)) : "NULL");
+                    src->view_src->name, ggml_backend_sched_get_tensor_backend(sched, src->view_src) ?
+                        ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, src->view_src)) : "NULL");
            }
        }
    }
@@ -1392,18 +1518,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 #endif

    // create copies of the graph for each split
-    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
+    // TODO: avoid this copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS, false);
    for (int i = 0; i < sched->n_splits; i++) {
        struct ggml_backend_sched_split * split = &sched->splits[i];
        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);

+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
        for (int j = 0; j < split->n_inputs; j++) {
            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id];
+            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id][sched->cur_copy];

            // add a dependency to the input source so that it is not freed before the copy is done
            struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
+            input_dep->src[0] = input;
            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(input);
            graph_copy->nodes[graph_copy->n_nodes++] = input_dep;

@@ -1417,18 +1545,56 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
        }
    }
+
+    if (sched->n_copies > 1) {
+        // add input copies as leafs so that they are allocated first
+        for (int i = 0; i < sched->n_graph_inputs; i++) {
+            struct ggml_tensor * input = sched->graph_inputs[i];
+            size_t id = hash_id(input);
+            int backend_id = tensor_backend_id(input);
+            for (int c = 0; c < sched->n_copies; c++) {
+                struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+            }
+        }
+
+        for (int i = 0; i < sched->n_splits; i++) {
+            struct ggml_backend_sched_split * split = &sched->splits[i];
+            int backend_id = split->backend_id;
+            for (int j = 0; j < split->n_inputs; j++) {
+                struct ggml_tensor * input = split->inputs[j];
+                size_t id = hash_id(input);
+                for (int c = 0; c < sched->n_copies; c++) {
+                    struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                    sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+                }
+            }
+        }
+    }
+
+    // add leafs from the original graph
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        graph_copy->leafs[graph_copy->n_leafs++] = leaf;
+    }
+
    sched->graph = graph_copy;
 }

 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
-    // ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids);
+    // allocate graph
    if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+        // the re-allocation may cause the split inputs to be moved to a different address
+        ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_sched: failed to allocate graph, reserving\n");
+        fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__);
 #endif
-        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids);
+        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
        if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
-            fprintf(stderr, "ggml_backend_sched: failed to allocate graph\n");
+            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
            return false;
        }
    }
@@ -1437,9 +1603,6 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 }

 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
-    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
-    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
-
    struct ggml_backend_sched_split * splits = sched->splits;

    for (int i = 0; i < sched->n_splits; i++) {
@@ -1448,34 +1611,36 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
        ggml_backend_t split_backend = sched->backends[split_backend_id];

        // copy the input tensors to the split backend
-        uint64_t copy_start_us = ggml_time_us();
        for (int j = 0; j < split->n_inputs; j++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id];
+            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];

-            GGML_ASSERT(input->buffer != NULL);
-            GGML_ASSERT(input_cpy->buffer != NULL);
+            if (input->flags & GGML_TENSOR_FLAG_INPUT) {
+                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy(input, input_cpy);
+            } else {
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                    ggml_backend_synchronize(input_backend);
+                }

-            ggml_backend_tensor_copy_async(split_backend, input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+            }
        }
-        //ggml_backend_synchronize(split_backend); // necessary to measure copy time
-        int64_t copy_end_us = ggml_time_us();
-        copy_us[split_backend_id] += copy_end_us - copy_start_us;

-#if 0
-        char split_filename[GGML_MAX_NAME];
-        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
-        ggml_graph_dump_dot(split->graph, NULL, split_filename);
-#endif
-
-
-        uint64_t compute_start_us = ggml_time_us();
        if (!sched->callback_eval) {
-            enum ggml_status ec = ggml_backend_graph_compute(split_backend, &split->graph);
+            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
            if (ec != GGML_STATUS_SUCCESS) {
                return ec;
            }
-            //ggml_backend_synchronize(split_backend); // necessary to measure compute time
        } else {
            // similar to ggml_backend_compare_graph_backend
            for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
@@ -1494,11 +1659,14 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s

                struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);

-                enum ggml_status ec = ggml_backend_graph_compute(split_backend, &gv);
+                enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
                if (ec != GGML_STATUS_SUCCESS) {
                    return ec;
                }

+                // TODO: pass backend to the callback, then the user can decide if they want to synchronize
+                ggml_backend_synchronize(split_backend);
+
                if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
                    break;
                }
@@ -1506,39 +1674,54 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                j0 = j1;
            }
        }
-        uint64_t compute_end_us = ggml_time_us();
-        compute_us[split_backend_id] += compute_end_us - compute_start_us;
-    }

-#if 0
-    // per-backend timings
-    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (copy_us[i] > 0 || compute_us[i] > 0) {
-            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
+        // record the event of this copy
+        if (split->n_inputs > 0) {
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
+            }
        }
    }
-#endif
+
+    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;

    return GGML_STATUS_SUCCESS;
 }

-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_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_ASSERT(n_backends > 0);
-    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
+    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
+    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU

    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);

    // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), graph_size);
+    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), graph_size);

    sched->n_backends = n_backends;
-    for (int i = 0; i < n_backends; i++) {
-        sched->backends[i] = backends[i];
-        sched->bufts[i] = bufts ? bufts[i] : ggml_backend_get_default_buffer_type(backends[i]);
+
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+
+    GGML_ASSERT(sched->n_copies <= GGML_SCHED_MAX_COPIES);
+
+    for (int b = 0; b < n_backends; b++) {
+        sched->backends[b] = backends[b];
+        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
+        GGML_ASSERT(ggml_backend_buft_supports_backend(sched->bufts[b], backends[b]));
+        if (sched->n_copies > 1) {
+            for (int c = 0; c < sched->n_copies; c++) {
+                sched->events[b][c] = ggml_backend_event_new(backends[b]);
+            }
+        }
    }

    sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
@@ -1552,12 +1735,18 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
    if (sched == NULL) {
        return;
    }
+    for (int b = 0; b < sched->n_backends; b++) {
+        for (int c = 0; c < sched->n_copies; c++) {
+            ggml_backend_event_free(sched->events[b][c]);
+        }
+    }
    ggml_gallocr_free(sched->galloc);
    ggml_free(sched->ctx);
    free(sched->hash_set.keys);
    free(sched->tensor_backend_id);
    free(sched->tensor_copies);
    free(sched->node_backend_ids);
+    free(sched->leaf_backend_ids);
    free(sched);
 }

@@ -1569,34 +1758,63 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
    memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);

    sched->is_reset = true;
+    sched->is_alloc = false;
 }

 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
    ggml_backend_sched_split_graph(sched, measure_graph);

-    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids)) {
+    // TODO: extract this to a separate function
+    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
        return false;
    }

    ggml_backend_sched_reset(sched);
+    ggml_backend_sched_synchronize(sched);
+
+    return true;
+}
+
+bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
+
+    ggml_backend_sched_split_graph(sched, graph);
+
+    if (!ggml_backend_sched_alloc_splits(sched)) {
+        return false;
+    }
+
+    sched->is_alloc = true;
+
    return true;
 }

 enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+    enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
+    ggml_backend_sched_synchronize(sched);
+    return err;
+}

-    if (!sched->is_reset) {
+enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    if (!sched->is_reset && !sched->is_alloc) {
        ggml_backend_sched_reset(sched);
    }

-    ggml_backend_sched_split_graph(sched, graph);
-    if (!ggml_backend_sched_alloc_splits(sched)) {
-        return GGML_STATUS_ALLOC_FAILED;
+    if (!sched->is_alloc) {
+        if (!ggml_backend_sched_alloc_graph(sched, graph)) {
+            return GGML_STATUS_ALLOC_FAILED;
+        }
    }

    return ggml_backend_sched_compute_splits(sched);
 }

+void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_backend_synchronize(sched->backends[i]);
+    }
+}
+
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
    sched->callback_eval = callback;
    sched->callback_eval_user_data = user_data;
@@ -1606,19 +1824,24 @@ int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
    return sched->n_splits;
 }

+int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    return sched->n_copies;
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
    int backend_index = ggml_backend_sched_backend_id(sched, backend);
    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
    return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
 }

-void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
    int backend_index = ggml_backend_sched_backend_id(sched, backend);
    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
    tensor_backend_id(node) = backend_index;
 }

-ggml_backend_t ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
    int backend_index = tensor_backend_id(node);
    if (backend_index == -1) {
        return NULL;
@@ -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;

@@ -72,11 +73,24 @@ extern "C" {
    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);

+    GGML_API bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API bool ggml_backend_supports_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
@@ -2,7 +2,13 @@
 #include "ggml.h"
 #include "ggml-backend-impl.h"

+#if defined(GGML_USE_HIPBLAS)
+#define GGML_COMMON_DECL_HIP
+#define GGML_COMMON_IMPL_HIP
+#else
+#define GGML_COMMON_DECL_CUDA
 #define GGML_COMMON_IMPL_CUDA
+#endif
 #include "ggml-common.h"

 #include <algorithm>
@@ -66,6 +72,7 @@
 #define cudaEventCreateWithFlags hipEventCreateWithFlags
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
 #define cudaEvent_t hipEvent_t
 #define cudaEventDestroy hipEventDestroy
 #define cudaFree hipFree
@@ -75,6 +82,7 @@
 #define cudaGetDeviceProperties hipGetDeviceProperties
 #define cudaGetErrorString hipGetErrorString
 #define cudaGetLastError hipGetLastError
+#define cudaLaunchHostFunc hipLaunchHostFunc
 #ifdef GGML_HIP_UMA
 #define cudaMalloc hipMallocManaged
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
@@ -98,6 +106,7 @@
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
 #define cudaStreamFireAndForget hipStreamFireAndForget
 #define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
 #define cudaStreamSynchronize hipStreamSynchronize
 #define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
 #define cudaStream_t hipStream_t
@@ -359,66 +368,6 @@ typedef void (*ggml_cuda_op_flatten_t)(
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t 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 {
-    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 {
-    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 {
-    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 {
-    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 {
-    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 {
-    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_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
 typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
 typedef void (*load_tiles_cuda_t)(
@@ -428,174 +377,6 @@ typedef float (*vec_dot_q_mul_mat_cuda_t)(
    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
    const int * __restrict__ y_qs, const 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 {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    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 {
-    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
-    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 {
-    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(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    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 {
-    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 {
-    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 {
-    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
-    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 {
-    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 {
-    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");
-
-// 2.5625 bpw quants
-#define QR2_S 8
-#define QI2_S (QK_K / (4*QR2_S))
-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;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
-#define QR3_XXS 8
-#define QI3_XXS (QK_K / (4*QR3_XXS))
-typedef struct {
-    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 {
-    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 {
-    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_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
-
-#define QK4_NL 32
-#define QR4_NL 2
-#define QI4_NL (QK4_NL / (4*QR4_NL))
-typedef struct {
-    half 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
-#define QR4_XS QR4_NL
-#define QI4_XS QI4_NL
-#else
-// QR4_XS = 8 is very slightly faster than QR4_XS = 4
-#define QR4_XS 8
-#define QI4_XS (QK_K / (4*QR4_XS))
-typedef struct {
-    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_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
 #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

@@ -1722,22 +1503,15 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
    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 float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 0xf) + 1);
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int grid32[2]; const int8_t * q = (const int8_t *)grid32;
-    grid32[0] = *((const int *)(iq1s_grid_gpu + (x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8))));
-    grid32[1] = __vsub4((grid32[0] >>  4) & 0x0f0f0f0f, 0x01010101);
-    grid32[0] = __vsub4(grid32[0] & 0x0f0f0f0f, 0x01010101);
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
    for (int j = 0; j < 8; ++j) {
-        y[j] = d * q[j];
+        y[j] = d * (q[j] + delta);
    }
-#else
-    const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)));
-    for (int j = 0; j < 4; ++j) {
-        y[j+0] = d * ((grid[j] & 0xf) - 1);
-        y[j+4] = d * ((grid[j] >>  4) - 1);
-    }
-#endif
 #else
    assert(false);
 #endif
@@ -3577,7 +3351,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR2_K; ++ i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
    }

    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@@ -3699,7 +3473,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR3_K; ++i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
    }

    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@@ -3868,7 +3642,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(

    for (int i = 0; i < QR4_K; ++i) {
        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = __low2half(bq8i->ds);
+        d8[i] = __low2float(bq8i->ds);

        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
        u[2*i+0] = q8[0];
@@ -4233,7 +4007,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 #pragma unroll
    for (int i = 0; i < QR6_K; ++i) {
        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
+        d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
    }

    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@@ -4560,22 +4334,25 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
    const int * q8 = (const int *)bq8_1[ib32].qs;
    for (int l = 0; l < 4; ++l) {
        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
-        int grid0 = __vsub4(grid[0] & 0x0f0f0f0f, 0x01010101);
-        int grid1 = __vsub4((grid[0] >> 4) & 0x0f0f0f0f, 0x01010101);
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
        sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi));
    }
 #else
-    const int8_t   * q8 = bq8_1[ib32].qs;
+    const int8_t * q8 = bq8_1[ib32].qs;
    for (int l = 0; l < 4; ++l) {
        const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
        for (int j = 0; j < 4; ++j) {
-            sumi += q8[j] * ((grid[j] & 0xf) - 1) + q8[j+4] * ((grid[j] >>  4) - 1);
+            sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
        }
        q8 += 8;
    }
 #endif
-    const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
-    return d * sumi * (2*(bq1->qh[ib32] >> 12) + 1);
+    const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA;
+    const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1);
+    const float d = d1q * __low2float (bq8_1[ib32].ds);
+    const float m = d1q * __high2float(bq8_1[ib32].ds);
+    return d * sumi + m * delta;
 #else
    assert(false);
    return 0.f;
@@ -4767,7 +4544,7 @@ static __device__ __forceinline__ void mul_mat_q(
                    *dsi_dst = *dsi_src;
                } else {
                    float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = __low2half(*dsi_src);
+                    *dfi_dst = __low2float(*dsi_src);
                }
            }

@@ -10867,8 +10644,20 @@ GGML_CALL void ggml_cuda_get_device_description(int device, char * description,
 #define UNUSED GGML_UNUSED

 struct ggml_backend_cuda_context {
+    explicit ggml_backend_cuda_context(int device) :
+        device(device),
+        name(GGML_CUDA_NAME + std::to_string(device)) {
+    }
+
+    ~ggml_backend_cuda_context() {
+        if (copy_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(copy_event));
+        }
+    }
+
    int device;
    std::string name;
+    cudaEvent_t copy_event = nullptr;
 };

 // cuda buffer
@@ -10958,9 +10747,8 @@ GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }

 GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
@@ -10969,26 +10757,25 @@ GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

    ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }

 GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
    if (ggml_backend_buffer_is_cuda(src->buffer)) {
        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
-        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
-
-        ggml_cuda_set_device(src_ctx->device);
-        CUDA_CHECK(cudaDeviceSynchronize());
-        ggml_cuda_set_device(dst_ctx->device);
-        CUDA_CHECK(cudaDeviceSynchronize());
-        CUDA_CHECK(cudaMemcpy((char *)dst->data, (const char *)src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice));
-        CUDA_CHECK(cudaDeviceSynchronize());
-
+        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
+        if (src_ctx->device == dst_ctx->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
+        } else {
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread));
+        }
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
        return true;
    }
    return false;
+
+    UNUSED(buffer);
 }

 GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
@@ -11233,7 +11020,11 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buf
        }

        const char * buf_host = (const char *)data + offset_split;
-        CUDA_CHECK(cudaMemcpy(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice));
+        CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
    }
 }

@@ -11267,7 +11058,11 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buf
        }

        char * buf_host = (char *)data + offset_split;
-        CUDA_CHECK(cudaMemcpy(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost));
+        CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < g_device_count; ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
    }
 }

@@ -11446,6 +11241,10 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
    return &ggml_backend_cuda_buffer_type_host;
 }

+//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) {
+//    return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+//}
+
 // backend

 GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
@@ -11469,8 +11268,9 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer

 GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);

    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
@@ -11478,22 +11278,61 @@ GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend,

 GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;

-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);

    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
 }

-GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
-    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));

-    if (dst->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && ggml_backend_buffer_is_cuda(src->buffer)) {
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx->device][0]));
-        return true;
+    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    if (!ggml_backend_buffer_is_cuda(src->buffer)) {
+        return false;
    }

-    return false;
+    if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
+        return false;
+    }
+
+    // device -> device
+    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
+    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
+
+    if (backend_src != backend_dst) {
+        ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+        ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+        GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
+
+        if (!cuda_ctx_src->copy_event) {
+            ggml_cuda_set_device(cuda_ctx_src->device);
+            CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
+        }
+
+        // copy on src stream
+        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0]));
+        } else {
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), g_cudaStreams[cuda_ctx_src->device][0]));
+        }
+
+        // record event on src stream
+        CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, g_cudaStreams[cuda_ctx_src->device][0]));
+
+        // wait on dst stream for the copy to complete
+        CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx_dst->device][0], cuda_ctx_src->copy_event, 0));
+    } else {
+        // src and dst are on the same backend
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx_dst->device][0]));
+    }
+    return true;
 }

 GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
@@ -11670,6 +11509,52 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
    UNUSED(backend);
 }

+static ggml_backend_event_t ggml_backend_cuda_event_new(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    ggml_cuda_set_device(cuda_ctx->device);
+
+    cudaEvent_t event;
+    CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
+
+    return new ggml_backend_event {
+        /* .backend = */ backend,
+        /* .context = */ event,
+    };
+}
+
+static void ggml_backend_cuda_event_free(ggml_backend_event_t event) {
+    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
+
+    delete event;
+}
+
+static void ggml_backend_cuda_event_record(ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)event->backend->context;
+
+    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, g_cudaStreams[cuda_ctx->device][0]));
+}
+
+static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    if (ggml_backend_is_cuda(event->backend)) {
+        CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[cuda_ctx->device][0], (cudaEvent_t)event->context, 0));
+    } else {
+        // untested
+        auto wait_fn = [](void * user_data) {
+            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
+            ggml_backend_event_synchronize(event);
+        };
+
+        CUDA_CHECK(cudaLaunchHostFunc(g_cudaStreams[cuda_ctx->device][0], wait_fn, event));
+    }
+}
+
+static void ggml_backend_cuda_event_synchronize(ggml_backend_event_t event) {
+    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+}
+
 static ggml_backend_i ggml_backend_cuda_interface = {
    /* .get_name                = */ ggml_backend_cuda_name,
    /* .free                    = */ ggml_backend_cuda_free,
@@ -11683,6 +11568,11 @@ static ggml_backend_i ggml_backend_cuda_interface = {
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
    /* .supports_op             = */ ggml_backend_cuda_supports_op,
+    /* .event_new               = */ ggml_backend_cuda_event_new,
+    /* .event_free              = */ ggml_backend_cuda_event_free,
+    /* .event_record            = */ ggml_backend_cuda_event_record,
+    /* .event_wait              = */ ggml_backend_cuda_event_wait,
+    /* .event_synchronize       = */ ggml_backend_cuda_event_synchronize,
 };

 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -11701,10 +11591,11 @@ GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
    // not strictly necessary, but it may reduce the overhead of the first graph_compute
    ggml_cuda_set_main_device(device);

-    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context {
-        /* .device = */ device,
-        /* .name   = */ GGML_CUDA_NAME + std::to_string(device),
-    };
+    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
+    if (ctx == nullptr) {
+        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    }

    ggml_backend_t cuda_backend = new ggml_backend {
        /* .guid      = */ ggml_backend_cuda_guid(),
@@ -1951,6 +1951,11 @@ 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,
+    /* .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 +2837,11 @@ 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,
+    /* .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
@@ -202,24 +202,29 @@ 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>
@@ -3144,6 +3149,7 @@ namespace dpct

 } // COPY from DPCT head files

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

@@ -3312,66 +3318,6 @@ 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 +3334,6 @@ 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

@@ -4891,10 +4706,9 @@ 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 +4717,15 @@ 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 +7621,27 @@ 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 +8461,7 @@ 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 +10223,7 @@ 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 +10231,7 @@ 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 +10971,11 @@ 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(
@@ -17432,13 +17249,18 @@ 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,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -5693,6 +5693,11 @@ 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,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };

 static ggml_guid_t ggml_backend_vk_guid() {
@@ -11560,8 +11560,6 @@ 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 +11567,7 @@ 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 +11577,25 @@ 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 +11606,6 @@ 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 +11613,32 @@ 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 +11649,6 @@ 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 +11656,32 @@ 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));
    }
 }

@@ -17796,7 +17814,7 @@ 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 +17895,12 @@ 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 +17908,6 @@ 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 +18125,7 @@ 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 +18135,7 @@ 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 +18183,7 @@ 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 +18248,7 @@ 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 +363,9 @@ 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_COUNT,
    };

@@ -383,20 +383,20 @@ 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"
@@ -661,6 +662,9 @@ class GGMLQuantizationType(IntEnum):
    IQ3_S   = 21
    IQ2_S   = 22
    IQ4_XS  = 23
+    I8      = 24
+    I16     = 25
+    I32     = 26


 class GGUFEndian(IntEnum):
@@ -727,6 +731,9 @@ 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),
 }


@@ -746,6 +753,7 @@ 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
@@ -248,6 +248,15 @@ class GGUFReader:
            elif ggml_type == GGMLQuantizationType.F16:
                item_count = n_elems
                item_type = np.float16
+            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
            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 +204,18 @@ 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.float32:
+                dtype = GGMLQuantizationType.F32
+            elif tensor_dtype == np.float16:
+                dtype = GGMLQuantizationType.F16
+            elif tensor_dtype == np.int8:
+                dtype = GGMLQuantizationType.I8
+            elif tensor_dtype == np.int16:
+                dtype = GGMLQuantizationType.I16
+            elif tensor_dtype == np.int32:
+                dtype = GGMLQuantizationType.I32
+            else:
+                raise ValueError("Only F32, F16, I8, I16, I32 tensors are supported for now")
        else:
            dtype = raw_dtype
        self.ti_data += self._pack("I", dtype)
@@ -313,6 +321,9 @@ 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)

@@ -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);
@@ -456,7 +459,7 @@ 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 +479,12 @@ 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 +653,11 @@ 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 +716,7 @@ 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 +725,7 @@ 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) {
Author	SHA1	Message	Date
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