metal : improve FA + improve MoE (#12612)

* ggml : FA with different K, V head sizes (CPU)

ggml-ci

* metal : add FA with HS=192

* metal : extend FA to support different K and V head sizes

ggml-ci

* metal : add FA vector kernels for heads K 192 and V 128

ggml-ci

* ggml : restrict op on other backends to equal head sizes

ggml-ci

* metal : optimize FA-vec kernel

ggml-ci

* metal : FA remove mq registers

* metal : improve MoE mul_mat_id condition

ggml-ci

* metal : fix comments + remove unnecessary addition

ggml-ci

* metal : avoid too much shared memory usage with mul_mat_id

ggml-ci

.github/workflows/build.yml

@@ -676,6 +676,35 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
+  macOS-latest-cmake-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
   macOS-latest-swift:
     runs-on: macos-latest
 
@@ -774,7 +803,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.3.261.1
+      VULKAN_VERSION: 1.4.304.1
 
     strategy:
       matrix:
@@ -1379,7 +1408,7 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          zip -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
+          zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

CMakeLists.txt

@@ -29,6 +29,8 @@ else()
     set(LLAMA_STANDALONE OFF)
 endif()
 
+option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
+
 if (EMSCRIPTEN)
     set(BUILD_SHARED_LIBS_DEFAULT OFF)
 
@@ -145,7 +147,13 @@ endif()
 # 3rd-party
 #
 
-if (NOT TARGET ggml)
+if (LLAMA_USE_SYSTEM_GGML)
+    message(STATUS "Using system-provided libggml, skipping ggml build")
+    find_package(ggml REQUIRED)
+    add_library(ggml ALIAS ggml::ggml)
+endif()
+
+if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()

build-xcframework.sh

@@ -432,8 +432,8 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos --config Release -- -quiet
 
@@ -445,8 +445,8 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
 

ci/README.md

@@ -26,4 +26,43 @@ GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 # with SYCL support
 source /opt/intel/oneapi/setvars.sh
 GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with MUSA support
+GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```
+
+## Running MUSA CI in a Docker Container
+
+Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
+
+### 1. Create a local directory to store cached models, configuration files and venv:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-cache
+```
+
+### 2. Create a local directory to store CI run results:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-results
+```
+
+### 3. Start a Docker container and run the CI:
+
+```bash
+docker run --privileged -it \
+    -v $HOME/llama.cpp/ci-cache:/ci-cache \
+    -v $HOME/llama.cpp/ci-results:/ci-results \
+    -v $PWD:/ws -w /ws \
+    mthreads/musa:rc3.1.1-devel-ubuntu22.04
+```
+
+Inside the container, execute the following commands:
+
+```bash
+apt update -y && apt install -y bc cmake git python3.10-venv time unzip wget
+git config --global --add safe.directory /ws
+GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
+```
+
+This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.

ci/run.sh

@@ -16,6 +16,9 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with MUSA support
+# GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -52,13 +55,22 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
-
+    # Use only main GPU
+    export ONEAPI_DEVICE_SELECTOR="level_zero:0"
+    # Enable sysman for correct memory reporting
+    export ZES_ENABLE_SYSMAN=1
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
+
+if [ ! -z ${GG_BUILD_MUSA} ]; then
+    # Use qy1 by default (MTT S80)
+    MUSA_ARCH=${MUSA_ARCH:-21}
+    CMAKE_EXTRA="-DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
+fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -808,7 +820,7 @@ export LLAMA_LOG_PREFIX=1
 export LLAMA_LOG_TIMESTAMPS=1
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
+    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models
     rm -rf ${SRC}/models-mnt
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
@@ -826,8 +838,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-
-test $ret -eq 0 && gg_run ctest_debug
+if [ -z ${GG_BUILD_SYCL} ]; then
+    # SYCL build breaks with debug build flags
+    test $ret -eq 0 && gg_run ctest_debug
+fi
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -835,7 +849,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        test $ret -eq 0 && gg_run test_scripts_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run test_scripts_debug
+        fi
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -846,7 +862,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
             test $ret -eq 0 && gg_run pythia_2_8b
             #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        test $ret -eq 0 && gg_run ctest_with_model_debug
+        if [ -z ${GG_BUILD_SYCL} ]; then
+            test $ret -eq 0 && gg_run ctest_with_model_debug
+        fi
         test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi

cmake/common.cmake

@@ -1,3 +1,5 @@
+include("ggml/cmake/common.cmake")
+
 function(llama_add_compile_flags)
     if (LLAMA_FATAL_WARNINGS)
         if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")

common/CMakeLists.txt

@@ -114,8 +114,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.6.12:
-        GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
+        # v0.7.10:
+        GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/arg.cpp

@@ -764,7 +764,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CTX_SIZE"));
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
-        string_format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict),
+        string_format(
+            ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL
+                ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)"
+                : "number of tokens to predict (default: %d, -1 = infinity)",
+            params.n_predict),
         [](common_params & params, int value) {
             params.n_predict = value;
         }
@@ -849,6 +853,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_excludes({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-sysf", "--system-prompt-file"}, "FNAME",
+        "a file containing the system prompt (default: none)",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
+                params.system_prompt.pop_back();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--in-file"}, "FNAME",
         "an input file (repeat to specify multiple files)",
@@ -1871,7 +1889,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -1961,7 +1979,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
-        string_format("ip address to listen (default: %s)", params.hostname.c_str()),
+        string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
         [](common_params & params, const std::string & value) {
             params.hostname = value;
         }

common/common.cpp

@@ -955,8 +955,8 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
@@ -1033,6 +1033,8 @@ struct common_init_result common_init_from_params(common_params & params) {
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
+        llama_set_warmup(lctx, true);
+
         std::vector<llama_token> tmp;
         llama_token bos = llama_vocab_bos(vocab);
         llama_token eos = llama_vocab_eos(vocab);
@@ -1060,9 +1062,10 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
+        llama_set_warmup(lctx, false);
     }
 
     iparams.model.reset(model);

common/llguidance.cpp

@@ -11,25 +11,24 @@ struct llama_sampler_llg {
     std::string         grammar_kind;
     std::string         grammar_data;
     LlgTokenizer *      tokenizer;
-    LlgConstraint *     grammar;
-    LlgMaskResult       llg_res;
-    bool                has_llg_res;
+    LlgMatcher *        grammar;
 };
 
-static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
-                                             const char * grammar_data) {
+static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
+                                          const char * grammar_data) {
     LlgConstraintInit cinit;
     llg_constraint_init_set_defaults(&cinit, tokenizer);
     const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
     if (log_level && *log_level) {
         cinit.log_stderr_level = atoi(log_level);
     }
-    auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
-    if (llg_get_error(c)) {
-        LOG_ERR("llg error: %s\n", llg_get_error(c));
-        llg_free_constraint(c);
+    auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data);
+    if (llg_matcher_get_error(c)) {
+        LOG_ERR("llg error: %s\n", llg_matcher_get_error(c));
+        llg_free_matcher(c);
         return nullptr;
     }
+
     return c;
 }
 
@@ -40,39 +39,29 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
 static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        LlgCommitResult res;
-        llg_commit_token(ctx->grammar, token, &res);
-        ctx->has_llg_res = false;
+        llg_matcher_consume_token(ctx->grammar, token);
     }
 }
 
 static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
     if (ctx->grammar) {
-        if (!ctx->has_llg_res) {
-            if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
-                ctx->has_llg_res = true;
+        const uint32_t * mask = llg_matcher_get_mask(ctx->grammar);
+        if (mask == nullptr) {
+            if (llg_matcher_compute_mask(ctx->grammar) == 0) {
+                mask = llg_matcher_get_mask(ctx->grammar);
             } else {
-                LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
-                llg_free_constraint(ctx->grammar);
+                LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar));
+                llg_free_matcher(ctx->grammar);
                 ctx->grammar = nullptr;
+                return;
             }
         }
-        if (ctx->has_llg_res) {
-            if (ctx->llg_res.is_stop) {
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
-            } else {
-                const uint32_t * mask = ctx->llg_res.sample_mask;
-                for (size_t i = 0; i < cur_p->size; ++i) {
-                    auto token = cur_p->data[i].id;
-                    if ((mask[token / 32] & (1 << (token % 32))) == 0) {
-                        cur_p->data[i].logit = -INFINITY;
-                    }
-                }
+
+        for (size_t i = 0; i < cur_p->size; ++i) {
+            auto token = cur_p->data[i].id;
+            if ((mask[token / 32] & (1 << (token % 32))) == 0) {
+                cur_p->data[i].logit = -INFINITY;
             }
         }
     }
@@ -80,14 +69,9 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array
 
 static void llama_sampler_llg_reset(llama_sampler * smpl) {
     auto * ctx = (llama_sampler_llg *) smpl->ctx;
-    if (!ctx->grammar) {
-        return;
+    if (ctx->grammar) {
+        llg_matcher_reset(ctx->grammar);
     }
-
-    auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
-    llg_free_constraint(ctx->grammar);
-    ctx->grammar     = grammar_new;
-    ctx->has_llg_res = false;
 }
 
 static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
@@ -102,7 +86,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
         if (ctx->grammar) {
             result_ctx->grammar_kind = ctx->grammar_kind;
             result_ctx->grammar_data = ctx->grammar_data;
-            result_ctx->grammar      = llg_clone_constraint(ctx->grammar);
+            result_ctx->grammar      = llg_clone_matcher(ctx->grammar);
             result_ctx->tokenizer    = llg_clone_tokenizer(ctx->tokenizer);
         }
     }
@@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) {
     const auto * ctx = (llama_sampler_llg *) smpl->ctx;
 
     if (ctx->grammar) {
-        llg_free_constraint(ctx->grammar);
+        llg_free_matcher(ctx->grammar);
         llg_free_tokenizer(ctx->tokenizer);
     }
 
@@ -239,9 +223,11 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ grammar_data,
             /* .tokenizer    = */ tokenizer,
             /* .grammar      = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
+        if (ctx->grammar) {
+            GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 ==
+                        llg_matcher_get_mask_byte_size(ctx->grammar));
+        }
     } else {
         *ctx = {
             /* .vocab        = */ vocab,
@@ -249,15 +235,12 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g
             /* .grammar_data = */ {},
             /* .tokenizer    = */ nullptr,
             /* .grammar      = */ nullptr,
-            /* .llg_res      = */ {},
-            /* .has_llg_res  = */ false,
         };
     }
 
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_llg_i,
-        /* .ctx   = */ ctx
-    );
+        /* .ctx   = */ ctx);
 }
 
 #else

common/sampling.cpp

@@ -208,6 +208,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                                                         trigger_patterns_c.data(), trigger_patterns_c.size(),
                                                         trigger_tokens.data(), trigger_tokens.size())
              :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
     }
 
     auto * result = new common_sampler {

common/speculative.cpp

@@ -173,7 +173,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         prompt.clear();
     } else {
@@ -192,14 +192,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -180,7 +180,8 @@ class Model:
             extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
             missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
             if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                 f"Missing tensors: {missing}")
             else:
                 raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
                                  f"Missing tensors: {missing}\n"
@@ -528,6 +529,8 @@ class Model:
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
 
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
@@ -537,13 +540,13 @@ class Model:
                 if token in added_vocab:
                     # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
+                    if not added_tokens_decoder[i].normalized:
                         previous_token = token
                         token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
-                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
                         # NOTE: this was added for Gemma.
@@ -702,6 +705,9 @@ class Model:
         if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
             # ref: https://huggingface.co/Xenova/gpt-4o
             res = "gpt-4o"
+        if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f":
+            # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k
+            res = "superbpe"
 
         if res is None:
             logger.warning("\n")
@@ -908,6 +914,40 @@ class Model:
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_rwkv_world(self):
+        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
+        vocab_size = self.hparams.get("vocab_size", 65536)
+
+        tokens: list[bytes] = ['<s>'.encode("utf-8")]
+        toktypes: list[int] = [gguf.TokenType.CONTROL]
+
+        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
+            lines = f.readlines()
+            for line in lines:
+                parts = line.split(' ')
+                assert len(parts) >= 3
+                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
+                token = token.encode("utf-8") if isinstance(token, str) else token
+                assert isinstance(token, bytes)
+                assert len(token) == token_len
+                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
+                tokens.append(token_text.encode("utf-8"))
+                toktypes.append(gguf.TokenType.NORMAL)
+        remainder = vocab_size - len(tokens)
+        assert remainder >= 0
+        for i in range(len(tokens), vocab_size):
+            tokens.append(f"[PAD{i}]".encode("utf-8"))
+            toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("rwkv")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int):
         tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf"
         logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
@@ -1065,13 +1105,6 @@ class BloomModel(Model):
 
         tensors.append((self.map_tensor_name(name), data_torch))
 
-        if name == "word_embeddings.weight":
-            assert self.tensor_names is not None
-
-            # TODO: tie them at runtime, don't duplicate in the model file
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -1713,6 +1746,25 @@ class LlamaModel(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("Mistral3ForConditionalGeneration")
+class Mistral3Model(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+
+    # we need to merge the text_config into the root level of hparams
+    def __init__(self, *args, **kwargs):
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
+        if "text_config" in hparams:
+            hparams = {**hparams, **hparams["text_config"]}
+            kwargs["hparams"] = hparams
+        super().__init__(*args, **kwargs)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        name = name.replace("language_model.", "")
+        if "multi_modal_projector" in name or "vision_tower" in name:
+            return []
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("DeciLMForCausalLM")
 class DeciModel(Model):
     model_arch = gguf.MODEL_ARCH.DECI
@@ -2217,7 +2269,7 @@ class Qwen2Model(Model):
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
 
-@Model.register("Qwen2VLForConditionalGeneration")
+@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN2VL
 
@@ -2370,10 +2422,6 @@ class GPT2Model(Model):
 
         tensors.append((new_name, data_torch))
 
-        # note: GPT2 output is tied to (same as) wte in original model
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -2703,21 +2751,26 @@ class CodeShellModel(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
+    _has_tok_embd = False
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
         new_name = self.map_tensor_name(name)
 
-        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+        # assuming token_embd.weight is seen before output.weight
+        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
+            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
+                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
+                self.tensor_names.remove("transformer.wte.weight")
+        elif new_name == tok_embd_name:
+            self._has_tok_embd = True
 
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            assert self.tensor_names is not None
-
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                # copy tok_embd.weight to output.weight
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
-        return tensors
+        return [(new_name, data_torch)]
 
 
 @Model.register("InternLM2ForCausalLM")
@@ -3332,7 +3385,7 @@ class Gemma3Model(Model):
 
     # we need to merge the text_config into the root level of hparams
     def __init__(self, *args, **kwargs):
-        hparams = Model.load_hparams(kwargs["dir_model"])
+        hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0])
         if "text_config" in hparams:
             hparams = {**hparams, **hparams["text_config"]}
             kwargs["hparams"] = hparams
@@ -3412,38 +3465,7 @@ class Rwkv6Model(Model):
     model_arch = gguf.MODEL_ARCH.RWKV6
 
     def set_vocab(self):
-        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
-        vocab_size = self.hparams.get("vocab_size", 65536)
-
-        tokens: list[bytes] = ['<s>'.encode("utf-8")]
-        toktypes: list[int] = [gguf.TokenType.CONTROL]
-
-        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
-            lines = f.readlines()
-            for line in lines:
-                parts = line.split(' ')
-                assert len(parts) >= 3
-                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
-                token = token.encode("utf-8") if isinstance(token, str) else token
-                assert isinstance(token, bytes)
-                assert len(token) == token_len
-                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
-                tokens.append(token_text.encode("utf-8"))
-                toktypes.append(gguf.TokenType.NORMAL)
-        remainder = vocab_size - len(tokens)
-        assert remainder >= 0
-        for i in range(len(tokens), vocab_size):
-            tokens.append(f"[PAD{i}]".encode("utf-8"))
-            toktypes.append(gguf.TokenType.UNUSED)
-
-        self.gguf_writer.add_tokenizer_model("rwkv")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
-        # hack: Add '\n\n' as the EOT token to make it chat normally
-        special_vocab._set_special_token("eot", 261)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        self._set_vocab_rwkv_world()
 
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
@@ -3565,6 +3587,168 @@ class RWKV6Qwen2Model(Rwkv6Model):
             yield (new_name, data)
 
 
+@Model.register("Rwkv7ForCausalLM", "RWKV7ForCausalLM")
+class Rwkv7Model(Model):
+    model_arch = gguf.MODEL_ARCH.RWKV7
+
+    def set_vocab(self):
+        self._set_vocab_rwkv_world()
+
+    def calc_lora_rank(self, hidden_size, exponent, multiplier):
+        return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        try:
+            head_size = self.hparams["head_size"]
+            layer_norm_eps = self.hparams["layer_norm_epsilon"]
+        except KeyError:
+            head_size = self.hparams["head_dim"]
+            layer_norm_eps = self.hparams["norm_eps"]
+        hidden_size = self.hparams["hidden_size"]
+        intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else (hidden_size * 4)
+
+        # ICLR: In-Context-Learning-Rate
+        try:
+            lora_rank_decay = self.hparams["lora_rank_decay"] if self.hparams["lora_rank_decay"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["lora_rank_iclr"] if self.hparams["lora_rank_iclr"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["lora_rank_value_residual_mix"] if self.hparams["lora_rank_value_residual_mix"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["lora_rank_gate"] if self.hparams["lora_rank_gate"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+        except KeyError:
+            lora_rank_decay = self.hparams["decay_low_rank_dim"] if self.hparams["decay_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_iclr = self.hparams["a_low_rank_dim"] if self.hparams["a_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8)
+            lora_rank_value_residual_mix = self.hparams["v_low_rank_dim"] if self.hparams["v_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3)
+            lora_rank_gate = self.hparams["gate_low_rank_dim"] if self.hparams["gate_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6)
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    lerp_weights: dict[int, dict[str, Tensor]] = {}
+    lora_needs_transpose: bool = True
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # unify tensor names here to make life easier
+        name = name.replace("blocks", "layers").replace("ffn", "feed_forward")
+        name = name.replace("self_attn", "attention").replace("attn", "attention")
+        name = name.replace("time_mixer.", "")
+        # lora layer names in fla-hub's impl
+        if "_lora.lora" in name:
+            self.lora_needs_transpose = False
+        name = name.replace("_lora.lora.0.weight", "1.weight")
+        name = name.replace("_lora.lora.2.weight", "2.weight")
+        name = name.replace("_lora.lora.2.bias", "0.weight")
+
+        name = name.replace("feed_forward_norm", "ln2")
+        name = name.replace("g_norm", "ln_x")
+
+        if "attention.v" in name and "value" not in self.map_tensor_name(name) and bid == 0:
+            # some models have dummy v0/v1/v2 on first layer while others don't
+            # ignore them all since they are not used
+            return
+
+        wkv_has_gate = self.hparams.get("wkv_has_gate", True)
+        lerp_list = ["r", "w", "k", "v", "a", "g"] if wkv_has_gate else ["r", "w", "k", "v", "a"]
+
+        if bid is not None and "attention.x_" in name:
+            if "attention.x_x" in name:
+                # already concatenated
+                new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                data = data_torch.reshape(len(lerp_list), 1, 1, -1)
+                yield (new_name, data)
+            else:
+                try:
+                    self.lerp_weights[bid][name] = data_torch
+                except KeyError:
+                    self.lerp_weights[bid] = {name: data_torch}
+                if all(f"model.layers.{bid}.attention.x_{i}" in self.lerp_weights[bid].keys() for i in lerp_list):
+                    new_name = f"blk.{bid}.time_mix_lerp_fused.weight"
+                    data = torch.stack([self.lerp_weights[bid][f"model.layers.{bid}.attention.x_{i}"] for i in lerp_list], dim=0)
+                    yield (new_name, data)
+            return
+        else:
+            data_torch = data_torch.squeeze()
+            new_name = self.map_tensor_name(name)
+
+            if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
+                new_name += ".weight"
+
+            if self.lora_needs_transpose and any(
+                new_name.endswith(t) for t in [
+                    "time_mix_w1.weight", "time_mix_w2.weight",
+                    "time_mix_a1.weight", "time_mix_a2.weight",
+                    "time_mix_v1.weight", "time_mix_v2.weight",
+                    "time_mix_g1.weight", "time_mix_g2.weight",
+                ]
+            ):
+                data_torch = data_torch.transpose(0, 1)
+
+            if 'r_k' in new_name:
+                data_torch = data_torch.flatten()
+
+            if bid == 0 and "time_mix_a" in new_name:
+                # dummy v0/v1/v2 on first layer
+                # easist way to make llama happy
+                yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch)
+
+            yield (new_name, data_torch)
+
+
+@Model.register("RwkvHybridForCausalLM")
+class ARwkv7Model(Rwkv7Model):
+    model_arch = gguf.MODEL_ARCH.ARWKV7
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        hidden_size = self.hparams["hidden_size"]
+        head_size = self.hparams["head_size"]
+        rms_norm_eps = self.hparams["rms_norm_eps"]
+        intermediate_size = self.hparams["intermediate_size"]
+        wkv_has_gate = self.hparams["wkv_has_gate"]
+        assert self.hparams["wkv_version"] == 7
+
+        # ICLR: In-Context-Learning-Rate
+        lora_rank_decay = 64
+        lora_rank_iclr = 64
+        lora_rank_value_residual_mix = 32
+        lora_rank_gate = 128 if wkv_has_gate else 0
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_decay_lora_rank(lora_rank_decay)
+        self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr)
+        self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix)
+        self.gguf_writer.add_gate_lora_rank(lora_rank_gate)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_token_shift_count(1)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+
 @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
@@ -3619,8 +3803,6 @@ class MambaModel(Model):
     _tok_embd = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
         tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
 
@@ -3630,6 +3812,10 @@ class MambaModel(Model):
             logger.debug("A_log --> A ==> " + new_name)
             data_torch = -torch.exp(data_torch)
 
+        # [4 1 8192 1] -> [4 8192 1 1]
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
         # assuming token_embd.weight is seen before output.weight
         if self._tok_embd is not None and new_name == output_name:
             if torch.equal(self._tok_embd, data_torch):
@@ -4233,6 +4419,29 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("PLMForCausalLM")
+class PLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.PLM
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+
 @Model.register("T5WithLMHeadModel")
 @Model.register("T5ForConditionalGeneration")
 @Model.register("MT5ForConditionalGeneration")
@@ -5174,7 +5383,7 @@ def main() -> None:
             logger.error(f"Model {model_architecture} is not supported")
             sys.exit(1)
 
-        model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out,
+        model_instance = model_class(dir_model, output_type, fname_out,
                                      is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
                                      eager=args.no_lazy,
                                      metadata_override=args.metadata, model_name=args.model_name,

convert_hf_to_gguf_update.py

@@ -110,6 +110,7 @@ models = [
     {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
+    {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
 ]
 
 

docs/backend/CUDA-FEDORA.md

@@ -14,9 +14,7 @@ In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox
 - [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment)
 - [Installing Essential Development Tools](#installing-essential-development-tools)
 - [Adding the CUDA Repository](#adding-the-cuda-repository)
-- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs)
-- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts)
-- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs)
+- [Installing Nvidia Driver Libraries](#installing-nvidia-driver-libraries)
 - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package)
 - [Configuring the Environment](#configuring-the-environment)
 - [Verifying the Installation](#verifying-the-installation)
@@ -67,7 +65,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for
    sudo dnf distro-sync
    ```
 
-2. **Install the Default Text Editor (Optional):**
+2. **Install **Vim** the default text editor (Optional):**
 
    ```bash
    sudo dnf install vim-default-editor --allowerasing
@@ -97,36 +95,48 @@ After adding the repository, synchronize the package manager again:
 sudo dnf distro-sync
 ```
 
-## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+## Installing Nvidia Driver Libraries
 
-We need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go).
+First, we need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go):
 
 ```bash
 ls -la /usr/lib64/libcuda.so.1
 ```
 
+### If *`libcuda.so.1`* is missing:
+
+```
+ls: cannot access '/usr/lib64/libcuda.so.1': No such file or directory
+```
+
 **Explanation:**
+The host dose not supply the CUDA drivers, **install them now:**
 
-- `nvidia-driver-libs` and `nvidia-driver-cuda-libs` contains necessary NVIDIA driver libraries required by CUDA,
-  on hosts with NVIDIA drivers installed the Fedora Container will supply the host libraries.
-
-### Install Nvidia Driver Libraries on Guest (if `libcuda.so.1` was NOT found).
+#### Install the Nvidia Driver Libraries on Guest:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found).
+### If *`libcuda.so.1`* exists:
+```
+lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07
+```
 
-If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files.
+**Explanation:**
+The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:**
 
-#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies)
+#### Update the Toolbox RPM Database to include the Host-Supplied Libraries:
+
+Note: we do not actually install the libraries, we just update the DB so that the guest system knows they are supplied by the host.
+
+##### 1. Download `nvidia-` parts that are supplied by the host RPM's (with dependencies)
 
 ```bash
-sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-#### 2. Update the RPM database to assume the installation of these packages.
+##### 2. Update the RPM database to assume the installation of these packages.
 
 ```bash
 sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
@@ -134,23 +144,26 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/*
 
 **Note:**
 
-- The `--justdb` option only updates the RPM database, without touching the filesystem.
+- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere.
 
-#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs`
+##### Check that the RPM Database has been correctly updated:
+
+**Note:** This is the same command as in the *"Install the Nvidia Driver Libraries on Guest"* for if *`libcuda.so.1`* was missing.
 
-After manually installing the dependencies, run:
 
 ```bash
-sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs
+sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced
 ```
 
-You should receive a message indicating the package is already installed:
+*(this time it will not install anything, as the database things that these packages are already installed)*
 
 ```
 Updating and loading repositories:
 Repositories loaded.
-Package "nvidia-driver-libs-3:570.86.10-1.fc41.x86_64" is already installed.
-Package "nvidia-driver-cuda-libs-3:570.86.10-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-driver-cuda-libs-3:570.124.06-1.fc41.x86_64" is already installed.
+Package "nvidia-persistenced-3:570.124.06-1.fc41.x86_64" is already installed.
 
 Nothing to do.
 ```
@@ -207,9 +220,9 @@ You should see output similar to:
 ```
 nvcc: NVIDIA (R) Cuda compiler driver
 Copyright (c) 2005-2025 NVIDIA Corporation
-Built on Wed_Jan_15_19:20:09_PST_2025
-Cuda compilation tools, release 12.8, V12.8.61
-Build cuda_12.8.r12.8/compiler.35404655_0
+Built on Fri_Feb_21_20:23:50_PST_2025
+Cuda compilation tools, release 12.8, V12.8.93
+Build cuda_12.8.r12.8/compiler.35583870_0
 ```
 
 This output confirms that the CUDA compiler is accessible and indicates the installed version.

docs/backend/SYCL.md

@@ -237,6 +237,15 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
 cmake --build buildWithCublas --config Release
 ```
 
+**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
+
+```sh
+git clone https://github.com/oneapi-src/oneDNN.git
+cd oneDNN
+cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake --build build-nvidia --config Release
+```
+
 - **Adding support to AMD GPUs**
 
 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
@@ -327,10 +336,10 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -660,8 +669,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
-| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)          | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
+| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
@@ -671,6 +681,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 

docs/build.md

@@ -132,12 +132,14 @@ You may find the official downloads here: [NVIDIA developer site](https://develo
 
 
 #### Compile and run inside a Fedora Toolbox Container
-We also have a [guide](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
+We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/).
 
 **Recommended for:**
-
-- ***Particularly*** *convenient* for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
-- Toolbox is installed by default: [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde).
+- ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
+  - (there are no supported CUDA packages for these systems)
+- ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads).
+  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system)
+- ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean.
 - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download)
 
 
@@ -189,7 +191,7 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
@@ -216,6 +218,7 @@ By default, all supported compute capabilities are enabled. To customize this be
 
 ```bash
 cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21"
+cmake --build build --config Release
 ```
 
 This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time.
@@ -433,6 +436,26 @@ llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
 
 For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
 
+## Arm® KleidiAI™
+KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
+
+To enable KleidiAI, go to the llama.cpp directory and build using CMake
+```bash
+cmake -B build -DGGML_CPU_KLEIDIAI=ON
+cmake --build build --config Release
+```
+You can verify that KleidiAI is being used by running
+```bash
+./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
+```
+If KleidiAI is enabled, the ouput will contain a line similar to:
+```
+load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
+```
+KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`.
+
+Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
+
 ## Android
 
 To read documentation for how to build on Android, [click here](./android.md)

docs/install.md

@@ -9,6 +9,13 @@ brew install llama.cpp
 ```
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
+## MacPorts
+
+```sh
+sudo port install llama.cpp
+```
+see also: https://ports.macports.org/port/llama.cpp/details/
+
 ## Nix
 
 On Mac and Linux, the Nix package manager can be used via

examples/batched-bench/batched-bench.cpp

@@ -132,7 +132,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_clear(ctx);
+                llama_kv_self_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +141,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/cvector-generator/cvector-generator.cpp

@@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;

examples/embedding/embedding.cpp

@@ -38,7 +38,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const struct llama_model * model = llama_get_model(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/gritlm/gritlm.cpp

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 

examples/imatrix/imatrix.cpp

@@ -495,7 +495,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

examples/infill/infill.cpp

@@ -332,8 +332,8 @@ int main(int argc, char ** argv) {
                 LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/llama-bench/llama-bench.cpp

@@ -1578,7 +1578,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // cool off before the test
         if (params.delay) {
@@ -1618,7 +1618,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_self_clear(ctx);
 
             uint64_t t_start = get_time_ns();
 

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

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
+    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.swiftui/README.md

@@ -16,7 +16,7 @@ Open `llama.swiftui.xcodeproj` project in Xcode and you should be able to build
 a simulator or a real device.
 
 To use the framework with a different project, the XCFramework can be added to the project by
-adding `build-ios/llama.xcframework` by dragging and dropping it into the project navigator, or
+adding `build-apple/llama.xcframework` by dragging and dropping it into the project navigator, or
 by manually selecting the framework in the "Frameworks, Libraries, and Embedded Content" section
 of the project settings.
 

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

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_cache_clear(context)
+        llama_kv_self_clear(context)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/llava/clip.cpp

@@ -2989,7 +2989,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     assert(itype < GGML_TYPE_COUNT);
     ggml_type type = static_cast<ggml_type>(itype);
 
-    auto * ctx_clip = clip_model_load(fname_inp, 2);
+    auto * ctx_clip = clip_init(fname_inp, clip_context_params{
+        /* use_gpu */   false,
+        /* verbosity */ 2,
+    });
 
     const auto & ctx_src = ctx_clip->ctx_gguf;
     const auto & ctx_data = ctx_clip->ctx_data;

examples/llava/gemma3-cli.cpp

@@ -309,7 +309,7 @@ int main(int argc, char ** argv) {
             }
             if (line == "/clear") {
                 ctx.n_past = 0;
-                llama_kv_cache_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
                 LOG("Chat history cleared\n\n");
                 continue;
             }

examples/lookahead/lookahead.cpp

@@ -96,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -438,17 +438,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_cache_seq_keep(ctx, seq_id_best);
-            llama_kv_cache_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_cache_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_kv_self_seq_keep(ctx, seq_id_best);
+            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -192,7 +192,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/main/README.md

@@ -27,12 +27,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 ##### Input prompt (One-and-done)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```bash
-./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```bash
+./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 ##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -44,12 +56,24 @@ Once downloaded, place your model in the models folder in llama.cpp.
 
 ##### Input prompt (One-and-done)
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time"
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time"
 ```
 ##### Conversation mode (Allow for continuous interaction with the model)
 
 ```powershell
-./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma
+```
+
+##### Conversation mode using built-in jinja chat template
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja
+```
+
+##### One-and-done query using jinja with custom system prompt and a starting prompt
+
+```powershell
+./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello"
 ```
 
 #### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it):
@@ -77,6 +101,8 @@ The `llama-cli` program provides several ways to interact with the LLaMA models
 
 -   `--prompt PROMPT`: Provide a prompt directly as a command-line option.
 -   `--file FNAME`: Provide a file containing a prompt or multiple prompts.
+-   `--system-prompt PROMPT`: Provide a system prompt (will otherwise use the default one in the chat template (if provided)).
+-   `--system-prompt-file FNAME`: Provide a file containing a system prompt.
 -   `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
 
 ## Interaction
@@ -89,7 +115,10 @@ In interactive mode, users can participate in text generation by injecting their
 
 -   `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
 -   `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
--   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false)
+-   `-cnv,  --conversation`:  Run the program in conversation mode (does not print special tokens and suffix/prefix, use default or provided chat template) (default: true if chat template found)
+-   `-no-cnv`:  Disable conversation mode (default: false)
+-   `-st, --single-turn`:  Only process a single conversation turn (user input) and then exit.
+-   `--jinja`:  Enable jinja chat template parser, will use the model's built-in template or a user-provided one (default: false)
 -   `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
 
 By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
@@ -125,6 +154,8 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t
 
  Example usage: `--chat-template gemma`
 
+`--chat-template-file FNAME`:  Load a custom jinja chat template from an external file, useful if the model contains outdated or incompatible template, some examples can be found in models/templates. Up-to-date chat templates can be downloaded from Hugging Face using scripts/get_chat_template.py
+
 ## Context Management
 
 During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.

examples/main/main.cpp

@@ -354,7 +354,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);
     }
 
     LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
@@ -602,8 +602,8 @@ int main(int argc, char ** argv) {
                     LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_self_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -626,9 +626,9 @@ int main(int argc, char ** argv) {
                     LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_self_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 

examples/parallel/parallel.cpp

@@ -202,7 +202,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -234,9 +234,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
+                llama_kv_self_seq_rm(ctx, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -372,8 +372,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
+                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -133,11 +133,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_update  (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -167,12 +167,12 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+      //llama_kv_self_defrag (ctx);
+        llama_kv_self_update (ctx);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -198,12 +198,12 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+          //llama_kv_self_defrag (ctx);
+            llama_kv_self_update (ctx);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1203,7 +1203,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1575,7 +1575,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
             return;
         }
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1765,7 +1765,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

examples/quantize-stats/quantize-stats.cpp

@@ -1,6 +1,6 @@
 #include "ggml.h"
 #include "llama.h"
-#include "llama-context.h"
+#include "llama-model.h"
 #include "common.h"
 
 #include <algorithm>
@@ -328,7 +328,7 @@ int main(int argc, char ** argv) {
         }
     }
 
-    const auto & tensors = llama_internal_get_tensor_map(ctx);
+    const auto & tensors = llama_internal_get_tensor_map(model);
 
     // check layer tensors
     int included_layers = 0;

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/rpc/CMakeLists.txt

@@ -1,2 +1,4 @@
-add_executable(rpc-server rpc-server.cpp)
-target_link_libraries(rpc-server PRIVATE ggml llama)
+set(TARGET rpc-server)
+add_executable(${TARGET} rpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE ggml)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/rpc/README.md

@@ -72,3 +72,14 @@ $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name
 
 This way you can offload model layers to both local and remote devices.
 
+### Local cache
+
+The RPC server can use a local cache to store large tensors and avoid transferring them over the network.
+This can speed up model loading significantly, especially when using large models.
+To enable the cache, use the `-c` option:
+
+```bash
+$ bin/rpc-server -c
+```
+
+By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.

examples/rpc/rpc-server.cpp

@@ -1,3 +1,7 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
 #include "ggml-cpu.h"
 
 #ifdef GGML_USE_CUDA
@@ -18,26 +22,142 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  define DIRECTORY_SEPARATOR '\\'
+#  include <locale>
 #  include <windows.h>
+#  include <fcntl.h>
+#  include <io.h>
 #else
+#  define DIRECTORY_SEPARATOR '/'
 #  include <unistd.h>
+#  include <sys/stat.h>
 #endif
+#include <codecvt>
 #include <string>
 #include <stdio.h>
+#include <vector>
+#include <filesystem>
+
+namespace fs = std::filesystem;
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+// returns true if successful, false otherwise
+static bool fs_create_directory_with_parents(const std::string & path) {
+#ifdef _WIN32
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring wpath = converter.from_bytes(path);
+
+    // if the path already exists, check whether it's a directory
+    const DWORD attributes = GetFileAttributesW(wpath.c_str());
+    if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+        return true;
+    }
+
+    size_t pos_slash = 0;
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
+        const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
+
+        const bool success = CreateDirectoryW(test, NULL);
+        if (!success) {
+            const DWORD error = GetLastError();
+
+            // if the path already exists, ensure that it's a directory
+            if (error == ERROR_ALREADY_EXISTS) {
+                const DWORD attributes = GetFileAttributesW(subpath.c_str());
+                if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+                    return false;
+                }
+            } else {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#else
+    // if the path already exists, check whether it's a directory
+    struct stat info;
+    if (stat(path.c_str(), &info) == 0) {
+        return S_ISDIR(info.st_mode);
+    }
+
+    size_t pos_slash = 1; // skip leading slashes for directory creation
+
+    // process path from front to back, procedurally creating directories
+    while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
+        const std::string subpath = path.substr(0, pos_slash);
+        struct stat info;
+
+        // if the path already exists, ensure that it's a directory
+        if (stat(subpath.c_str(), &info) == 0) {
+            if (!S_ISDIR(info.st_mode)) {
+                return false;
+            }
+        } else {
+            // create parent directories
+            const int ret = mkdir(subpath.c_str(), 0755);
+            if (ret != 0) {
+                return false;
+            }
+        }
+
+        pos_slash += 1;
+    }
+
+    return true;
+#endif // _WIN32
+}
+
+// NOTE: this is copied from common.cpp to avoid linking with libcommon
+static std::string fs_get_cache_directory() {
+    std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
+    if (getenv("LLAMA_CACHE")) {
+        cache_directory = std::getenv("LLAMA_CACHE");
+    } else {
+#ifdef __linux__
+        if (std::getenv("XDG_CACHE_HOME")) {
+            cache_directory = std::getenv("XDG_CACHE_HOME");
+        } else {
+            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        }
+#elif defined(__APPLE__)
+        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
+#elif defined(_WIN32)
+        cache_directory = std::getenv("LOCALAPPDATA");
+#endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
+        cache_directory += "llama.cpp";
+    }
+    return ensure_trailing_slash(cache_directory);
+}
 
 struct rpc_server_params {
     std::string host        = "127.0.0.1";
     int         port        = 50052;
     size_t      backend_mem = 0;
+    bool        use_cache   = false;
 };
 
 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
     fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
-    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
-    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
+    fprintf(stderr, "  -h, --help                show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
+    fprintf(stderr, "  -c,      --cache          enable local file cache\n");
     fprintf(stderr, "\n");
 }
 
@@ -58,6 +178,8 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
             if (params.port <= 0 || params.port > 65535) {
                 return false;
             }
+        } else if (arg == "-c" || arg == "--cache") {
+            params.use_cache = true;
         } else if (arg == "-m" || arg == "--mem") {
             if (++i >= argc) {
                 return false;
@@ -164,8 +286,20 @@ int main(int argc, char * argv[]) {
     } else {
         get_backend_memory(&free_mem, &total_mem);
     }
-    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
-    ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
+    const char * cache_dir = nullptr;
+    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
+    if (params.use_cache) {
+        if (!fs_create_directory_with_parents(cache_dir_str)) {
+            fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
+            return 1;
+        }
+        cache_dir = cache_dir_str.c_str();
+    }
+    printf("Starting RPC server\n");
+    printf("  endpoint       : %s\n", endpoint.c_str());
+    printf("  local cache    : %s\n", cache_dir ? cache_dir : "n/a");
+    printf("  backend memory : %zu MB\n", free_mem / (1024 * 1024));
+    ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem);
     ggml_backend_free(backend);
     return 0;
 }

examples/run/run.cpp

@@ -38,24 +38,6 @@
 }
 #endif
 
-GGML_ATTRIBUTE_FORMAT(1, 2)
-static std::string fmt(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    const int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::string buf;
-    buf.resize(size);
-    const int size2 = vsnprintf(const_cast<char *>(buf.data()), buf.size() + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-
-    return buf;
-}
-
 GGML_ATTRIBUTE_FORMAT(1, 2)
 static int printe(const char * fmt, ...) {
     va_list args;
@@ -79,6 +61,7 @@ class Opt {
         ctx_params           = llama_context_default_params();
         model_params         = llama_model_default_params();
         context_size_default = ctx_params.n_batch;
+        n_threads_default    = ctx_params.n_threads;
         ngl_default          = model_params.n_gpu_layers;
         common_params_sampling sampling;
         temperature_default = sampling.temp;
@@ -104,6 +87,7 @@ class Opt {
 
         ctx_params.n_batch        = context_size >= 0 ? context_size : context_size_default;
         ctx_params.n_ctx          = ctx_params.n_batch;
+        ctx_params.n_threads = ctx_params.n_threads_batch = n_threads >= 0 ? n_threads : n_threads_default;
         model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default;
         temperature               = temperature >= 0 ? temperature : temperature_default;
 
@@ -116,12 +100,12 @@ class Opt {
     std::string chat_template_file;
     std::string          user;
     bool                 use_jinja   = false;
-    int                  context_size = -1, ngl = -1;
+    int                  context_size = -1, ngl = -1, n_threads = -1;
     float                temperature = -1;
     bool                 verbose     = false;
 
   private:
-    int   context_size_default = -1, ngl_default = -1;
+    int   context_size_default = -1, ngl_default = -1, n_threads_default = -1;
     float temperature_default = -1;
     bool  help                = false;
 
@@ -159,53 +143,94 @@ class Opt {
         return 0;
     }
 
+    int parse_options_with_value(int argc, const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
+            if (handle_option_with_value(argc, argv, i, context_size) == 1) {
+                return 1;
+            }
+        } else if (options_parsing &&
+                   (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
+            if (handle_option_with_value(argc, argv, i, ngl) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && (strcmp(argv[i], "-t") == 0 || strcmp(argv[i], "--threads") == 0)) {
+            if (handle_option_with_value(argc, argv, i, n_threads) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
+            if (handle_option_with_value(argc, argv, i, temperature) == 1) {
+                return 1;
+            }
+        } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0) {
+            if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
+                return 1;
+            }
+            use_jinja = true;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_options(const char ** argv, int & i, bool & options_parsing) {
+        if (options_parsing && (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
+            verbose = true;
+        } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
+            use_jinja = true;
+        } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
+            help = true;
+            return 0;
+        } else if (options_parsing && strcmp(argv[i], "--") == 0) {
+            options_parsing = false;
+        } else {
+            return 2;
+        }
+
+        return 0;
+    }
+
+    int parse_positional_args(const char ** argv, int & i, int & positional_args_i) {
+        if (positional_args_i == 0) {
+            if (!argv[i][0] || argv[i][0] == '-') {
+                return 1;
+            }
+
+            ++positional_args_i;
+            model_ = argv[i];
+        } else if (positional_args_i == 1) {
+            ++positional_args_i;
+            user = argv[i];
+        } else {
+            user += " " + std::string(argv[i]);
+        }
+
+        return 0;
+    }
+
     int parse(int argc, const char ** argv) {
         bool options_parsing   = true;
         for (int i = 1, positional_args_i = 0; i < argc; ++i) {
-            if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) {
-                if (handle_option_with_value(argc, argv, i, context_size) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
-                if (handle_option_with_value(argc, argv, i, ngl) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing && strcmp(argv[i], "--temp") == 0) {
-                if (handle_option_with_value(argc, argv, i, temperature) == 1) {
-                    return 1;
-                }
-            } else if (options_parsing &&
-                       (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
-                verbose = true;
-            } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
-                use_jinja = true;
-            } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0){
-                if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) {
-                    return 1;
-                }
-                use_jinja = true;
-            } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
-                help = true;
-                return 0;
-            } else if (options_parsing && strcmp(argv[i], "--") == 0) {
-                options_parsing = false;
-            } else if (positional_args_i == 0) {
-                if (!argv[i][0] || argv[i][0] == '-') {
-                    return 1;
-                }
+            int ret = parse_options_with_value(argc, argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
 
-                ++positional_args_i;
-                model_ = argv[i];
-            } else if (positional_args_i == 1) {
-                ++positional_args_i;
-                user = argv[i];
-            } else {
-                user += " " + std::string(argv[i]);
+            ret = parse_options(argv, i, options_parsing);
+            if (ret == 0) {
+                continue;
+            } else if (ret == 1) {
+                return ret;
+            }
+
+            if (parse_positional_args(argv, i, positional_args_i)) {
+                return 1;
             }
         }
 
-        if (model_.empty()){
+        if (model_.empty()) {
             return 1;
         }
 
@@ -232,6 +257,8 @@ class Opt {
             "      Number of GPU layers (default: %d)\n"
             "  --temp <value>\n"
             "      Temperature (default: %.1f)\n"
+            "  -t, --threads <value>\n"
+            "      Number of threads to use during generation (default: %d)\n"
             "  -v, --verbose, --log-verbose\n"
             "      Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n"
             "  -h, --help\n"
@@ -260,7 +287,7 @@ class Opt {
             "  llama-run file://some-file3.gguf\n"
             "  llama-run --ngl 999 some-file4.gguf\n"
             "  llama-run --ngl 999 some-file5.gguf Hello World\n",
-            context_size_default, ngl_default, temperature_default);
+            context_size_default, ngl_default, temperature_default, n_threads_default);
     }
 };
 
@@ -480,11 +507,11 @@ class HttpClient {
         int secs = static_cast<int>(seconds) % 60;
 
         if (hrs > 0) {
-            return fmt("%dh %02dm %02ds", hrs, mins, secs);
+            return string_format("%dh %02dm %02ds", hrs, mins, secs);
         } else if (mins > 0) {
-            return fmt("%dm %02ds", mins, secs);
+            return string_format("%dm %02ds", mins, secs);
         } else {
-            return fmt("%ds", secs);
+            return string_format("%ds", secs);
         }
     }
 
@@ -499,7 +526,7 @@ class HttpClient {
             }
         }
 
-        return fmt("%.2f %s", dbl_size, suffix[i]);
+        return string_format("%.2f %s", dbl_size, suffix[i]);
     }
 
     static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t,
@@ -533,7 +560,9 @@ class HttpClient {
         return (now_downloaded_plus_file_size * 100) / total_to_download;
     }
 
-    static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast<long int>(percentage)); }
+    static std::string generate_progress_prefix(curl_off_t percentage) {
+        return string_format("%3ld%% |", static_cast<long int>(percentage));
+    }
 
     static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) {
         const auto                          now             = std::chrono::steady_clock::now();
@@ -544,9 +573,9 @@ class HttpClient {
     static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download,
                                                 double speed, double estimated_time) {
         const int width = 10;
-        return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width,
-                   human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width,
-                   human_readable_time(estimated_time).c_str());
+        return string_format("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(),
+                             width, human_readable_size(total_to_download).c_str(), width,
+                             human_readable_size(speed).c_str(), width, human_readable_time(estimated_time).c_str());
     }
 
     static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) {
@@ -891,7 +920,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;
 
     const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
     prompt_tokens.resize(n_prompt_tokens);
@@ -907,7 +936,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
     const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");

examples/save-load-state/save-load-state.cpp

@@ -15,7 +15,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    print_build_info();
+    common_init();
 
     if (params.n_predict < 0) {
         params.n_predict = 16;
@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_cache_clear(ctx3);
+        llama_kv_self_clear(ctx3);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/server/server.cpp

@@ -489,8 +489,12 @@ struct result_timings {
     double predicted_per_token_ms;
     double predicted_per_second;
 
+    // Optional speculative metrics - only included when > 0
+    int32_t draft_n = 0;
+    int32_t draft_n_accepted = 0;
+
     json to_json() const {
-        return {
+        json base = {
             {"prompt_n",               prompt_n},
             {"prompt_ms",              prompt_ms},
             {"prompt_per_token_ms",    prompt_per_token_ms},
@@ -501,6 +505,13 @@ struct result_timings {
             {"predicted_per_token_ms", predicted_per_token_ms},
             {"predicted_per_second",   predicted_per_second},
         };
+
+        if (draft_n > 0) {
+            base["draft_n"] = draft_n;
+            base["draft_n_accepted"] = draft_n_accepted;
+        }
+
+        return base;
     }
 };
 
@@ -830,6 +841,11 @@ struct server_task_result_cmpl_final : server_task_result {
             ret.push_back({"timings", timings.to_json()});
         }
 
+        // extra fields for debugging purposes
+        if (verbose) {
+            ret["__verbose"] = to_json_non_oaicompat();
+        }
+
         return ret;
     }
 };
@@ -1294,6 +1310,10 @@ struct server_slot {
 
     std::function<void(int)> callback_on_release;
 
+    // Speculative decoding stats
+    int32_t n_draft_total = 0;      // Total draft tokens generated
+    int32_t n_draft_accepted = 0;   // Draft tokens actually accepted
+
     void reset() {
         SLT_DBG(*this, "%s", "\n");
 
@@ -1310,6 +1330,10 @@ struct server_slot {
 
         generated_tokens.clear();
         generated_token_probs.clear();
+
+        // clear speculative decoding stats
+        n_draft_total = 0;
+        n_draft_accepted = 0;
     }
 
     bool is_non_causal() const {
@@ -1376,6 +1400,12 @@ struct server_slot {
         timings.predicted_per_token_ms = t_token_generation / n_decoded;
         timings.predicted_per_second = 1e3 / t_token_generation * n_decoded;
 
+        // Add speculative metrics
+        if (n_draft_total > 0) {
+            timings.draft_n = n_draft_total;
+            timings.draft_n_accepted = n_draft_accepted;
+        }
+
         return timings;
     }
 
@@ -1423,6 +1453,15 @@ struct server_slot {
                 t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second,
                 t_token_generation, n_decoded, t_gen, n_gen_second,
                 t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded);
+
+        if (n_draft_total > 0) {
+            const float draft_ratio = (float) n_draft_accepted / n_draft_total;
+            SLT_INF(*this,
+                    "\n"
+                    "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n",
+                    draft_ratio, n_draft_accepted, n_draft_total
+            );
+        }
     }
 
     json to_json() const {
@@ -1872,6 +1911,10 @@ struct server_context {
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
 
+            // force F16 KV cache for the draft model for extra performance
+            params_dft.cache_type_k = GGML_TYPE_F16;
+            params_dft.cache_type_v = GGML_TYPE_F16;
+
             llama_init_dft = common_init_from_params(params_dft);
 
             model_dft = llama_init_dft.model.get();
@@ -1892,10 +1935,6 @@ struct server_context {
             cparams_dft = common_context_params_to_llama(params_dft);
             cparams_dft.n_batch = n_ctx_dft;
 
-            // force F16 KV cache for the draft model for extra performance
-            cparams_dft.type_k = GGML_TYPE_F16;
-            cparams_dft.type_v = GGML_TYPE_F16;
-
             // the context is not needed - we will create one for each slot
             llama_init_dft.context.reset();
         }
@@ -2040,6 +2079,18 @@ struct server_context {
         return ret;
     }
 
+    bool can_be_detokenized(const struct llama_context * ctx, const std::vector<llama_token> & tokens) {
+        const llama_model * model = llama_get_model(ctx);
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const int32_t n_vocab = llama_vocab_n_tokens(vocab);
+        for (const auto & token : tokens) {
+            if (token < 0 || token >= n_vocab) {
+                return false;
+            }
+        }
+        return true;
+    }
+
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot.reset();
         slot.id_task       = task.id;
@@ -2054,6 +2105,11 @@ struct server_context {
             slot.lora = task.params.lora;
         }
 
+        bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens);
+        if (!can_detokenize) {
+            send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST);
+            return false;
+        }
         SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@@ -2096,7 +2152,7 @@ struct server_context {
         SRV_DBG("%s", "clearing KV cache\n");
 
         // clear the entire KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
         clean_kv_cache = false;
     }
 
@@ -2638,8 +2694,8 @@ struct server_context {
                     res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                     res->t_start             = metrics.t_start;
 
-                    res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx);
-                    res->kv_cache_used_cells   = llama_get_kv_cache_used_cells(ctx);
+                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
+                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);
 
                     res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                     res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
@@ -2755,7 +2811,7 @@ struct server_context {
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
+                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
                     slot->cache_tokens.clear();
 
                     auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2823,8 +2879,8 @@ struct server_context {
 
                 SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
+                llama_kv_self_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
+                llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 if (slot.params.cache_prompt) {
                     for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@@ -3015,8 +3071,8 @@ struct server_context {
 
                                             const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
 
-                                            llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
-                                            llama_kv_cache_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
+                                            llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
+                                            llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@@ -3054,9 +3110,9 @@ struct server_context {
                     }
 
                     // keep only the common part
-                    if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
                         // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
+                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
 
                         // there is no common part left
                         slot.n_past = 0;
@@ -3268,6 +3324,9 @@ struct server_context {
 
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
+                // keep track of total number of tokens generated in the draft
+                slot.n_draft_total += draft.size();
+
                 // ignore small drafts
                 if (slot.params.speculative.n_min > (int) draft.size()) {
                     SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
@@ -3293,10 +3352,13 @@ struct server_context {
                 slot.n_past    += ids.size();
                 slot.n_decoded += ids.size();
 
+                // update how many tokens out of draft was accepted
+                slot.n_draft_accepted += ids.size() - 1;
+
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
 
-                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
 
                 for (size_t i = 0; i < ids.size(); ++i) {
                     completion_token_output result;
@@ -4437,15 +4499,24 @@ int main(int argc, char ** argv) {
         llama_backend_free();
     };
 
-    // bind HTTP listen port
     bool was_bound = false;
-    if (params.port == 0) {
-        int bound_port = svr->bind_to_any_port(params.hostname);
-        if ((was_bound = (bound_port >= 0))) {
-            params.port = bound_port;
-        }
+    if (string_ends_with(std::string(params.hostname), ".sock")) {
+        LOG_INF("%s: setting address family to AF_UNIX\n", __func__);
+        svr->set_address_family(AF_UNIX);
+        // bind_to_port requires a second arg, any value other than 0 should
+        // simply get ignored
+        was_bound = svr->bind_to_port(params.hostname, 8080);
     } else {
-        was_bound = svr->bind_to_port(params.hostname, params.port);
+        LOG_INF("%s: binding port with default address family\n", __func__);
+        // bind HTTP listen port
+        if (params.port == 0) {
+            int bound_port = svr->bind_to_any_port(params.hostname);
+            if ((was_bound = (bound_port >= 0))) {
+                params.port = bound_port;
+            }
+        } else {
+            was_bound = svr->bind_to_port(params.hostname, params.port);
+        }
     }
 
     if (!was_bound) {

examples/server/tests/utils.py

@@ -302,7 +302,7 @@ class ServerPreset:
         server.model_hf_repo = "ggml-org/models"
         server.model_hf_file = "tinyllamas/stories260K.gguf"
         server.model_alias = "tinyllama-2"
-        server.n_ctx = 256
+        server.n_ctx = 512
         server.n_batch = 32
         server.n_slots = 2
         server.n_predict = 64

examples/server/utils.hpp

@@ -621,7 +621,9 @@ static json oaicompat_completion_params_parse(
 
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
     llama_params["prompt"]           = chat_params.prompt;
-    llama_params["grammar"]          = chat_params.grammar;
+    if (!chat_params.grammar.empty()) {
+        llama_params["grammar"] = chat_params.grammar;
+    }
     llama_params["grammar_lazy"]     = chat_params.grammar_lazy;
     auto grammar_triggers = json::array();
     for (const auto & trigger : chat_params.grammar_triggers) {

examples/server/webui/src/components/ChatScreen.tsx

@@ -99,13 +99,9 @@ export default function ChatScreen() {
     canvasData,
     replaceMessageAndGenerate,
   } = useAppContext();
-  const [inputMsg, setInputMsg] = useState(prefilledMsg.content());
-  const inputRef = useRef<HTMLTextAreaElement>(null);
+  const textarea = useOptimizedTextarea(prefilledMsg.content());
 
-  const { extraContext, clearExtraContext } = useVSCodeContext(
-    inputRef,
-    setInputMsg
-  );
+  const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
   // TODO: improve this when we have "upload file" feature
   const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
 
@@ -135,9 +131,10 @@ export default function ChatScreen() {
   };
 
   const sendNewMessage = async () => {
-    if (inputMsg.trim().length === 0 || isGenerating(currConvId ?? '')) return;
-    const lastInpMsg = inputMsg;
-    setInputMsg('');
+    const lastInpMsg = textarea.value();
+    if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? ''))
+      return;
+    textarea.setValue('');
     scrollToBottom(false);
     setCurrNodeId(-1);
     // get the last message node
@@ -146,13 +143,13 @@ export default function ChatScreen() {
       !(await sendMessage(
         currConvId,
         lastMsgNodeId,
-        inputMsg,
+        lastInpMsg,
         currExtra,
         onChunk
       ))
     ) {
       // restore the input message if failed
-      setInputMsg(lastInpMsg);
+      textarea.setValue(lastInpMsg);
     }
     // OK
     clearExtraContext();
@@ -195,16 +192,13 @@ export default function ChatScreen() {
       // send the prefilled message if needed
       sendNewMessage();
     } else {
-      // otherwise, focus on the input and move the cursor to the end
-      if (inputRef.current) {
-        inputRef.current.focus();
-        inputRef.current.selectionStart = inputRef.current.value.length;
-      }
+      // otherwise, focus on the input
+      textarea.focus();
     }
     prefilledMsg.clear();
     // no need to keep track of sendNewMessage
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [inputRef]);
+  }, [textarea.ref]);
 
   // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg)
   const pendingMsgDisplay: MessageDisplay[] =
@@ -258,9 +252,7 @@ export default function ChatScreen() {
           <textarea
             className="textarea textarea-bordered w-full"
             placeholder="Type a message (Shift+Enter to add a new line)"
-            ref={inputRef}
-            value={inputMsg}
-            onChange={(e) => setInputMsg(e.target.value)}
+            ref={textarea.ref}
             onKeyDown={(e) => {
               if (e.nativeEvent.isComposing || e.keyCode === 229) return;
               if (e.key === 'Enter' && e.shiftKey) return;
@@ -280,11 +272,7 @@ export default function ChatScreen() {
               Stop
             </button>
           ) : (
-            <button
-              className="btn btn-primary ml-2"
-              onClick={sendNewMessage}
-              disabled={inputMsg.trim().length === 0}
-            >
+            <button className="btn btn-primary ml-2" onClick={sendNewMessage}>
               Send
             </button>
           )}
@@ -298,3 +286,43 @@ export default function ChatScreen() {
     </div>
   );
 }
+
+export interface OptimizedTextareaValue {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  useEffect(() => {
+    if (textareaRef.current && savedInitValue) {
+      textareaRef.current.value = savedInitValue;
+      setSavedInitValue('');
+    }
+  }, [textareaRef, savedInitValue, setSavedInitValue]);
+
+  return {
+    value: () => {
+      return textareaRef.current?.value ?? savedInitValue;
+    },
+    setValue: (value: string) => {
+      if (textareaRef.current) {
+        textareaRef.current.value = value;
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        // focus and move the cursor to the end
+        textareaRef.current.focus();
+        textareaRef.current.selectionStart = textareaRef.current.value.length;
+      }
+    },
+    ref: textareaRef,
+  };
+}

examples/server/webui/src/utils/llama-vscode.ts

@@ -1,5 +1,6 @@
 import { useEffect, useState } from 'react';
 import { MessageExtraContext } from './types';
+import { OptimizedTextareaValue } from '../components/ChatScreen';
 
 // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
 // Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@@ -14,10 +15,7 @@ interface SetTextEvData {
  * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n  return 123' }, '*');
  */
 
-export const useVSCodeContext = (
-  inputRef: React.RefObject<HTMLTextAreaElement>,
-  setInputMsg: (text: string) => void
-) => {
+export const useVSCodeContext = (textarea: OptimizedTextareaValue) => {
   const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
     null
   );
@@ -27,20 +25,20 @@ export const useVSCodeContext = (
     const handleMessage = (event: MessageEvent) => {
       if (event.data?.command === 'setText') {
         const data: SetTextEvData = event.data;
-        setInputMsg(data?.text);
+        textarea.setValue(data?.text);
         if (data?.context && data.context.length > 0) {
           setExtraContext({
             type: 'context',
             content: data.context,
           });
         }
-        inputRef.current?.focus();
+        textarea.focus();
       }
     };
 
     window.addEventListener('message', handleMessage);
     return () => window.removeEventListener('message', handleMessage);
-  }, [inputRef, setInputMsg]);
+  }, [textarea]);
 
   // Add a keydown listener that sends the "escapePressed" message to the parent window
   useEffect(() => {

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_get_kv_cache_used_cells(ctx) == 0;
+        const bool is_first = llama_kv_self_used_cells(ctx) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_get_kv_cache_used_cells(ctx);
+            int n_ctx_used = llama_kv_self_used_cells(ctx);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");

examples/speculative-simple/speculative-simple.cpp

@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -331,11 +331,11 @@ int main(int argc, char ** argv) {
                         }
 
                         active_seqs.erase(s);
-                        for(int i = 0; i < n_seq_dft; i++) {
+                        for (int i = 0; i < n_seq_dft; i++) {
                             if (i == s) {
                                 continue;
                             }
-                            if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
+                            if (drafts[i].active && drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
                                 // synchronize active status for sequences with the same drafted token
                                 drafts[i].active = drafts[i].active && accept;
                                 if (!drafts[i].active) {
@@ -420,14 +420,14 @@ int main(int argc, char ** argv) {
             {
                 LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
 
-                llama_kv_cache_seq_keep(ctx_dft, s_keep);
-                llama_kv_cache_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_dft, 0);
+                llama_kv_self_seq_keep(ctx_dft, s_keep);
+                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_dft, 0);
 
-                llama_kv_cache_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, s_keep);
-                llama_kv_cache_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, 0);
+                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
+                llama_kv_self_seq_keep(ctx_tgt, s_keep);
+                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +444,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
             // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode(ctx_dft, batch_dft);
 
@@ -503,8 +503,8 @@ int main(int argc, char ** argv) {
                     if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
-                        llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
+                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
 
                         // all previous tokens from this branch are now also part of the new branch
                         for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -585,9 +585,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_cache_seq_keep(ctx_tgt, 0);
+            llama_kv_self_seq_keep(ctx_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
             }
 
             // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());

examples/tts/tts.cpp

@@ -87,11 +87,11 @@ struct wav_header {
     uint32_t data_size;
 };
 
-static void save_wav16(const std::string & fname, const std::vector<float> & data, int sample_rate) {
+static bool save_wav16(const std::string & fname, const std::vector<float> & data, int sample_rate) {
     std::ofstream file(fname, std::ios::binary);
     if (!file) {
-        LOG_ERR("%s: Failed to open file '%s' for writing", __func__, fname.c_str());
-        return;
+        LOG_ERR("%s: Failed to open file '%s' for writing.\n", __func__, fname.c_str());
+        return false;
     }
 
     wav_header header;
@@ -108,7 +108,7 @@ static void save_wav16(const std::string & fname, const std::vector<float> & dat
         file.write(reinterpret_cast<const char*>(&pcm_sample), sizeof(pcm_sample));
     }
 
-    file.close();
+    return file.good();
 }
 
 static void fill_hann_window(int length, bool periodic, float * output) {
@@ -536,6 +536,7 @@ static std::string audio_data_from_speaker(json speaker, const outetts_version t
 int main(int argc, char ** argv) {
     common_params params;
 
+    params.out_file = "output.wav";
     params.prompt = "";
 
     params.n_predict = 4096;
@@ -570,6 +571,10 @@ int main(int argc, char ** argv) {
     model_ttc = llama_init_ttc.model.get();
     ctx_ttc   = llama_init_ttc.context.get();
 
+    if (model_ttc == nullptr || ctx_ttc == nullptr) {
+        return ENOENT;
+    }
+
     const llama_vocab * vocab = llama_model_get_vocab(model_ttc);
 
     // TODO: refactor in a common struct
@@ -585,6 +590,10 @@ int main(int argc, char ** argv) {
     model_cts = llama_init_cts.model.get();
     ctx_cts   = llama_init_cts.context.get();
 
+    if (model_cts == nullptr || ctx_cts == nullptr) {
+        return ENOENT;
+    }
+
     std::vector<common_sampler *> smpl(n_parallel);
     for (int i = 0; i < n_parallel; ++i) {
         params.sampling.no_perf = (i != 0);
@@ -1060,8 +1069,6 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
     }
 #endif
 
-    const std::string fname = "output.wav";
-
     const int n_sr = 24000; // sampling rate
 
     // zero out first 0.25 seconds
@@ -1072,11 +1079,15 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
     LOG_INF("%s: time for spectral ops: %.3f ms\n", __func__, (ggml_time_us() - t_spec_start) / 1000.0f);
     LOG_INF("%s: total time:            %.3f ms\n", __func__, (ggml_time_us() - t_main_start) / 1000.0f);
 
-    save_wav16(fname, audio, n_sr);
+    int retval = 0;
 
-    LOG_INF("%s: audio written to file '%s'\n", __func__, fname.c_str());
+    if (save_wav16(params.out_file, audio, n_sr)) {
+        LOG_INF("%s: audio written to file '%s'\n", __func__, params.out_file.c_str());
+    } else {
+        retval = ENOENT;
+    }
 
     llama_backend_free();
 
-    return 0;
+    return retval;
 }

ggml/CMakeLists.txt

@@ -123,10 +123,12 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
-set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
+set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
 if (WIN32)
@@ -186,6 +188,7 @@ option(GGML_OPENMP                          "ggml: use OpenMP"
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
+option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING

ggml/cmake/GitVars.cmake

@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

ggml/cmake/common.cmake

@@ -0,0 +1,26 @@
+function(ggml_get_flags CCID CCVER)
+    set(C_FLAGS "")
+    set(CXX_FLAGS "")
+
+    if (CCID MATCHES "Clang")
+        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
+        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
+
+        if (
+            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
+            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
+        )
+            list(APPEND C_FLAGS -Wdouble-promotion)
+        endif()
+    elseif (CCID STREQUAL "GNU")
+        set(C_FLAGS   -Wdouble-promotion)
+        set(CXX_FLAGS -Wno-array-bounds)
+
+        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
+            list(APPEND CXX_FLAGS -Wextra-semi)
+        endif()
+    endif()
+
+    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
+    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
+endfunction()

ggml/cmake/ggml-config.cmake.in

@@ -5,7 +5,7 @@
 
 set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
 set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
-set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
+#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
 
 find_package(Threads REQUIRED)
 

ggml/include/ggml-rpc.h

@@ -17,7 +17,9 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c
 
 GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint,
+                                                    const char * cache_dir,
+                                                    size_t free_mem, size_t total_mem);
 
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 

ggml/include/ggml.h

@@ -454,6 +454,7 @@ extern "C" {
         GGML_OP_RMS_NORM,
         GGML_OP_RMS_NORM_BACK,
         GGML_OP_GROUP_NORM,
+        GGML_OP_L2_NORM,
 
         GGML_OP_MUL_MAT,
         GGML_OP_MUL_MAT_ID,
@@ -502,6 +503,7 @@ extern "C" {
         GGML_OP_ADD_REL_POS,
         GGML_OP_RWKV_WKV6,
         GGML_OP_GATED_LINEAR_ATTN,
+        GGML_OP_RWKV_WKV7,
 
         GGML_OP_UNARY,
 
@@ -1095,6 +1097,18 @@ extern "C" {
             int                   n_groups,
             float                 eps);
 
+    // l2 normalize along rows
+    // used in rwkv v7
+    GGML_API struct ggml_tensor * ggml_l2_norm(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
+    GGML_API struct ggml_tensor * ggml_l2_norm_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 eps);
+
     // a - x
     // b - dy
     GGML_API struct ggml_tensor * ggml_rms_norm_back(
@@ -1777,11 +1791,11 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd, n_batch,     n_head,    1]
-    // k:    [n_embd, n_kv,        n_head_kv, 1]
-    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    1]
+    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
+    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1890,6 +1904,16 @@ extern "C" {
             struct ggml_tensor  * state,
             float scale);
 
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv7(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * r,
+            struct ggml_tensor  * w,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * state);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);

ggml/src/CMakeLists.txt

@@ -1,4 +1,5 @@
 include(CheckCXXCompilerFlag)
+include("../cmake/common.cmake")
 
 add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES})
 
@@ -24,33 +25,6 @@ if (NOT MSVC)
     endif()
 endif()
 
-function(ggml_get_flags CCID CCVER)
-    set(C_FLAGS "")
-    set(CXX_FLAGS "")
-
-    if (CCID MATCHES "Clang")
-        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
-        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
-
-        if (
-            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
-            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
-        )
-            list(APPEND C_FLAGS -Wdouble-promotion)
-        endif()
-    elseif (CCID STREQUAL "GNU")
-        set(C_FLAGS   -Wdouble-promotion)
-        set(CXX_FLAGS -Wno-array-bounds)
-
-        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
-            list(APPEND CXX_FLAGS -Wextra-semi)
-        endif()
-    endif()
-
-    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
-    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
-endfunction()
-
 if (GGML_FATAL_WARNINGS)
     if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
         list(APPEND C_FLAGS   -Werror)
@@ -102,7 +76,11 @@ if (GGML_CCACHE)
             set(GGML_CCACHE_VARIANT sccache)
         endif()
         # TODO: should not be set globally
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32)
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl")
+        else ()
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        endif ()
         set(ENV{CCACHE_SLOPPINESS} time_macros)
         message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
     else()
@@ -351,6 +329,10 @@ if (CMAKE_SYSTEM_NAME MATCHES "Android")
     target_link_libraries(ggml-base PRIVATE dl)
 endif()
 
+if(CMAKE_SYSTEM_NAME MATCHES "visionOS")
+    target_compile_definitions(ggml-base PUBLIC _DARWIN_C_SOURCE)
+endif()
+
 if (BUILD_SHARED_LIBS)
     foreach (target ggml-base ggml)
         set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON)

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2790,10 +2790,14 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
                 (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
                 output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
                 output_ne_offset);
+            int64_t antiquantGroupSize = 0;
+            if (src0->ne[0] > QK8_0) {
+                antiquantGroupSize = QK8_0;
+            }
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                 acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
-                nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
+                nullptr, nullptr, nullptr, antiquantGroupSize, acl_output_tensor,
                 &workspaceSize, &executor));
             if (workspaceAddr == nullptr) {
                 workspaceAddr = workspace_allocator.alloc(workspaceSize);
@@ -2833,7 +2837,7 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
 
                 ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                     acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
-                    nullptr, nullptr, nullptr, nullptr, QK8_0,
+                    nullptr, nullptr, nullptr, nullptr, antiquantGroupSize,
                     acl_output_tensor, &workspaceSize, &executor));
                 ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
                     workspaceAddr, workspaceSize, executor, ctx.stream()));

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1689,11 +1689,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
                 case GGML_TYPE_Q8_0:
-                    // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
-                    if (op->src[0]->ne[0] <= QK8_0) {
-                        return false;
-                    }
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                 case GGML_TYPE_Q4_0:

ggml/src/ggml-cpu/CMakeLists.txt

@@ -287,17 +287,31 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 endif()
             endif()
         endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
         message(STATUS "PowerPC detected")
-        execute_process(COMMAND bash -c "grep POWER /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER_M)
-        if (${POWER_M} MATCHES "POWER10")
-            list(APPEND ARCH_FLAGS -mcpu=power10)
-        elseif (${POWER_M} MATCHES "POWER9")
-            list(APPEND ARCH_FLAGS -mcpu=power9)
-        elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-            list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+        if (GGML_NATIVE)
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+                file(READ "/proc/cpuinfo" POWER10_M)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
+                execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
+            endif()
+
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
+
+            if (EXTRACTED_NUMBER GREATER_EQUAL 10)
+                list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
+            elseif (EXTRACTED_NUMBER EQUAL 9)
+                list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
+            elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+                list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
+            else()
+                list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
+            endif()
         else()
-            list(APPEND ARCH_FLAGS -mcpu=powerpc64 -mtune=native)
+            if (GGML_CPU_POWERPC_CPUTYPE)
+                list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
+            endif()
         endif()
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         message(STATUS "loongarch64 detected")
@@ -312,7 +326,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
         message(STATUS "RISC-V detected")
         if (GGML_RVV)
-            list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+            if (GGML_RV_ZFH)
+                list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -DGGML_RV_ZFH -mabi=lp64d)
+            else()
+                list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+            endif()
         endif()
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
         message(STATUS "s390x detected")
@@ -351,9 +369,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.3.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.5.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "060bd2dc64642b091f461cc8dd7426d9")
+        set(KLEIDIAI_ARCHIVE_MD5  "ea22e1aefb800e9bc8c74d91633cc58e")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3110,17 +3110,17 @@ static void ggml_compute_forward_dup_same_cont(
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    // parallelize by elements
-    const int ne = ggml_nelements(dst);
-    const int dr = (ne + nth - 1) / nth;
-    const int ie0 = dr * ith;
-    const int ie1 = MIN(ie0 + dr, ne);
+    // parallelize by blocks
+    const int nk = ggml_nelements(src0)/ggml_blck_size(src0->type);
+    const int dr = (nk + nth - 1) / nth;
+    const int k0 = dr * ith;
+    const int k1 = MIN(k0 + dr, nk);
 
-    if (ie0 < ie1) {
+    if (k0 < k1) {
         memcpy(
-            ((char *)  dst->data + ie0*nb0),
-            ((char *) src0->data + ie0*nb0),
-            (ie1 - ie0) * nb0);
+            ((char *)  dst->data + k0*nb0),
+            ((char *) src0->data + k0*nb0),
+            (k1 - k0) * nb0);
     }
 }
 
@@ -4055,7 +4055,6 @@ static void ggml_compute_forward_dup_f32(
 static void ggml_compute_forward_dup_bytes(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
-
     const struct ggml_tensor * src0 = dst->src[0];
 
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
@@ -4069,10 +4068,10 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     const size_t type_size = ggml_type_size(src0->type);
+
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -4082,10 +4081,10 @@ static void ggml_compute_forward_dup_bytes(
     const int ir1 = MIN(ir0 + dr, nr);
 
     if (src0->type == dst->type &&
-        ne00 == ne0 &&
+        ggml_are_same_shape(src0, dst) &&
         nb00 == type_size && nb0 == type_size) {
         // copy by rows
-        const size_t rs = ne00 * type_size;
+        const size_t rs = ggml_row_size(src0->type, ne00);
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 for (int64_t i01 = ir0; i01 < ir1; i01++) {
@@ -4140,17 +4139,20 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     // dst counters
-
-    int64_t i10 = 0;
+    int64_t k10 = 0;
     int64_t i11 = 0;
     int64_t i12 = 0;
     int64_t i13 = 0;
 
+    // number of blocks in a row
+    const int64_t nk00 = ne00 / ggml_blck_size(src0->type);
+    const int64_t nk0  = ne0  / ggml_blck_size(dst->type);
+
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
-            i10 += ne00 * ir0;
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * ir0;
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -4162,14 +4164,14 @@ static void ggml_compute_forward_dup_bytes(
                 }
             }
             for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+                for (int64_t k00 = 0; k00 < nk00; k00++) {
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
                     memcpy(dst_ptr, src0_ptr, type_size);
 
-                    if (++i10 == ne0) {
-                        i10 = 0;
+                    if (++k10 == nk0) {
+                        k10 = 0;
                         if (++i11 == ne1) {
                             i11 = 0;
                             if (++i12 == ne2) {
@@ -4182,9 +4184,9 @@ static void ggml_compute_forward_dup_bytes(
                     }
                 }
             }
-            i10 += ne00 * (ne01 - ir1);
-            while (i10 >= ne0) {
-                i10 -= ne0;
+            k10 += nk00 * (ne01 - ir1);
+            while (k10 >= nk0) {
+                k10 -= nk0;
                 if (++i11 == ne1) {
                     i11 = 0;
                     if (++i12 == ne2) {
@@ -8548,6 +8550,69 @@ static void ggml_compute_forward_group_norm(
     }
 }
 
+// ggml_compute_forward_l2_norm
+
+static void ggml_compute_forward_l2_norm_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    GGML_ASSERT(eps >= 0.0f);
+
+    // TODO: optimize
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
+                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+
+                ggml_float sum = 0.0;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    sum += (ggml_float)(x[i00] * x[i00]);
+                }
+
+                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
+
+                memcpy(y, x, ne00 * sizeof(float));
+
+                const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
+
+                ggml_vec_scale_f32(ne00, y, scale);
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_l2_norm(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_l2_norm_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_mul_mat
 
 static void ggml_compute_forward_mul_mat_one_chunk(
@@ -12173,10 +12238,11 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t D = neq0;
-    const int64_t N = neq1;
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
 
-    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne0 == DV);
     GGML_ASSERT(ne2 == N);
 
     // input tensor rows must be contiguous
@@ -12184,12 +12250,11 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(nbk0 == ggml_type_size(k->type));
     GGML_ASSERT(nbv0 == ggml_type_size(v->type));
 
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev0 == D);
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
 
     GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nev0 == D);
 
     // dst cannot be transposed or permuted
     GGML_ASSERT(nb0 == sizeof(float));
@@ -12255,15 +12320,15 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         float S = 0.0f;      // sum
         float M = -INFINITY; // maximum KQ value
 
-        float       * VKQ32 = (float       *) params->wdata + ith*(3*D + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator
-        float       * V32   =                 (VKQ32 + 1*D); // (temporary) FP32 V buffer
-        ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*D); // (temporary) FP16 VKQ accumulator
-        ggml_fp16_t * Q_q   = (ggml_fp16_t *) (VKQ32 + 2*D); // (temporary) buffer for Q converted to quantized/FP16
+        float       * VKQ32 = (float       *) params->wdata + ith*(1*DK + 2*DV + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator
+        float       * V32   =                 (VKQ32 + 1*DV); // (temporary) FP32 V buffer
+        ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*DV); // (temporary) FP16 VKQ accumulator
+        ggml_fp16_t * Q_q   = (ggml_fp16_t *) (VKQ32 + 2*DV); // (temporary) buffer for Q converted to quantized/FP16
 
         if (v->type == GGML_TYPE_F16) {
-            memset(VKQ16, 0, D*sizeof(ggml_fp16_t));
+            memset(VKQ16, 0, DV*sizeof(ggml_fp16_t));
         } else {
-            memset(VKQ32, 0, D*sizeof(float));
+            memset(VKQ32, 0, DV*sizeof(float));
         }
 
         const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
@@ -12277,7 +12342,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int iv2 = iq2 / rv2;
 
         const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3));
-        q_to_vec_dot(pq, Q_q, D);
+        q_to_vec_dot(pq, Q_q, DK);
 
         // online softmax / attention
         // loop over n_kv and n_head_kv
@@ -12291,7 +12356,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             float s; // KQ value
 
             const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3);
-            kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1);
+            kq_vec_dot(DK, &s, 0, k_data, 0, Q_q, 0, 1);
 
             s = s*scale; // scale KQ value
 
@@ -12315,14 +12380,14 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                     ms = expf(Mold - M);
 
                     // V = V*expf(Mold - M)
-                    ggml_vec_scale_f16(D, VKQ16, ms);
+                    ggml_vec_scale_f16(DV, VKQ16, ms);
                 } else {
                     // no new maximum, ms == 1.0f, vs != 1.0f
                     vs = expf(s - M);
                 }
 
                 // V += v*expf(s - M)
-                ggml_vec_mad_f16(D, VKQ16, (const ggml_fp16_t *) v_data, vs);
+                ggml_vec_mad_f16(DV, VKQ16, (const ggml_fp16_t *) v_data, vs);
             } else {
                 if (s > M) {
                     // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
@@ -12330,30 +12395,30 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                     ms = expf(Mold - M);
 
                     // V = V*expf(Mold - M)
-                    ggml_vec_scale_f32(D, VKQ32, ms);
+                    ggml_vec_scale_f32(DV, VKQ32, ms);
                 } else {
                     // no new maximum, ms == 1.0f, vs != 1.0f
                     vs = expf(s - M);
                 }
 
-                v_to_float(v_data, V32, D);
+                v_to_float(v_data, V32, DV);
 
                 // V += v*expf(s - M)
-                ggml_vec_mad_f32(D, VKQ32, V32, vs);
+                ggml_vec_mad_f32(DV, VKQ32, V32, vs);
             }
 
             S = S*ms + vs; // scale and increment sum with partial sum
         }
 
         if (v->type == GGML_TYPE_F16) {
-            for (int64_t d = 0; d < D; ++d) {
+            for (int64_t d = 0; d < DV; ++d) {
                 VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]);
             }
         }
 
         // V /= S
         const float S_inv = 1.0f/S;
-        ggml_vec_scale_f32(D, VKQ32, S_inv);
+        ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
         // dst indices
         const int i1 = iq1;
@@ -13604,6 +13669,184 @@ static void ggml_compute_forward_gla(
     }
 }
 
+// ggml_compute_forward_rwkv_wkv7
+
+static void ggml_compute_forward_rwkv_wkv7_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const int64_t T = dst->src[1]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t HEADS = dst->src[1]->ne[1];
+    const int64_t n_seqs = dst->src[6]->ne[1];
+    const int64_t head_size = C / HEADS;
+
+    float * dst_data = (float *) dst->data;
+    float * state = ((float *) dst->data) + C * T;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    if (ith >= HEADS) {
+        return;
+    }
+
+    const int h_start = (HEADS * ith) / nth;
+    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                (HEADS * (ith + 1)) / nth : HEADS;
+
+    float * r = (float *) dst->src[0]->data;
+    float * w = (float *) dst->src[1]->data;
+    float * k = (float *) dst->src[2]->data;
+    float * v = (float *) dst->src[3]->data;
+    float * a = (float *) dst->src[4]->data;
+    float * b = (float *) dst->src[5]->data;
+
+    int64_t t_stride = HEADS * head_size; // Same to C
+
+    int64_t h_stride = C / HEADS;
+    GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
+    int64_t h_stride_2d = head_size * head_size;
+
+    #if defined(GGML_SIMD)
+        for (int64_t t = 0; t < T; t++) {
+            int64_t t_offset = t * t_stride;
+            int64_t state_offset = head_size * C * (t / (T / n_seqs));
+            float * state_cur = state + state_offset;
+            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                int64_t h_offset = h * h_stride;
+                int64_t t_h_offset = t_offset + h_offset;
+                int64_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t ii = 0; ii < head_size; ii++) {
+                    int64_t t_h_i_offset = t_h_offset + ii;
+                    int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
+
+                    GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
+
+                    float sa = 0;
+                    {
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
+                            for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                                ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
+                                ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
+                                sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
+                            }
+                        }
+                        GGML_F32_VEC_REDUCE(sa, sum);
+                    }
+
+                    GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
+
+                    int64_t j = 0;
+                    GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                    for (; j < head_size; j += GGML_F32_STEP) {
+                        for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                            int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
+                            int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
+
+                            GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
+                            GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
+                            GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
+                            GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
+
+                            k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
+
+                            GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
+                            // kv + s * decay + sa * b
+                            state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
+                            state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
+                            GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
+
+                            result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
+                        }
+                    }
+                    GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
+
+                    // There shouldn't be left-overs though.
+                    for (; j < head_size; j++) {
+                        int64_t t_h_j_offset = t_h_offset + j;
+                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float r_val = r[t_h_j_offset];
+                        float w_val = w[t_h_j_offset];
+                        float k_val = k[t_h_j_offset];
+                        float b_val = b[t_h_j_offset];
+                        float kv_val = v[t_h_i_offset] * k_val;
+
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                        dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                    }
+                }
+            }
+        }
+    #else
+        for (int64_t t = 0; t < T; t++) {
+            int64_t t_offset = t * t_stride;
+            int64_t state_offset = head_size * C * (t / (T / n_seqs));
+            float * state_cur = state + state_offset;
+            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+
+            for (int64_t h = h_start; h < h_end; h++) {
+                int64_t h_offset = h * h_stride;
+                int64_t t_h_offset = t_offset + h_offset;
+                int64_t h_2d_offset = h * h_stride_2d;
+
+                for (int64_t i = 0; i < head_size; i++) {
+                    int64_t t_h_i_offset = t_h_offset + i;
+                    int64_t h_2d_i_offset = h_2d_offset + i * h_stride;
+
+                    float v_val = v[t_h_i_offset];
+
+                    float sa = 0, result = 0;
+                    for (int64_t j = 0; j < head_size; j++) {
+                        sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j];
+                    }
+
+                    for (int64_t j = 0; j < head_size; j++) {
+                        int64_t t_h_j_offset = t_h_offset + j;
+                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                        float r_val = r[t_h_j_offset];
+                        float w_val = w[t_h_j_offset];
+                        float k_val = k[t_h_j_offset];
+                        float b_val = b[t_h_j_offset];
+                        float kv_val = v_val * k_val;
+                        float prev_state_val = state_prev[h_2d_i_j_offset];
+                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                        result += state_cur[h_2d_i_j_offset] * r_val;
+                    }
+                    dst_data[t_h_i_offset] = result;
+                }
+            }
+        }
+    #endif
+}
+
+
+static void ggml_compute_forward_rwkv_wkv7(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_rwkv_wkv7_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_map_unary
 
 static void ggml_compute_forward_map_unary_f32(
@@ -14067,7 +14310,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     }
 
     // extra_buffer op?
-    if (ggml_cpu_extra_compute_forward(params, tensor)) return;
+    if (ggml_cpu_extra_compute_forward(params, tensor)) {
+        return;
+    }
 
     switch (tensor->op) {
         case GGML_OP_DUP:
@@ -14170,6 +14415,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_group_norm(params, tensor);
             } break;
+        case GGML_OP_L2_NORM:
+            {
+                ggml_compute_forward_l2_norm(params, tensor);
+            } break;
         case GGML_OP_MUL_MAT:
             {
                 ggml_compute_forward_mul_mat(params, tensor);
@@ -14357,6 +14606,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_gla(params, tensor);
             } break;
+        case GGML_OP_RWKV_WKV7:
+            {
+                ggml_compute_forward_rwkv_wkv7(params, tensor);
+            } break;
         case GGML_OP_MAP_UNARY:
             {
                 ggml_unary_op_f32_t fun;
@@ -14582,6 +14835,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_CONCAT:
         case GGML_OP_MUL_MAT:
@@ -14648,14 +14902,15 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
+        case GGML_OP_RWKV_WKV6:
+        case GGML_OP_GATED_LINEAR_ATTN:
+        case GGML_OP_RWKV_WKV7:
             {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
-        case GGML_OP_RWKV_WKV6:
-        case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
         case GGML_OP_MAP_CUSTOM1_F32:
@@ -15022,7 +15277,6 @@ struct ggml_cplan ggml_graph_plan(
         size_t cur = 0;
 
         if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) {
-
             switch (node->op) {
                 case GGML_OP_CPY:
                 case GGML_OP_DUP:
@@ -15131,9 +15385,10 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
-                        const int64_t ne00 = node->src[0]->ne[0]; // D
+                        const int64_t ne10 = node->src[1]->ne[0]; // DK
+                        const int64_t ne20 = node->src[2]->ne[0]; // DV
 
-                        cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
+                        cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread)
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -51,11 +51,10 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
         },
         /* .lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
             /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .require_aligned_m_idx = */ true,
         },
         /* .rhs_info = */ {
             /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
@@ -100,7 +99,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
             /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
         },
         /* .rhs_info = */ {
             /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -144,7 +142,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
             /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
         },
         /* .rhs_info = */ {
             /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -189,7 +186,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
             /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
         },
         /* .rhs_info = */ {
             /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
@@ -233,7 +229,6 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
             /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
             /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
-            /* .require_aligned_m_idx = */ false,
         },
         /* .rhs_info = */ {
             /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -40,7 +40,6 @@ struct lhs_packing_info {
     size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
     void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
                       size_t lhs_stride, void* lhs_packed);
-    bool require_aligned_m_idx;
 };
 
 struct rhs_packing_info {

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -124,8 +124,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             size_t sr = kernel->get_sr();
 
             // Calculate number of columns to be processed per thread
-            const bool use_multithread = lhs_info->require_aligned_m_idx && m <= mr ? false : true;
-            const size_t num_m_per_thread = use_multithread ? kai_roundup(m, nth) / nth : m;
+            const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
             const size_t m_start = ith * num_m_per_thread;
             size_t m_to_process = num_m_per_thread;
             if ((m_start + m_to_process) > m) {
@@ -135,11 +134,11 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             if(m_start < m) {
                 // Transform LHS
                 const size_t src_stride        = src1->nb[1];
-                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(0, dst->src[1]->nb[1]));
+                const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
                 const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
                 void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
 
-                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, m_start, src_ptr, src_stride, lhs_packed_ptr);
+                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
             }
 
             ggml_barrier(params->threadpool);

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -55,6 +55,7 @@
 
 #include <atomic>
 #include <array>
+#include <type_traits>
 
 #ifdef _MSC_VER
 #define NOINLINE __declspec(noinline)
@@ -1092,13 +1093,403 @@ class tinyBLAS_Q0_PPC {
        }
     }
 
-    template<typename VA, typename VB>
-    void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    template<typename VA, typename VB, int size>
+    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, VA* vec, std::array<int, size>& comparray) {
         int64_t i, j;
         TA *aoffset = NULL;
         VA *vecOffset = NULL;
         TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
         TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
+        VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
+        VB t1, t2, t3, t4, t5, t6, t7, t8;
+        const vector signed char lowMask = vec_splats((signed char)0xF);
+        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+        const vector signed char v8 = vec_splats((signed char)0x8);
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
+        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
+        vector signed int vsum = {0};
+        vector signed int vsum2 = {0};
+
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
+
+                i = (cols >> 2);
+                if (i > 0) {
+                    do {
+                        c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                        c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                        c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                        c4[1] = reinterpret_cast<VB>(vec_xl(0, aoffset4->qs));
+                        c5[1] = reinterpret_cast<VB>(vec_xl(0, aoffset5->qs));
+                        c6[1] = reinterpret_cast<VB>(vec_xl(0, aoffset6->qs));
+                        c7[1] = reinterpret_cast<VB>(vec_xl(0, aoffset7->qs));
+                        c8[1] = reinterpret_cast<VB>(vec_xl(0, aoffset8->qs));
+
+                        c1[0] = vec_and(c1[1], lowMask);
+                        c1[1] = vec_sr(c1[1], v4);
+                        c1[0] = vec_sub(c1[0], v8);
+                        c1[1] = vec_sub(c1[1], v8);
+                        vsum = vec_sum4s(c1[0], vsum);
+                        vsum2 = vec_sum4s(c1[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c2[0] = vec_and(c2[1], lowMask);
+                        c2[1] = vec_sr(c2[1], v4);
+                        c2[0] = vec_sub(c2[0], v8);
+                        c2[1] = vec_sub(c2[1], v8);
+                        vsum = vec_sum4s(c2[0], vsum);
+                        vsum2 = vec_sum4s(c2[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c3[0] = vec_and(c3[1], lowMask);
+                        c3[1] = vec_sr(c3[1], v4);
+                        c3[0] = vec_sub(c3[0], v8);
+                        c3[1] = vec_sub(c3[1], v8);
+                        vsum = vec_sum4s(c3[0], vsum);
+                        vsum2 = vec_sum4s(c3[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c4[0] = vec_and(c4[1], lowMask);
+                        c4[1] = vec_sr(c4[1], v4);
+                        c4[0] = vec_sub(c4[0], v8);
+                        c4[1] = vec_sub(c4[1], v8);
+                        vsum = vec_sum4s(c4[0], vsum);
+                        vsum2 = vec_sum4s(c4[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c5[0] = vec_and(c5[1], lowMask);
+                        c5[1] = vec_sr(c5[1], v4);
+                        c5[0] = vec_sub(c5[0], v8);
+                        c5[1] = vec_sub(c5[1], v8);
+                        vsum = vec_sum4s(c5[0], vsum);
+                        vsum2 = vec_sum4s(c5[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[4] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c6[0] = vec_and(c6[1], lowMask);
+                        c6[1] = vec_sr(c6[1], v4);
+                        c6[0] = vec_sub(c6[0], v8);
+                        c6[1] = vec_sub(c6[1], v8);
+                        vsum = vec_sum4s(c6[0], vsum);
+                        vsum2 = vec_sum4s(c6[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[5] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c7[0] = vec_and(c7[1], lowMask);
+                        c7[1] = vec_sr(c7[1], v4);
+                        c7[0] = vec_sub(c7[0], v8);
+                        c7[1] = vec_sub(c7[1], v8);
+                        vsum = vec_sum4s(c7[0], vsum);
+                        vsum2 = vec_sum4s(c7[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[6] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        c8[0] = vec_and(c8[1], lowMask);
+                        c8[1] = vec_sr(c8[1], v4);
+                        c8[0] = vec_sub(c8[0], v8);
+                        c8[1] = vec_sub(c8[1], v8);
+                        vsum = vec_sum4s(c8[0], vsum);
+                        vsum2 = vec_sum4s(c8[1], vsum2);
+                        vsum = vec_add(vsum, vsum2);
+                        comparray[7] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                        vsum = vec_splats(0);
+                        vsum2 = vec_splats(0);
+
+                        t1 = vec_perm(c1[0], c2[0], swiz1);
+                        t2 = vec_perm(c1[0], c2[0], swiz2);
+                        t3 = vec_perm(c3[0], c4[0], swiz1);
+                        t4 = vec_perm(c3[0], c4[0], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset);
+                        vec_xst(t6, 0, vecOffset+16);
+                        vec_xst(t7, 0, vecOffset+32);
+                        vec_xst(t8, 0, vecOffset+48);
+
+                        t1 = vec_perm(c1[1], c2[1], swiz1);
+                        t2 = vec_perm(c1[1], c2[1], swiz2);
+                        t3 = vec_perm(c3[1], c4[1], swiz1);
+                        t4 = vec_perm(c3[1], c4[1], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+64);
+                        vec_xst(t6, 0, vecOffset+80);
+                        vec_xst(t7, 0, vecOffset+96);
+                        vec_xst(t8, 0, vecOffset+112);
+
+                        t1 = vec_perm(c5[0], c6[0], swiz1);
+                        t2 = vec_perm(c5[0], c6[0], swiz2);
+                        t3 = vec_perm(c7[0], c8[0], swiz1);
+                        t4 = vec_perm(c7[0], c8[0], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+128);
+                        vec_xst(t6, 0, vecOffset+144);
+                        vec_xst(t7, 0, vecOffset+160);
+                        vec_xst(t8, 0, vecOffset+176);
+
+                        t1 = vec_perm(c5[1], c6[1], swiz1);
+                        t2 = vec_perm(c5[1], c6[1], swiz2);
+                        t3 = vec_perm(c7[1], c8[1], swiz1);
+                        t4 = vec_perm(c7[1], c8[1], swiz2);
+                        t5 = vec_perm(t1, t3, swiz3);
+                        t6 = vec_perm(t1, t3, swiz4);
+                        t7 = vec_perm(t2, t4, swiz3);
+                        t8 = vec_perm(t2, t4, swiz4);
+                        vec_xst(t5, 0, vecOffset+192);
+                        vec_xst(t6, 0, vecOffset+208);
+                        vec_xst(t7, 0, vecOffset+224);
+                        vec_xst(t8, 0, vecOffset+240);
+
+                        aoffset1 += lda;
+                        aoffset2 += lda;
+                        aoffset3 += lda;
+                        aoffset4 += lda;
+                        aoffset5 += lda;
+                        aoffset6 += lda;
+                        aoffset7 += lda;
+                        aoffset8 += lda;
+                        vecOffset += 256;
+                        i--;
+                    } while (i > 0);
+                }
+                j--;
+            } while (j > 0);
+        }
+
+        if (rows & 4) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            aoffset4 = aoffset3 + lda;
+            aoffset += 4 * lda;
+
+            i = (cols >> 2);
+            if (i > 0) {
+                do {
+                    c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                    c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                    c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                    c4[1] = reinterpret_cast<VB>(vec_xl(0, aoffset4->qs));
+
+                    c1[0] = vec_and(c1[1], lowMask);
+                    c1[1] = vec_sr(c1[1], v4);
+                    c1[0] = vec_sub(c1[0], v8);
+                    c1[1] = vec_sub(c1[1], v8);
+                    vsum = vec_sum4s(c1[0], vsum);
+                    vsum2 = vec_sum4s(c1[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c2[0] = vec_and(c2[1], lowMask);
+                    c2[1] = vec_sr(c2[1], v4);
+                    c2[0] = vec_sub(c2[0], v8);
+                    c2[1] = vec_sub(c2[1], v8);
+                    vsum = vec_sum4s(c2[0], vsum);
+                    vsum2 = vec_sum4s(c2[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c3[0] = vec_and(c3[1], lowMask);
+                    c3[1] = vec_sr(c3[1], v4);
+                    c3[0] = vec_sub(c3[0], v8);
+                    c3[1] = vec_sub(c3[1], v8);
+                    vsum = vec_sum4s(c3[0], vsum);
+                    vsum2 = vec_sum4s(c3[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c4[0] = vec_and(c4[1], lowMask);
+                    c4[1] = vec_sr(c4[1], v4);
+                    c4[0] = vec_sub(c4[0], v8);
+                    c4[1] = vec_sub(c4[1], v8);
+                    vsum = vec_sum4s(c4[0], vsum);
+                    vsum2 = vec_sum4s(c4[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats( 0);
+
+                    t1 = vec_perm(c1[0], c2[0], swiz1);
+                    t2 = vec_perm(c1[0], c2[0], swiz2);
+                    t3 = vec_perm(c3[0], c4[0], swiz1);
+                    t4 = vec_perm(c3[0], c4[0], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset);
+                    vec_xst(t6, 0, vecOffset+16);
+                    vec_xst(t7, 0, vecOffset+32);
+                    vec_xst(t8, 0, vecOffset+48);
+
+                    t1 = vec_perm(c1[1], c2[1], swiz1);
+                    t2 = vec_perm(c1[1], c2[1], swiz2);
+                    t3 = vec_perm(c3[1], c4[1], swiz1);
+                    t4 = vec_perm(c3[1], c4[1], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset+64);
+                    vec_xst(t6, 0, vecOffset+80);
+                    vec_xst(t7, 0, vecOffset+96);
+                    vec_xst(t8, 0, vecOffset+112);
+
+                    aoffset1 += lda;
+                    aoffset2 += lda;
+                    aoffset3 += lda;
+                    aoffset4 += lda;
+                    vecOffset += 128;
+                    i--;
+                } while (i > 0);
+            }
+        }
+
+        if (rows & 3) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            i = (cols >> 2);
+            if (i > 0) {
+                do {
+                    switch(rows) {
+                        case 3: c3[1] = reinterpret_cast<VB>(vec_xl(0, aoffset3->qs));
+                        case 2: c2[1] = reinterpret_cast<VB>(vec_xl(0, aoffset2->qs));
+                        case 1: c1[1] = reinterpret_cast<VB>(vec_xl(0, aoffset1->qs));
+                            break;
+                    }
+                    c1[0] = vec_and(c1[1], lowMask);
+                    c1[1] = vec_sr(c1[1], v4);
+                    c1[0] = vec_sub(c1[0], v8);
+                    c1[1] = vec_sub(c1[1], v8);
+                    vsum = vec_sum4s(c1[0], vsum);
+                    vsum2 = vec_sum4s(c1[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c2[0] = vec_and(c2[1], lowMask);
+                    c2[1] = vec_sr(c2[1], v4);
+                    c2[0] = vec_sub(c2[0], v8);
+                    c2[1] = vec_sub(c2[1], v8);
+                    vsum = vec_sum4s(c2[0], vsum);
+                    vsum2 = vec_sum4s(c2[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c3[0] = vec_and(c3[1], lowMask);
+                    c3[1] = vec_sr(c3[1], v4);
+                    c3[0] = vec_sub(c3[0], v8);
+                    c3[1] = vec_sub(c3[1], v8);
+                    vsum = vec_sum4s(c3[0], vsum);
+                    vsum2 = vec_sum4s(c3[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    c4[0] = vec_and(c4[1], lowMask);
+                    c4[1] = vec_sr(c4[1], v4);
+                    c4[0] = vec_sub(c4[0], v8);
+                    c4[1] = vec_sub(c4[1], v8);
+                    vsum = vec_sum4s(c4[0], vsum);
+                    vsum2 = vec_sum4s(c4[1], vsum2);
+                    vsum = vec_add(vsum, vsum2);
+                    comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+                    vsum = vec_splats(0);
+                    vsum2 = vec_splats(0);
+
+                    t1 = vec_perm(c1[0], c2[0], swiz1);
+                    t2 = vec_perm(c1[0], c2[0], swiz2);
+                    t3 = vec_perm(c3[0], c4[0], swiz1);
+                    t4 = vec_perm(c3[0], c4[0], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset);
+                    vec_xst(t6, 0, vecOffset+16);
+                    vec_xst(t7, 0, vecOffset+32);
+                    vec_xst(t8, 0, vecOffset+48);
+
+                    t1 = vec_perm(c1[1], c2[1], swiz1);
+                    t2 = vec_perm(c1[1], c2[1], swiz2);
+                    t3 = vec_perm(c3[1], c4[1], swiz1);
+                    t4 = vec_perm(c3[1], c4[1], swiz2);
+                    t5 = vec_perm(t1, t3, swiz3);
+                    t6 = vec_perm(t1, t3, swiz4);
+                    t7 = vec_perm(t2, t4, swiz3);
+                    t8 = vec_perm(t2, t4, swiz4);
+                    vec_xst(t5, 0, vecOffset+64);
+                    vec_xst(t6, 0, vecOffset+80);
+                    vec_xst(t7, 0, vecOffset+96);
+                    vec_xst(t8, 0, vecOffset+112);
+                    aoffset1 += lda;
+                    aoffset2 += lda;
+                    aoffset3 += lda;
+                    vecOffset += 128;
+                    i--;
+                } while(i > 0);
+            }
+        }
+    }
+
+    template<typename VA, typename VB>
+    void packNormal(const TB* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+        int64_t i, j;
+        TB *aoffset = NULL;
+        VA *vecOffset = NULL;
+        TB *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
+        TB *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
         __vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
         VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0};
         VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0};
@@ -1111,24 +1502,24 @@ class tinyBLAS_Q0_PPC {
         vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
         vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
 
-        aoffset = const_cast<TA*>(a);
+        aoffset = const_cast<TB*>(a);
         vecOffset = vec;
         j = (rows >> 3);
         if (j > 0) {
             do {
-            aoffset1 = aoffset;
-            aoffset2 = aoffset1 + lda;
-            aoffset3 = aoffset2 + lda;
-            aoffset4 = aoffset3 + lda;
-            aoffset5 = aoffset4 + lda;
-            aoffset6 = aoffset5 + lda;
-            aoffset7 = aoffset6 + lda;
-            aoffset8 = aoffset7 + lda;
-            aoffset += 8 * lda;
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
 
-            i = (cols >> 3);
-            if (i > 0) {
-               do {
+                i = (cols >> 3);
+                if (i > 0) {
+                do {
                     C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
                     C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
                     C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
@@ -1156,10 +1547,10 @@ class tinyBLAS_Q0_PPC {
                     t7 = vec_perm(t2, t4, swiz3);
                     t8 = vec_perm(t2, t4, swiz4);
                     if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                     }
                     vec_xst(t5, 0, vecOffset);
                     vec_xst(t6, 0, vecOffset+16);
@@ -1175,10 +1566,10 @@ class tinyBLAS_Q0_PPC {
                     t7 = vec_perm(t2, t4, swiz3);
                     t8 = vec_perm(t2, t4, swiz4);
                     if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                     }
                     vec_xst(t5, 0, vecOffset+64);
                     vec_xst(t6, 0, vecOffset+80);
@@ -1194,10 +1585,10 @@ class tinyBLAS_Q0_PPC {
                     t7 = vec_perm(t2, t4, swiz3);
                     t8 = vec_perm(t2, t4, swiz4);
                     if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                     }
                     vec_xst(t5, 0, vecOffset+128);
                     vec_xst(t6, 0, vecOffset+144);
@@ -1213,10 +1604,10 @@ class tinyBLAS_Q0_PPC {
                     t7 = vec_perm(t2, t4, swiz3);
                     t8 = vec_perm(t2, t4, swiz4);
                     if (flip == true) {
-                       t5 = vec_xor(t5, xor_vector);
-                       t6 = vec_xor(t6, xor_vector);
-                       t7 = vec_xor(t7, xor_vector);
-                       t8 = vec_xor(t8, xor_vector);
+                        t5 = vec_xor(t5, xor_vector);
+                        t6 = vec_xor(t6, xor_vector);
+                        t7 = vec_xor(t7, xor_vector);
+                        t8 = vec_xor(t8, xor_vector);
                     }
                     vec_xst(t5, 0, vecOffset+192);
                     vec_xst(t6, 0, vecOffset+208);
@@ -1240,11 +1631,11 @@ class tinyBLAS_Q0_PPC {
     }
 
     if (rows & 4) {
-            aoffset1 = aoffset;
-            aoffset2 = aoffset1 + lda;
-            aoffset3 = aoffset2 + lda;
-            aoffset4 = aoffset3 + lda;
-            aoffset += 4 * lda;
+        aoffset1 = aoffset;
+        aoffset2 = aoffset1 + lda;
+        aoffset3 = aoffset2 + lda;
+        aoffset4 = aoffset3 + lda;
+        aoffset += 4 * lda;
 
         i = (cols >> 3);
             if (i > 0) {
@@ -1311,7 +1702,7 @@ class tinyBLAS_Q0_PPC {
             aoffset2 = aoffset1 + lda;
             aoffset3 = aoffset2 + lda;
             i = (cols >> 3);
-        if (i > 0) {
+            if (i > 0) {
                 do {
                     switch(rows) {
                         case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
@@ -1527,13 +1918,18 @@ class tinyBLAS_Q0_PPC {
     void KERNEL_4x8(int64_t ii, int64_t jj) {
         vec_t vec_A[8], vec_B[16] = {0};
         acc_t acc_0, acc_1;
-        std::array<int, 4> comparray;
+        std::array<int, 4> comparray {};
         vector float fin_res[8] = {0};
         vector float vs[8] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
             __builtin_mma_xxsetaccz(&acc_0);
             __builtin_mma_xxsetaccz(&acc_1);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 4>((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+            }
             packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
             for(int x = 0; x < 8; x++) {
                 __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1545,15 +1941,17 @@ class tinyBLAS_Q0_PPC {
                     *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
                 }
             }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 4; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 4; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
             }
             compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
             compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
@@ -1565,13 +1963,18 @@ class tinyBLAS_Q0_PPC {
     void KERNEL_8x4(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[8] = {0};
         acc_t acc_0, acc_1;
-        std::array<int, 8> comparray;
+        std::array<int, 8> comparray {};
         vector float fin_res[8] = {0};
         vector float vs[8] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
             __builtin_mma_xxsetaccz(&acc_0);
             __builtin_mma_xxsetaccz(&acc_1);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            }
             packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
             for(int x = 0; x < 8; x++) {
                 __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1582,15 +1985,17 @@ class tinyBLAS_Q0_PPC {
                     *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
                 }
             }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 8; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 8; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
             }
             compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
             compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
@@ -1602,15 +2007,20 @@ class tinyBLAS_Q0_PPC {
     void KERNEL_8x8(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[16] = {0};
         acc_t acc_0, acc_1, acc_2, acc_3;
-        std::array<int, 8> comparray;
+        std::array<int, 8> comparray {};
         vector float fin_res[16] = {0};
         vector float vs[16] = {0};
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
             __builtin_mma_xxsetaccz(&acc_0);
             __builtin_mma_xxsetaccz(&acc_1);
             __builtin_mma_xxsetaccz(&acc_2);
             __builtin_mma_xxsetaccz(&acc_3);
-            packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            if (std::is_same_v<TA, block_q4_0>) {
+               packNormalInt4<int8_t, vector signed char, 8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+            } else {
+               packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+            }
             packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
             for(int x = 0; x < 8; x++) {
                 __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1624,15 +2034,17 @@ class tinyBLAS_Q0_PPC {
                     *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
                 }
             }
-            auto aoffset = A+(ii*lda)+l;
-            for (int i = 0; i < 8; i++) {
-                comparray[i] = 0;
-                int ca = 0;
-                const int8_t *at = aoffset->qs;
-                for (int j = 0; j < 32; j++)
-                    ca += (int)*at++;
-                comparray[i] = ca;
-                aoffset += lda;
+            if (!isAblock_q4) {
+                auto aoffset = A+(ii*lda)+l;
+                for (int i = 0; i < 8; i++) {
+                    comparray[i] = 0;
+                    int ca = 0;
+                    auto *at = aoffset->qs;
+                    for (int j = 0; j < 32; j++)
+                        ca += (int)*at++;
+                    comparray[i] = ca;
+                    aoffset += lda;
+                }
             }
             compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
             compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
@@ -1653,16 +2065,17 @@ class tinyBLAS_Q0_PPC {
         int64_t duty = (tiles + nth - 1) / nth;
         int64_t start = duty * ith;
         int64_t end = start + duty;
-        vec_t vec_A[8], vec_B[8] = {0};
+        vec_t vec_A[8] = {0}, vec_B[8] = {0};
         vector signed int vec_C[4];
         acc_t acc_0;
+        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
 
         if (end > tiles)
             end = tiles;
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            std::array<int, RM> comparray;
+            std::array<int, 4> comparray{};
             vector float res[4] = {0};
             vector float fin_res[4] = {0};
             vector float vs[4] = {0};
@@ -1673,7 +2086,11 @@ class tinyBLAS_Q0_PPC {
                 __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
                 __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
                 __builtin_mma_xxsetaccz(&acc_0);
-                packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                if (isAblock_q4) {
+                   packNormalInt4<int8_t, vector signed char, 4>((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray);
+                } else {
+                   packNormal<int8_t, vector signed char>((const TB*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                }
                 packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
                 for(int x = 0; x < 8; x+=4) {
                     __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
@@ -1687,17 +2104,18 @@ class tinyBLAS_Q0_PPC {
                     }
                 }
                 __builtin_mma_disassemble_acc(vec_C, &acc_0);
-                auto aoffset = A+(ii*lda)+l;
-                for (int i = 0; i < RM; i++) {
-                    comparray[i] = 0;
-                    int ca = 0;
-                    const int8_t *at = aoffset->qs;
-                    for (int j = 0; j < 32; j++)
-                        ca += (int)*at++;
-                    comparray[i] = ca;
-                    aoffset += lda;
+                if (!isAblock_q4) {
+                    auto aoffset = A+(ii*lda)+l;
+                    for (int i = 0; i < RM; i++) {
+                        comparray[i] = 0;
+                        int ca = 0;
+                        auto *at = aoffset->qs;
+                        for (int j = 0; j < 32; j++)
+                            ca += (int)*at++;
+                        comparray[i] = ca;
+                        aoffset += lda;
+                    }
                 }
-
                 for (int i = 0; i < RM; i++) {
                     CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0));
                     res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
@@ -2013,6 +2431,7 @@ class tinyBLAS_PPC {
             }
         }
     }
+
     void KERNEL_4x4(int64_t ii, int64_t jj) {
         vec_t vec_A[4], vec_B[4], vec_C[4];
         acc_t acc_0;
@@ -2259,15 +2678,27 @@ class tinyBLAS_PPC {
             vec_t vec_C[4];
             acc_t acc_0;
             __builtin_mma_xxsetaccz(&acc_0);
-            vec_t vec_A[4], vec_B[4];
+            vec_t vec_A[4] {0}, vec_B[4] = {0};
             for (int l=0; l<k; l+=4) {
-                if (RN >= 4 && RM == 1) {
+                /* 'GEMV Forwarding' concept is used in first two conditional loops.
+                 * when one of the matrix has a single row/column, the elements are
+                 * broadcasted, instead of using packing routine to prepack the
+                 * matrix elements.
+                 */
+                if (RM == 1) {
                     TA* a = const_cast<TA*>(A+(ii)*lda+l);
-                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
+                    packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
                     vec_A[0] = (vec_t)vec_xl(0,a);
                     vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1));
                     vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2));
                     vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3));
+                } else if (RN == 1) {
+                    packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
+                    TB* b = const_cast<TB*>(B+(jj)*ldb+l);
+                    vec_B[0] = (vec_t)vec_xl(0,b);
+                    vec_B[1] = (vec_t)vec_splats(*((TB*)&vec_B+1));
+                    vec_B[2] = (vec_t)vec_splats(*((TB*)&vec_B+2));
+                    vec_B[3] = (vec_t)vec_splats(*((TB*)&vec_B+3));
                 } else {
                     packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
                     packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
@@ -2371,8 +2802,10 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
     assert(params->ith < params->nth);
 
     // only enable sgemm for prompt processing
+#if !defined(__MMA__)
     if (n < 2)
         return false;
+#endif
 
     if (Ctype != GGML_TYPE_F32)
         return false;
@@ -2503,8 +2936,8 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
-
 #elif defined(__MMA__)
+    //TO-DO: Remove this condition once gemv forwarding is enabled.
         if (n < 8 && n != 4)
            return false;
         if (m < 8 && m != 4)
@@ -2516,7 +2949,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
-
 #else
         return false;
 #endif
@@ -2541,6 +2973,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
+#elif defined(__MMA__)
+    //TO-DO: Remove this condition once gemv forwarding is enabled.
+        if (n < 8 && n != 4)
+           return false;
+        if (m < 8 && m != 4)
+           return false;
+        tinyBLAS_Q0_PPC<block_q4_0, block_q8_0, float> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            params->ith, params->nth};
+        tb.matmul(m, n);
+        return true;
 #else
         return false;
 #endif

ggml/src/ggml-cuda/common.cuh

@@ -41,15 +41,18 @@
 #define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 #define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
 
-#define GGML_CUDA_CC_PASCAL       600
-#define GGML_CUDA_CC_DP4A         610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define GGML_CUDA_CC_VOLTA        700
-#define GGML_CUDA_CC_TURING       750
-#define GGML_CUDA_CC_AMPERE       800
-#define GGML_CUDA_CC_ADA_LOVELACE 890
-#define GGML_CUDA_CC_OFFSET_AMD   0x1000000
+#define GGML_CUDA_CC_PASCAL          600
+#define GGML_CUDA_CC_DP4A            610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define GGML_CUDA_CC_VOLTA           700
+#define GGML_CUDA_CC_TURING          750
+#define GGML_CUDA_CC_AMPERE          800
+#define GGML_CUDA_CC_ADA_LOVELACE    890
+#define GGML_CUDA_CC_OFFSET_AMD      0x1000000
+#define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000
+#define GGML_CUDA_CC_IS_NVIDIA(cc)   (cc < GGML_CUDA_CC_OFFSET_MTHREADS)
 
-// GCN/CNDA, wave size is 64
+// AMD
+// GCN/CDNA, wave size is 64
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
@@ -57,21 +60,32 @@
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x910)  // MI210, minimum acc register renameing
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
 
-// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
+// RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
 
 #define GGML_CUDA_CC_IS_AMD(cc)   (cc >= GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
 #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
 
-#define GGML_CUDA_CC_QY1        210
-#define GGML_CUDA_CC_QY2        220
+// Moore Threads
+#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
+
+#define GGML_CUDA_CC_QY1  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG   (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD
+
+#define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
+#define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT)
+#define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
@@ -197,9 +211,9 @@ typedef float2 dfloat2;
 #define FP16_MMA_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -209,21 +223,21 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
 
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
 }
 
 static bool fast_fp16_available(const int cc) {
-    return fp16_available(cc) && cc != 610;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fast_fp16_hardware_available(const int cc) {
-    return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
 }
 
 // Any FP16 tensor core instructions are available for ggml code.
@@ -231,20 +245,20 @@ static bool fp16_mma_available(const int cc) {
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
     return false;
 #else
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ||
-        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA ||
-        GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
 static bool new_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
 
 static bool cp_async_available(const int cc) {
@@ -397,7 +411,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4)
     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(RDNA1) || defined(__gfx900__)
     int tmp1;
@@ -678,7 +692,7 @@ struct ggml_tensor_extra_gpu {
 };
 
 
-#if ((CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)) || defined(GGML_HIP_GRAPHS)
+#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS))
 #define USE_CUDA_GRAPH
 #endif
 

ggml/src/ggml-cuda/fattn-common.cuh

@@ -606,48 +606,47 @@ static __global__ void flash_attn_stream_k_fixup(
     *dst = dst_val / rowsum;
 }
 
-template<int D, int parallel_blocks> // D == head size
+template<int D> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
+        float * __restrict__ dst,
+        const int parallel_blocks) {
+    VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.z*blockIdx.x;
+    dst       +=                 D * gridDim.z*blockIdx.x;
 
     const int tid = threadIdx.x;
     __builtin_assume(tid < D);
 
-    __shared__ float2 meta[parallel_blocks];
+    extern __shared__ float2 meta[];
     if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid];
     }
 
     __syncthreads();
 
     float kqmax = meta[0].x;
-#pragma unroll
     for (int l = 1; l < parallel_blocks; ++l) {
         kqmax = max(kqmax, meta[l].x);
     }
 
     float VKQ_numerator   = 0.0f;
     float VKQ_denominator = 0.0f;
-#pragma unroll
     for (int l = 0; l < parallel_blocks; ++l) {
         const float diff = meta[l].x - kqmax;
         const float KQ_max_scale = expf(diff);
         const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
         *((uint32_t *) &KQ_max_scale) &= ftz_mask;
 
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid];
         VKQ_denominator += KQ_max_scale * meta[l].y;
     }
 
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+    dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
 }
 
 static void on_no_fattn_vec_case(const int D) {
@@ -671,12 +670,10 @@ static void on_no_fattn_vec_case(const int D) {
     }
 }
 
-// parallel_blocks == 0 is stream-k decomposition
-template <int D, int ncols1, int ncols2, int parallel_blocks, int KQ_stride>
+template <int D, int ncols1, int ncols2, int KQ_stride>
 void launch_fattn(
-    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel,
-    const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V,
-    const int warp_size = WARP_SIZE
+    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
+    const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
@@ -748,12 +745,14 @@ void launch_fattn(
         nb23 = nb23*bs*sizeof(half)/ts;
     }
 
+    int parallel_blocks = 1;
+
     const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
     const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
 
     const dim3 block_dim(warp_size, nwarps, 1);
     dim3 blocks_num;
-    if (parallel_blocks == 0) {
+    if (stream_k) {
         // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
         const int max_blocks = 2*nsm;
         const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
@@ -769,9 +768,43 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float));
     } else {
-        blocks_num.x = parallel_blocks*ntiles_x;
-        blocks_num.y = Q->ne[2];
-        blocks_num.z = Q->ne[3];
+        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
+        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+
+        int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
+        CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+
+        // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
+        parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);
+
+        // parallel_blocks must not be larger than what the tensor size allows:
+        parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
+
+        // If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects.
+        // Test whether parallel_blocks can be set to a higher value for better efficiency.
+        const int blocks_per_wave = nsm * max_blocks_per_sm;
+        int nwaves_best = 0;
+        int efficiency_percent_best = 0;
+        for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) {
+            const int nblocks_total = ntiles_total * parallel_blocks_test;
+            const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave;
+            const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);
+
+            // Stop trying configurations with more waves if we already have good efficiency to avoid excessive overhead.
+            if (efficiency_percent_best >= 90 && nwaves > nwaves_best) {
+                break;
+            }
+
+            if (efficiency_percent > efficiency_percent_best) {
+                nwaves_best = nwaves;
+                efficiency_percent_best = efficiency_percent;
+                parallel_blocks = parallel_blocks_test;
+            }
+        }
+
+        blocks_num.x = ntiles_x;
+        blocks_num.y = parallel_blocks;
+        blocks_num.z = Q->ne[2]*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -803,7 +836,7 @@ void launch_fattn(
         K_data,
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
-        (parallel_blocks) > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
+        !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
@@ -815,7 +848,7 @@ void launch_fattn(
     );
     CUDA_CHECK(cudaGetLastError());
 
-    if constexpr (parallel_blocks == 0) {
+    if (stream_k) {
         if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
             const dim3 block_dim_combine(D, 1, 1);
             const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};
@@ -824,13 +857,14 @@ void launch_fattn(
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
                 ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
         }
-    } else if constexpr (parallel_blocks > 1) {
+    } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(D, 1, 1);
-        const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+        const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
+        const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
-        flash_attn_combine_results<D, parallel_blocks>
-            <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+        flash_attn_combine_results<D>
+            <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
+            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
     }
     CUDA_CHECK(cudaGetLastError());
 }

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -970,7 +970,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
         fattn_kernel = flash_attn_ext_f16<D, ncols1, ncols2, nwarps, KQ_per_iter, ntiles, use_logit_softcap>;
     }
 
-    launch_fattn<D, ncols1, ncols2, 0, KQ_per_iter>(ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, true, true);
+    launch_fattn<D, ncols1, ncols2, KQ_per_iter>
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true);
 }
 
 

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f16(
 
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half   * maskh = (const half   *)  mask + ne11*ic0;
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -105,8 +104,7 @@ static __global__ void flash_attn_tile_ext_f16(
 
     __syncthreads();
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         half kqmax_new[ncols/nwarps];
@@ -271,16 +269,16 @@ static __global__ void flash_attn_tile_ext_f16(
             const int i0 = i00 + 2*threadIdx.x;
 
             half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val /= __half2half2(kqsum_j);
             }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val);
         }
 
-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -288,7 +286,7 @@ static __global__ void flash_attn_tile_ext_f16(
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
 
-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
@@ -296,15 +294,17 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    D             = 64;
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         case 128: {
             constexpr int    D             = 128;
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         default: {
             GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -324,37 +324,22 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten
 
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     }
 }

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f32(
 
     // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half   * maskh = (const half   *)  mask + ne11*ic0;
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
 
@@ -103,8 +102,7 @@ static __global__ void flash_attn_tile_ext_f32(
 
     __syncthreads();
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new[ncols/nwarps];
@@ -269,17 +267,17 @@ static __global__ void flash_attn_tile_ext_f32(
             const int i0 = i00 + 2*threadIdx.x;
 
             float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val.x /= kqsum_j;
                 dst_val.y /= kqsum_j;
             }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
         }
 
-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -287,7 +285,7 @@ static __global__ void flash_attn_tile_ext_f32(
 #endif // FLASH_ATTN_AVAILABLE
 }
 
-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
@@ -295,15 +293,17 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    D             = 64;
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         case 128: {
             constexpr int    D             = 128;
             constexpr int    nwarps        = 8;
             constexpr size_t nbytes_shared = 0;
-            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         default: {
             GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -320,37 +320,22 @@ void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_ten
 
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     }
 }

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,17 +55,16 @@ static __global__ void flash_attn_vec_ext_f16(
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
     constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio);
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio);
 
     const half * maskh = (const half   *)  mask + ne11*ic0;
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -172,8 +171,7 @@ static __global__ void flash_attn_vec_ext_f16(
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
@@ -283,29 +281,29 @@ static __global__ void flash_attn_vec_ext_f16(
         kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]);
 
         half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
    NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
 
-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, type_K, type_V, use_logit_softcap>;
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -325,65 +323,48 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     }
 }
 

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,16 +55,15 @@ static __global__ void flash_attn_vec_ext_f32(
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
     constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
     const half * maskh = (const half   *)  mask + ne11*ic0;
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -167,8 +166,7 @@ static __global__ void flash_attn_vec_ext_f32(
 
     float VKQ[ncols] = {0.0f};
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new_arr[ncols];
@@ -268,29 +266,29 @@ static __global__ void flash_attn_vec_ext_f32(
         kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
 
         float dst_val = VKQ[j_VKQ];
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
 
-template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, type_K, type_V, use_logit_softcap>;
     constexpr bool need_f16_K = D != 128;
     constexpr bool need_f16_V = D != 128 && D != 64;
     constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -307,65 +305,48 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     }
 }
 

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -18,7 +18,7 @@ namespace wmma = rocwmma;
 #endif // FP16_MMA_AVAILABLE
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
-template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
+template<int D, int ncols, int nwarps, int VKQ_stride, typename KQ_acc_t, bool use_logit_softcap>
 __launch_bounds__(nwarps*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
@@ -67,8 +67,7 @@ static __global__ void flash_attn_ext_f16(
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
-    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+    const int ic0 = ncols*blockIdx.x; // Index of the first Q/QKV column to work on.
 
     static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
     static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
@@ -91,16 +90,16 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
-    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.z              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.z / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
     const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
     const half2 slope2 = make_half2(slopef, slopef);
 
@@ -176,7 +175,7 @@ static __global__ void flash_attn_ext_f16(
     __syncthreads();
 
     // Iterate over ne11 == previous tokens:
-    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
         // Calculate tile of KQ:
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
@@ -395,7 +394,7 @@ static __global__ void flash_attn_ext_f16(
         if (ic0 + j_VKQ >= ne01) {
             return;
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
 
         float KQ_rowsum_j;
         if (std::is_same<KQ_acc_t, float>::value) {
@@ -411,13 +410,13 @@ static __global__ void flash_attn_ext_f16(
                 break;
             }
             float dst_val = VKQ[j_VKQ*D_padded + i];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val /= KQ_rowsum_j;
             }
-            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+            dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val;
         }
 
-        if (parallel_blocks == 1 || threadIdx.x != 0) {
+        if (gridDim.y == 1 || threadIdx.x != 0) {
             continue;
         }
 
@@ -428,7 +427,7 @@ static __global__ void flash_attn_ext_f16(
             dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
         }
         dst_meta_val.y = KQ_rowsum_j;
-        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+        dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
     }
 #else
    NO_DEVICE_CODE;
@@ -462,60 +461,26 @@ static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 template <int D, int cols_per_block, typename KQ_acc_t>
 void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
 
     constexpr int nwarps = 4;
 
     constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
-    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
-    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
 
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
-    if (4*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 4;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        }
-        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    if (2*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 2;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            fattn_kernel = flash_attn_ext_f16<
-                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
-        }
-        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    constexpr int parallel_blocks = 1;
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
         fattn_kernel = flash_attn_ext_f16<
-            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     } else {
         constexpr bool use_logit_softcap = true;
         fattn_kernel = flash_attn_ext_f16<
-            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     }
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
 }
 
 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/fattn.cu

@@ -253,7 +253,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-    if (cc >= GGML_CUDA_CC_OFFSET_AMD) {
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
 #if defined(GGML_HIP_ROCWMMA_FATTN)
         if (fp16_mma_available(cc)) {
             ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
@@ -281,13 +281,13 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     if (!fp16_mma_available(cc)) {
         if (prec == GGML_PREC_DEFAULT) {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                 ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
             } else {
                 ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
             }
         } else {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                 ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
             } else {
                 ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
@@ -296,17 +296,17 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
         return;
     }
 
-    const int gqa_ratio = Q->ne[2] / K->ne[2];
-    const bool mma_fast_for_bs1 = fp16_mma_available(cc) && gqa_ratio % 2 == 0 &&
-        K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16 && mask;
-    if (Q->ne[1] == 1 && Q->ne[0] % (2*warp_size) == 0 && !mma_fast_for_bs1) {
+    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
+    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool can_use_vector_kernel = (Q->ne[0] % (2*warp_size) == 0) && (prec == GGML_PREC_DEFAULT || Q->ne[0] <= 128);
+    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-            return;
-        } else if(Q->ne[0] <= 128) {
+        } else {
             ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            return;
         }
+        return;
     }
 
     // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -36,7 +36,7 @@
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
-#include "ggml-cuda/wkv6.cuh"
+#include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
 #include "ggml.h"
 
@@ -262,9 +262,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
                       id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff,
                       device_vmm ? "yes" : "no", prop.warpSize);
 #elif defined(GGML_USE_MUSA)
-        // TODO: refine the .cc to reflect MUSA's actual CC capabilities
+        // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs.
+        info.devices[id].warp_size = 32;
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
-        info.devices[id].cc = 100*prop.major + 10*prop.minor;
+        info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100;
+        info.devices[id].cc += prop.minor * 0x10;
         GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n",
                         id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
 #else
@@ -1186,11 +1188,11 @@ static void ggml_cuda_op_mul_mat_cublas(
     // ldc == nrows of the matrix that cuBLAS writes into
     int64_t ldc = id == ctx.device ? ne0 : row_diff;
 
-    const int compute_capability = ggml_cuda_info().devices[id].cc;
+    const int cc = ggml_cuda_info().devices[id].cc;
 
     const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
 
-    if (compute_capability >= GGML_CUDA_CC_VOLTA && use_fp16) {
+    if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
@@ -1214,7 +1216,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
         CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
 
-        if (GGML_CUDA_CC_IS_CDNA(compute_capability)) {
+        if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
             const float alpha = 1.0f;
             const float beta = 0.0f;
             CUBLAS_CHECK(
@@ -1757,7 +1759,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         beta  = &beta_f32;
     }
 
-    if (GGML_CUDA_CC_IS_CDNA(ggml_cuda_info().devices[ctx.device].cc)) {
+    int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
         cu_compute_type = CUBLAS_COMPUTE_32F;
         alpha = &alpha_f32;
         beta  = &beta_f32;
@@ -1834,7 +1838,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     }
 #endif
 
-    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
         const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
         to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
     }
@@ -2196,6 +2200,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GROUP_NORM:
             ggml_cuda_op_group_norm(ctx, dst);
             break;
+        case GGML_OP_L2_NORM:
+            ggml_cuda_op_l2_norm(ctx, dst);
+            break;
         case GGML_OP_CONCAT:
             ggml_cuda_op_concat(ctx, dst);
             break;
@@ -2304,6 +2311,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GATED_LINEAR_ATTN:
             ggml_cuda_op_gated_linear_attn(ctx, dst);
             break;
+        case GGML_OP_RWKV_WKV7:
+            ggml_cuda_op_rwkv_wkv7(ctx, dst);
+            break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             ggml_cuda_cross_entropy_loss_back(ctx, dst);
             break;
@@ -2610,13 +2620,15 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx,
 
 static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 
+#if CUDART_VERSION >= 12000
     cudaGraphExecUpdateResultInfo result_info;
-#ifdef __HIP_PLATFORM_AMD__
-    hipGraphNode_t errorNode;
-    hipError_t stat = hipGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
-#else
     cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
-#endif
+#else
+    cudaGraphNode_t errorNode;
+    cudaGraphExecUpdateResult result_info;
+    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
+#endif // CUDART_VERSION >= 12000
+
     if (stat == cudaErrorGraphExecUpdateFailure) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__);
@@ -3159,6 +3171,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             break;
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_L2_NORM:
             return true;
         case GGML_OP_RMS_NORM_BACK:
             return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
@@ -3213,11 +3226,22 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_GATED_LINEAR_ATTN:
+        case GGML_OP_RWKV_WKV7:
             return true;
         case GGML_OP_FLASH_ATTN_EXT: {
 #ifndef FLASH_ATTN_AVAILABLE
             return false;
 #endif // FLASH_ATTN_AVAILABLE
+            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
+                // different head sizes of K and V are not supported yet
+                return false;
+            }
+            if (op->src[0]->ne[0] == 192) {
+                return false;
+            }
+            if (op->src[0]->ne[3] != 1) {
+                return false;
+            }
             if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                 return false;
             }

ggml/src/ggml-cuda/mmq.cu

@@ -27,8 +27,8 @@ void ggml_cuda_op_mul_mat_q(
     // The stream-k decomposition is only faster for recent NVIDIA GPUs.
     // Also its fixup needs to allocate a temporary buffer in the memory pool.
     // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA &&
-        cc < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
     const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
@@ -145,9 +145,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     return true;
 #endif //GGML_CUDA_FORCE_MMQ
 
-    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
         return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }

ggml/src/ggml-cuda/mmq.cuh

@@ -90,7 +90,7 @@ struct tile_x_sizes {
 
 static int get_mmq_x_max_host(const int cc) {
     return new_mma_available(cc) ? 128 :
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ?
+        GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
 #else
@@ -123,8 +123,8 @@ static constexpr __device__ int get_mmq_x_max_device() {
 }
 
 static int get_mmq_y_host(const int cc) {
-    return cc >= GGML_CUDA_CC_OFFSET_AMD ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) :
-        (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? 128 : 64);
+    return GGML_CUDA_CC_IS_AMD(cc) ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) :
+        ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64);
 }
 
 static constexpr __device__ int get_mmq_y_device() {
@@ -2577,9 +2577,9 @@ static __device__ void mul_mat_q_process_tile(
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
+#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
 #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)
@@ -2772,14 +2772,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
     static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
     if (!shmem_limit_raised[id]) {
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         shmem_limit_raised[id] = true;
     }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
     const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
@@ -2832,7 +2832,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
     const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD;
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
 
     int mmq_x_best  = 0;
     int nparts_best = INT_MAX;

ggml/src/ggml-cuda/mmvq.cu

@@ -54,7 +54,7 @@ enum mmvq_parameter_table_id {
 };
 
 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
-#if defined(RDNA2) || defined(RDNA3)
+#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4)
     return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
     return MMVQ_PARAMETERS_GCN;
@@ -64,7 +64,7 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
 }
 
 static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
-    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) {
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
         return MMVQ_PARAMETERS_RDNA2;
     }
     if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {

ggml/src/ggml-cuda/norm.cu

@@ -201,6 +201,85 @@ static __global__ void rms_norm_back_f32(
     }
 }
 
+// template <int block_size>
+// static __global__ void l2_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+//     const int row = blockIdx.x*blockDim.y + threadIdx.y;
+//     const int tid = threadIdx.x;
+
+//     float tmp = 0.0f; // partial sum for thread in warp
+
+//     for (int col = tid; col < ncols; col += block_size) {
+//         const float xi = x[row*ncols + col];
+//         tmp += xi * xi;
+//     }
+
+//     // sum up partial sums
+//     tmp = warp_reduce_sum(tmp);
+//     if (block_size > WARP_SIZE) {
+//         __shared__ float s_sum[32];
+//         int warp_id = threadIdx.x / WARP_SIZE;
+//         int lane_id = threadIdx.x % WARP_SIZE;
+//         if (lane_id == 0) {
+//             s_sum[warp_id] = tmp;
+//         }
+//         __syncthreads();
+//         tmp = s_sum[lane_id];
+//         tmp = warp_reduce_sum(tmp);
+//     }
+
+//     // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html
+//     const float scale = rsqrtf(fmaxf(tmp, eps * eps));
+
+//     for (int col = tid; col < ncols; col += block_size) {
+//         dst[row*ncols + col] = scale * x[row*ncols + col];
+//     }
+// }
+
+template <int block_size>
+static __global__ void l2_norm_f32(
+        const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel,
+        const int64_t stride_sample, const float eps) {
+    const int nrows     = gridDim.x;
+    const int nchannels = gridDim.y;
+
+    const int row       = blockIdx.x;
+    const int channel   = blockIdx.y;
+    const int sample    = blockIdx.z;
+    const int tid       = threadIdx.x;
+
+    x   += sample*stride_sample + channel*stride_channel + row*stride_row;
+    dst += ((sample*nchannels + channel)*nrows + row)*ncols;
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[col];
+        tmp += xi * xi;
+    }
+
+    // sum up partial sums
+    tmp = warp_reduce_sum(tmp);
+    if constexpr (block_size > WARP_SIZE) {
+        static_assert(block_size == 1024, "unexpected block_size");
+        __shared__ float s_sum[32];
+        const int warp_id = threadIdx.x / WARP_SIZE;
+        const int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html
+    const float scale = rsqrtf(fmaxf(tmp, eps * eps));
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[col] = scale * x[col];
+    }
+}
+
 static void norm_f32_cuda(
         const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
         const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
@@ -248,6 +327,19 @@ static void rms_norm_back_f32_cuda(const float * grad, const float * xf, float *
     }
 }
 
+static void l2_norm_f32_cuda(
+        const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
+        const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) {
+    const dim3 blocks_num(nrows, nchannels, nsamples);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        l2_norm_f32<WARP_SIZE><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        l2_norm_f32<1024><<<blocks_num, block_dims, 0, stream>>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps);
+    }
+}
+
 void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *) src0->data;
@@ -340,3 +432,27 @@ void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * d
 
     rms_norm_back_f32_cuda(grad_d, src0f_d, dst_d, ne00, nrows, eps, stream);
 }
+
+void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *) src0->data;
+    float * dst_d = (float *) dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    const size_t ts0 = ggml_type_size(src0->type);
+    GGML_ASSERT(nb00 == ts0);
+    const int64_t s01 = nb01 / ts0;
+    const int64_t s02 = nb02 / ts0;
+    const int64_t s03 = nb03 / ts0;
+
+    l2_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream);
+}

ggml/src/ggml-cuda/norm.cuh

@@ -7,3 +7,5 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/vendors/hip.h

@@ -112,7 +112,7 @@
 #define cudaGraphExecDestroy hipGraphExecDestroy
 #define cudaGraphLaunch hipGraphLaunch
 #define cudaErrorGraphExecUpdateFailure hipErrorGraphExecUpdateFailure
-#define cudaGraphExecUpdateResultInfo hipGraphExecUpdateResult
+#define cudaGraphExecUpdateResult hipGraphExecUpdateResult
 #define cudaGraphNodeType hipGraphNodeType
 #define cudaGraphNodeTypeKernel hipGraphNodeTypeKernel
 #define cudaGraphInstantiate hipGraphInstantiate
@@ -129,6 +129,7 @@
 #define cudaGraph_t hipGraph_t
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
@@ -150,6 +151,10 @@
 #define CDNA
 #endif
 
+#if defined(__GFX12__)
+#define RDNA4
+#endif
+
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
     defined(__gfx1150__) || defined(__gfx1151__)
 #define RDNA3

ggml/src/ggml-cuda/vendors/musa.h

@@ -119,7 +119,7 @@
 #define cudaGraphExecDestroy musaGraphExecDestroy
 #define cudaGraphExec_t musaGraphExec_t
 #define cudaGraphExecUpdate musaGraphExecUpdate
-#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphExecUpdateResult musaGraphExecUpdateResult
 #define cudaGraphGetNodes musaGraphGetNodes
 #define cudaGraphInstantiate musaGraphInstantiate
 #define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
@@ -132,6 +132,8 @@
 #define cudaGraph_t musaGraph_t
 #define cudaKernelNodeParams musaKernelNodeParams
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamBeginCapture musaStreamBeginCapture
 #define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
 typedef mt_bfloat16 nv_bfloat16;

ggml/src/ggml-cuda/wkv.cu

@@ -0,0 +1,199 @@
+#include "common.cuh"
+#include "wkv.cuh"
+
+template <int block_size>
+static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
+    const int tid = threadIdx.x;
+    const int bid = blockIdx.x;
+
+    const int head_size = block_size;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    float state[head_size];
+    __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
+    }
+
+    __syncthreads();
+    _tf[tid] = tf[head_i * head_size + tid];
+    __syncthreads();
+
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
+        __syncthreads();
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        __syncthreads();
+
+        const float _v = v[t];
+        float y = 0;
+        for (int j = 0; j < head_size; j += 4) {
+            const float4& k = (float4&)(_k[j]);
+            const float4& r = (float4&)(_r[j]);
+            const float4& tf = (float4&)(_tf[j]);
+            const float4& td = (float4&)(_td[j]);
+            float4& s = (float4&)(state[j]);
+            float4 kv;
+
+            kv.x = k.x * _v;
+            kv.y = k.y * _v;
+            kv.z = k.z * _v;
+            kv.w = k.w * _v;
+
+            y += r.x * (tf.x * kv.x + s.x);
+            y += r.y * (tf.y * kv.y + s.y);
+            y += r.z * (tf.z * kv.z + s.z);
+            y += r.w * (tf.w * kv.w + s.w);
+
+            s.x = s.x * td.x + kv.x;
+            s.y = s.y * td.y + kv.y;
+            s.z = s.z * td.z + kv.z;
+            s.w = s.w * td.w + kv.w;
+        }
+        dst[t] = y;
+    }
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
+    }
+}
+
+template <int block_size>
+static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
+    const int tid = threadIdx.x;
+    const int bid = blockIdx.x;
+
+    const int head_size = block_size;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    float state[head_size];
+    __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
+
+#ifndef GGML_USE_MUSA
+    #pragma unroll
+#endif
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + tid * head_size + i];
+    }
+
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
+        __syncthreads();
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        __syncthreads();
+
+        float sa = 0;
+        #pragma unroll
+        for (int j = 0; j < head_size; j += 4)
+        {
+            const float4& a = (float4&)(_a[j]);
+            const float4& s = (float4&)(state[j]);
+            sa += a.x * s.x;
+            sa += a.y * s.y;
+            sa += a.z * s.z;
+            sa += a.w * s.w;
+        }
+
+        const float _v = v[t];
+        float y = 0;
+        for (int j = 0; j < head_size; j += 4) {
+            const float4& r = (float4&)(_r[j]);
+            const float4& w = (float4&)(_w[j]);
+            const float4& k = (float4&)(_k[j]);
+            const float4& b = (float4&)(_b[j]);
+            float4& s = (float4&)(state[j]);
+            float4 kv;
+
+            kv.x = k.x * _v;
+            kv.y = k.y * _v;
+            kv.z = k.z * _v;
+            kv.w = k.w * _v;
+
+            s.x = s.x * w.x + kv.x + sa * b.x;
+            s.y = s.y * w.y + kv.y + sa * b.y;
+            s.z = s.z * w.z + kv.z + sa * b.z;
+            s.w = s.w * w.w + kv.w + sa * b.w;
+
+            y += s.x * r.x;
+            y += s.y * r.y;
+            y += s.z * r.z;
+            y += s.w * r.w;
+        }
+        dst[t] = y;
+    }
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + tid * head_size + i] = state[i];
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const float * k_d  = (const float *)dst->src[0]->data;
+    const float * v_d  = (const float *)dst->src[1]->data;
+    const float * r_d  = (const float *)dst->src[2]->data;
+    const float * tf_d = (const float *)dst->src[3]->data;
+    const float * td_d = (const float *)dst->src[4]->data;
+    const float * s_d  = (const float *)dst->src[5]->data;
+
+    const int64_t B = dst->src[5]->ne[1];
+    const int64_t T = dst->src[0]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[1];
+
+    float * dst_d = (float *)dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
+
+    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+        rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+    } else {
+        rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const float * r_d = (const float *)dst->src[0]->data;
+    const float * w_d = (const float *)dst->src[1]->data;
+    const float * k_d = (const float *)dst->src[2]->data;
+    const float * v_d = (const float *)dst->src[3]->data;
+    const float * a_d = (const float *)dst->src[4]->data;
+    const float * b_d = (const float *)dst->src[5]->data;
+    const float * s_d = (const float *)dst->src[6]->data;
+
+    const int64_t B = dst->src[6]->ne[1];
+    const int64_t T = dst->src[0]->ne[2];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[1];
+
+    float * dst_d = (float *)dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
+
+    if (C / H == CUDA_WKV_BLOCK_SIZE) {
+        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    } else {
+        rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
+    }
+}

ggml/src/ggml-cuda/wkv.cuh

@@ -3,3 +3,5 @@
 #define CUDA_WKV_BLOCK_SIZE 64
 
 void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/wkv6.cu

@@ -1,89 +0,0 @@
-#include "common.cuh"
-#include "wkv6.cuh"
-
-static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
-    const int tid = threadIdx.x;
-    const int bid = blockIdx.x;
-
-    const int head_size = CUDA_WKV_BLOCK_SIZE;
-    const int batch_i = bid / H;
-    const int head_i = bid % H;
-    const int state_size = C * head_size;
-    const int n_seq_tokens = T / B;
-
-    float state[head_size];
-    __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
-
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
-    }
-
-    __syncthreads();
-    _tf[tid] = tf[head_i * head_size + tid];
-    __syncthreads();
-
-    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
-        __syncthreads();
-        _k[tid] = k[t];
-        _r[tid] = r[t];
-        _td[tid] = td[t];
-        __syncthreads();
-
-        const float _v = v[t];
-        float y = 0;
-        for (int j = 0; j < head_size; j += 4) {
-            const float4& k = (float4&)(_k[j]);
-            const float4& r = (float4&)(_r[j]);
-            const float4& tf = (float4&)(_tf[j]);
-            const float4& td = (float4&)(_td[j]);
-            float4& s = (float4&)(state[j]);
-            float4 kv;
-
-            kv.x = k.x * _v;
-            kv.y = k.y * _v;
-            kv.z = k.z * _v;
-            kv.w = k.w * _v;
-
-            y += r.x * (tf.x * kv.x + s.x);
-            y += r.y * (tf.y * kv.y + s.y);
-            y += r.z * (tf.z * kv.z + s.z);
-            y += r.w * (tf.w * kv.w + s.w);
-
-            s.x = s.x * td.x + kv.x;
-            s.y = s.y * td.y + kv.y;
-            s.z = s.z * td.z + kv.z;
-            s.w = s.w * td.w + kv.w;
-        }
-        dst[t] = y;
-    }
-
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
-    }
-}
-
-void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const float * k_d  = (const float *)dst->src[0]->data;
-    const float * v_d  = (const float *)dst->src[1]->data;
-    const float * r_d  = (const float *)dst->src[2]->data;
-    const float * tf_d = (const float *)dst->src[3]->data;
-    const float * td_d = (const float *)dst->src[4]->data;
-    const float * s_d  = (const float *)dst->src[5]->data;
-
-    const int64_t B = dst->src[5]->ne[1];
-    const int64_t T = dst->src[0]->ne[2];
-    const int64_t C = dst->ne[0];
-    const int64_t H = dst->src[0]->ne[1];
-
-    float * dst_d = (float *)dst->data;
-
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
-    GGML_ASSERT(C % H == 0);
-    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE); // The current cuda kernel is designed for RWKV6, HEAD_SIZE == 64
-
-    rwkv_wkv_f32<<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
-}

ggml/src/ggml-impl.h

@@ -381,6 +381,35 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size);
         return r;
     }
 
+#elif defined(__riscv) && defined(GGML_RV_ZFH)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        __asm__(
+            "fmv.h.x %[f], %[h]\n\t"
+            "fcvt.s.h %[f], %[f]"
+            : [f] "=&f" (f)
+            : [h] "r" (h)
+        );
+        return f;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __asm__(
+            "fcvt.h.s %[f], %[f]\n\t"
+            "fmv.x.h %[h], %[f]"
+            : [h] "=&r" (res)
+            : [f] "f" (f)
+        );
+        return res;
+    }
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #else
 
     // FP16 <-> FP32

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -1,6 +1,70 @@
 #ifndef GGML_METAL_IMPL
 #define GGML_METAL_IMPL
 
+// kernel parameters for mat-vec threadgroups
+//
+// N_R0: number of src0 rows to process per simdgroup
+// N_SG: number of simdgroups per threadgroup
+//
+// TODO: for optimal performance, become function of the device and work size
+
+#define N_R0_Q4_0 4
+#define N_SG_Q4_0 2
+
+#define N_R0_Q4_1 4
+#define N_SG_Q4_1 2
+
+#define N_R0_Q5_0 4
+#define N_SG_Q5_0 2
+
+#define N_R0_Q5_1 4
+#define N_SG_Q5_1 2
+
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+
+#define N_R0_Q2_K 4
+#define N_SG_Q2_K 2
+
+#define N_R0_Q3_K 2
+#define N_SG_Q3_K 2
+
+#define N_R0_Q4_K 4
+#define N_SG_Q4_K 2
+
+#define N_R0_Q5_K 2
+#define N_SG_Q5_K 2
+
+#define N_R0_Q6_K 1
+#define N_SG_Q6_K 2
+
+#define N_R0_IQ1_S 4
+#define N_SG_IQ1_S 2
+
+#define N_R0_IQ1_M 4
+#define N_SG_IQ1_M 2
+
+#define N_R0_IQ2_XXS 4
+#define N_SG_IQ2_XXS 2
+
+#define N_R0_IQ2_XS 4
+#define N_SG_IQ2_XS 2
+
+#define N_R0_IQ2_S 4
+#define N_SG_IQ2_S 2
+
+#define N_R0_IQ3_XXS 4
+#define N_SG_IQ3_XXS 2
+
+#define N_R0_IQ3_S 4
+#define N_SG_IQ3_S 2
+
+#define N_R0_IQ4_NL 2
+#define N_SG_IQ4_NL 2
+
+#define N_R0_IQ4_XS 2
+#define N_SG_IQ4_XS 2
+
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -155,9 +219,12 @@ typedef struct {
     int32_t  ne11;
     int32_t  ne_12_2; // assume K and V are same shape
     int32_t  ne_12_3;
-    uint64_t nb_12_1;
-    uint64_t nb_12_2;
-    uint64_t nb_12_3;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
     uint64_t nb31;
     int32_t  ne1;
     int32_t  ne2;
@@ -285,6 +352,13 @@ typedef struct {
     float    eps;
 } ggml_metal_kargs_rms_norm;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_l2_norm;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;

ggml/src/ggml-musa/CMakeLists.txt

@@ -67,10 +67,6 @@ if (MUSAToolkit_FOUND)
     add_compile_definitions(GGML_USE_MUSA)
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
-    if (GGML_CUDA_GRAPHS)
-        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()

ggml/src/ggml-opencl/CMakeLists.txt

@@ -25,124 +25,47 @@ endif ()
 if (GGML_OPENCL_EMBED_KERNELS)
     add_compile_definitions(GGML_OPENCL_EMBED_KERNELS)
 
-    set(OPENCL_CL_SOURCE_EMBED         "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl.cl.h")
-    set(OPENCL_MM_CL_SOURCE_EMBED      "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mm.cl.h")
-    set(OPENCL_CVT_CL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_cvt.cl.h")
+    set(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
+    file(MAKE_DIRECTORY     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    set(OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED             "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle.cl.h")
-    set(OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle_general.cl.h")
-    set(OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED          "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mul_mat_Ab_Bi_8x4.cl.h")
-    set(OPENCL_TRANSPOSE_16_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_16.cl.h")
-    set(OPENCL_TRANSPOSE_32_SOURCE_EMBED               "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32.cl.h")
-    set(OPENCL_TRANSPOSE_32_16_SOURCE_EMBED            "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32_16.cl.h")
-
-    set(EMBED_KERNEL_SCRIPT             "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
-    file(MAKE_DIRECTORY                 "${CMAKE_BINARY_DIR}/autogenerated")
-
-    include_directories("${CMAKE_BINARY_DIR}/autogenerated")
-
-    # Python must be accessible from command line
-    add_custom_command(
-        OUTPUT ${OPENCL_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl.cl
-            ${OPENCL_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_MM_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mm.cl
-            ${OPENCL_MM_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_mm.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_mm.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_CVT_CL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_cvt.cl
-            ${OPENCL_CVT_CL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_cvt.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_cvt.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle.cl
-            ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_gemv_noshuffle.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_gemv_noshuffle.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle_general.cl
-            ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_gemv_noshuffle_general.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_gemv_noshuffle_general.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl
-            ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_mul_mat_Ab_Bi_8x4.cl.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_16.cl
-            ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_16.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_16.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32.cl
-            ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_32.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_32.cl.h"
-    )
-
-    add_custom_command(
-        OUTPUT ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
-        COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32_16.cl
-            ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}
-        DEPENDS kernels/ggml-opencl_transpose_32_16.cl ${EMBED_KERNEL_SCRIPT}
-        COMMENT "Generate ggml-opencl_transpose_32_16.cl.h"
-    )
-
-    target_sources(${TARGET_NAME} PRIVATE
-                   ${OPENCL_CL_SOURCE_EMBED}
-                   ${OPENCL_MM_CL_SOURCE_EMBED}
-                   ${OPENCL_CVT_CL_SOURCE_EMBED}
-                   ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED}
-                   ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED}
-                   ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_16_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_32_SOURCE_EMBED}
-                   ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED})
-else ()
-    # copy ggml-opencl.cl to bin directory
-    configure_file(kernels/ggml-opencl.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_mm.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mm.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_cvt.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_cvt.cl COPYONLY)
-
-    configure_file(kernels/ggml-opencl_gemv_noshuffle.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_gemv_noshuffle_general.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle_general.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mul_mat_Ab_Bi_8x4.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_16.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_32.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32.cl COPYONLY)
-    configure_file(kernels/ggml-opencl_transpose_32_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32_16.cl COPYONLY)
+    target_include_directories(${TARGET_NAME} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 endif ()
+
+function(ggml_opencl_add_kernel KNAME)
+    set(KERN_HDR ${CMAKE_CURRENT_BINARY_DIR}/autogenerated/${KNAME}.cl.h)
+    set(KERN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernels/${KNAME}.cl)
+
+    if (GGML_OPENCL_EMBED_KERNELS)
+        message(STATUS "opencl: embedding kernel ${KNAME}")
+
+        # Python must be accessible from command line
+        add_custom_command(
+            OUTPUT ${KERN_HDR}
+            COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} ${KERN_SRC} ${KERN_HDR}
+            DEPENDS ${KERN_SRC} ${EMBED_KERNEL_SCRIPT}
+            COMMENT "Generate ${KERN_HDR}"
+        )
+
+        target_sources(${TARGET_NAME} PRIVATE ${KERN_HDR})
+    else ()
+        message(STATUS "opencl: adding kernel ${KNAME}")
+        configure_file(${KERN_SRC} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${KNAME}.cl COPYONLY)
+    endif ()
+endfunction()
+
+set(GGML_OPENCL_KERNELS
+    ggml-opencl
+    ggml-opencl_mm
+    ggml-opencl_cvt
+    ggml-opencl_gemv_noshuffle
+    ggml-opencl_gemv_noshuffle_general
+    ggml-opencl_mul_mat_Ab_Bi_8x4
+    ggml-opencl_transpose_16
+    ggml-opencl_transpose_32
+    ggml-opencl_transpose_32_16
+    ggml-opencl_im2col
+)
+
+foreach (K ${GGML_OPENCL_KERNELS})
+    ggml_opencl_add_kernel(${K})
+endforeach()

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -224,12 +224,14 @@ struct ggml_backend_opencl_context {
     cl_program program;
     cl_program program_1;
     cl_program program_2;
+    cl_program program_im2col;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
     cl_kernel kernel_scale;
     cl_kernel kernel_silu, kernel_silu_4;
     cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_clamp;
     cl_kernel kernel_norm;
@@ -239,6 +241,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
+    cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
     cl_kernel kernel_mul_mat_f32_f32;
     cl_kernel kernel_mul_mat_f16_f16;
@@ -252,6 +255,7 @@ struct ggml_backend_opencl_context {
               kernel_mul_mat_q4_0_f32_flat_img_v0;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -297,8 +301,27 @@ static int ggml_backend_opencl_n_devices = 0;
 struct ProfilingInfo {
     std::string op_name;
     std::string kernel_name;
-    // Kernel execution time in nanoseconds.
-    cl_ulong duration_ns;
+
+    cl_kernel kernel;
+    cl_event evt;
+
+    cl_ulong cmd_queued;
+    cl_ulong cmd_submit;
+    cl_ulong cmd_start;
+    cl_ulong cmd_end;
+    cl_ulong overhead_start;
+    cl_ulong overhead_end;
+    // For the times below, see spec for clGetEventProfilingInfo
+    // The time kernel spent in cmd queue - SUBMIT - QUEUED
+    cl_ulong cmd_queued_duration_ns;
+    // The time kernel spent for submission - START - SUBMIT
+    cl_ulong cmd_submit_duration_ns;
+    // Kernel execution time in nanoseconds - END - START
+    cl_ulong cmd_duration_ns;
+    // The time for the kernel to complete - COMPLETE - END
+    cl_ulong cmd_complete_duration_ns;
+    // Total time to finish the kernel - COMPELTE - QUEUED
+    cl_ulong cmd_total_duration_ns;
     // Global and local work sizes.
     size_t global_size[3];
     size_t local_size[3];
@@ -689,6 +712,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->kernel_silu_4             = clCreateKernel(backend_ctx->program, "kernel_silu_4", &err), err));
     CL_CHECK((backend_ctx->kernel_gelu               = clCreateKernel(backend_ctx->program, "kernel_gelu", &err), err));
     CL_CHECK((backend_ctx->kernel_gelu_4             = clCreateKernel(backend_ctx->program, "kernel_gelu_4", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick         = clCreateKernel(backend_ctx->program, "kernel_gelu_quick", &err), err));
+    CL_CHECK((backend_ctx->kernel_gelu_quick_4       = clCreateKernel(backend_ctx->program, "kernel_gelu_quick_4", &err), err));
     CL_CHECK((backend_ctx->kernel_relu               = clCreateKernel(backend_ctx->program, "kernel_relu", &err), err));
     CL_CHECK((backend_ctx->kernel_clamp              = clCreateKernel(backend_ctx->program, "kernel_clamp", &err), err));
     CL_CHECK((backend_ctx->kernel_norm               = clCreateKernel(backend_ctx->program, "kernel_norm", &err), err));
@@ -703,6 +728,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->kernel_rope_norm_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_neox_f32      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f32", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_neox_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f32     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_multi_f16     = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f32    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_rope_vision_f16    = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f16_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f16_f32        = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f32", &err), err));
     CL_CHECK((backend_ctx->kernel_cpy_f32_f16        = clCreateKernel(backend_ctx->program, "kernel_cpy_f32_f16", &err), err));
@@ -750,6 +779,19 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
 
     CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle     = clCreateKernel(backend_ctx->program_2, "kernel_convert_block_q4_0_noshuffle", &err), err));
 
+    // im2col kernels
+#ifdef GGML_OPENCL_EMBED_KERNELS
+    const std::string kernel_src_im2col {
+        #include "ggml-opencl_im2col.cl.h"
+    };
+#else
+    const std::string kernel_src_im2col = read_file("ggml-opencl_im2col.cl");
+#endif
+    backend_ctx->program_im2col = build_program_from_source(context, device, kernel_src_im2col.c_str(), compile_opts);
+
+    CL_CHECK((backend_ctx->kernel_im2col_f32 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f32", &err), err));
+    CL_CHECK((backend_ctx->kernel_im2col_f16 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f16", &err), err));
+
     // Kernels for Adreno
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -903,12 +945,56 @@ static void ggml_cl2_free(void) {
         return;
     }
 
+    // Populate profiling info
+    for (ProfilingInfo & info : g_profiling_info) {
+        cl_ulong cmd_queued;
+        cl_ulong cmd_submit;
+        cl_ulong cmd_start;
+        cl_ulong cmd_end;
+        cl_ulong cmd_complete;
+
+        CL_CHECK(clWaitForEvents(1, &info.evt));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL));
+        CL_CHECK(clGetEventProfilingInfo(
+            info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL));
+        CL_CHECK(clReleaseEvent(info.evt));
+
+        char kernel_name[512];
+        CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME,
+            sizeof(kernel_name), kernel_name, NULL));
+        info.kernel_name = kernel_name;
+
+        info.cmd_queued = cmd_queued;
+        info.cmd_submit = cmd_submit;
+        info.cmd_start  = cmd_start;
+        info.cmd_end    = cmd_end;
+
+        info.cmd_queued_duration_ns     = cmd_submit    - cmd_queued;
+        info.cmd_submit_duration_ns     = cmd_start     - cmd_submit;
+        info.cmd_duration_ns            = cmd_end       - cmd_start;
+        info.cmd_complete_duration_ns   = cmd_complete  - cmd_end;
+        info.cmd_total_duration_ns      = cmd_complete  - cmd_queued;
+    }
+
+    // Dump a csv
     float total_kernel_time = 0;
-    fprintf(fperf, "op name, kernel name, duration (ms), global size, local size, output size\n");
+    fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n");
     for (const ProfilingInfo & info : g_profiling_info) {
-        total_kernel_time += info.duration_ns/1.e6f;
-        fprintf(fperf, "%s,%s,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
-            info.op_name.c_str(), info.kernel_name.c_str(), info.duration_ns/1.e6f,
+        total_kernel_time += info.cmd_duration_ns/1.e6f;
+        fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
+            info.op_name.c_str(), info.kernel_name.c_str(),
+            info.cmd_queued_duration_ns/1.e6f,
+            info.cmd_submit_duration_ns/1.e6f,
+            info.cmd_duration_ns/1.e6f,
+            info.cmd_complete_duration_ns/1.e6f,
+            info.cmd_total_duration_ns/1.e6f,
             info.global_size[0], info.global_size[1], info.global_size[2],
             info.local_size[0], info.local_size[2], info.local_size[2],
             info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
@@ -916,6 +1002,27 @@ static void ggml_cl2_free(void) {
     fclose(fperf);
 
     GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time);
+
+    // Dump a simple chrome trace
+    FILE* ftrace = fopen("cl_trace.json", "w");
+    if (!ftrace) {
+        GGML_LOG_ERROR("Failed to open cl_trace.json\n");
+        return;
+    }
+
+    fprintf(ftrace, "[\n");
+    for (const ProfilingInfo & info : g_profiling_info) {
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            info.kernel_name.c_str(), info.cmd_queued/1000);
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n",
+            info.kernel_name.c_str(), info.cmd_submit/1000);
+
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            info.kernel_name.c_str(), info.cmd_start/1000);
+        fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n",
+            info.kernel_name.c_str(), info.cmd_end/1000);
+    }
+    fclose(ftrace);
 #endif
 }
 
@@ -1103,6 +1210,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_QUICK:
                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -1132,14 +1240,26 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->ne[3] == 1;
         case GGML_OP_ROPE: {
             const int mode = ((const int32_t *) op->op_params)[2];
-            if (mode & GGML_ROPE_TYPE_MROPE) {
+            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+            if (is_mrope && !is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                 return false;
             }
-            if (mode & GGML_ROPE_TYPE_VISION) {
+            if (is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
                 return false;
             }
             return true;
         }
+        case GGML_OP_IM2COL:
+            return true;
         default:
             return false;
     }
@@ -2062,25 +2182,14 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
 // Profiling utility
 //------------------------------------------------------------------------------
 #ifdef GGML_OPENCL_PROFILING
-void populateProfilingInfo(
+static void populateProfilingInfo(
         ProfilingInfo& info, cl_event evt, cl_kernel kernel,
         size_t global_size[3], size_t local_size[3],
         const ggml_tensor * tensor) {
-    cl_ulong start;
-    cl_ulong end;
-    CL_CHECK(clWaitForEvents(1, &evt));
-    CL_CHECK(clGetEventProfilingInfo(
-        evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL));
-    CL_CHECK(clGetEventProfilingInfo(
-        evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL));
+    info.op_name     = tensor->name;
+    info.kernel      = kernel;
+    info.evt         = evt;
 
-    char kernel_name[512];
-    CL_CHECK(clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME,
-        sizeof(kernel_name), kernel_name, NULL));
-
-    info.duration_ns = end - start;
-    info.op_name = tensor->name;
-    info.kernel_name = kernel_name;
     info.local_size[0]  = local_size[0];
     info.local_size[1]  = local_size[1];
     info.local_size[2]  = local_size[2];
@@ -2509,6 +2618,53 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }
 
+static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_quick_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_quick;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt);
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
+#endif
+}
+
 static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -3907,6 +4063,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     float attn_factor;
     float beta_fast;
     float beta_slow;
+    int32_t sections[4];
 
     memcpy(&freq_base,   (int32_t *) dst->op_params + 5, sizeof(float));
     memcpy(&freq_scale,  (int32_t *) dst->op_params + 6, sizeof(float));
@@ -3914,23 +4071,23 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int32_t)*4);
 
     const bool is_neox = mode & 2;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
 
     cl_kernel kernel;
 
-    if (!is_neox) {
-        switch (src0->type) {
-            case GGML_TYPE_F32:
-                kernel = backend_ctx->kernel_rope_norm_f32;
-                break;
-            case GGML_TYPE_F16:
-                kernel = backend_ctx->kernel_rope_norm_f16;
-                break;
-            default:
-                GGML_ASSERT(false);
-        };
-    } else {
+    if (is_neox) {
         switch (src0->type) {
             case GGML_TYPE_F32:
                 kernel = backend_ctx->kernel_rope_neox_f32;
@@ -3941,6 +4098,39 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
             default:
                 GGML_ASSERT(false);
         };
+    } else if (is_mrope && !is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_multi_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_multi_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    } else if (is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_vision_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_vision_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    } else {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_norm_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_norm_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
     }
 
     CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
@@ -3976,6 +4166,9 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
     CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
     CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    if (is_mrope || is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
+    }
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -3991,6 +4184,98 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
 #endif
 }
 
+static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // src0 - filter, src1 - input
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const cl_long IC = src1->ne[is_2D ? 2 : 1];
+    const cl_long IH = is_2D ? src1->ne[1] : 1;
+    const cl_long IW =         src1->ne[0];
+
+    const cl_long KH = is_2D ? src0->ne[1] : 1;
+    const cl_long KW =         src0->ne[0];
+
+    const cl_long OH = is_2D ? dst->ne[2] : 1;
+    const cl_long OW =         dst->ne[1];
+
+    // nb is byte offset, src is type float32
+    const cl_ulong delta_offset = src1->nb[is_2D ? 2 : 1]/4;
+    const cl_long  batch        = src1->ne[is_2D ? 3 : 2];
+    const cl_ulong batch_offset = src1->nb[is_2D ? 3 : 2]/4;
+
+    const cl_long pelements = OW*KW*KH;
+    const cl_long CHW       = IC*KH*KW;
+
+    cl_kernel kernel;
+
+    if(dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_im2col_f16;
+    } else {
+        kernel = backend_ctx->kernel_im2col_f32;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(cl_ulong), &batch_offset));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(cl_ulong), &delta_offset));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(cl_long),  &IW));
+    CL_CHECK(clSetKernelArg(kernel,   7, sizeof(cl_long),  &IH));
+    CL_CHECK(clSetKernelArg(kernel,   8, sizeof(cl_long),  &IC));
+    CL_CHECK(clSetKernelArg(kernel,   9, sizeof(cl_long),  &OW));
+    CL_CHECK(clSetKernelArg(kernel,  10, sizeof(cl_long),  &OH));
+    CL_CHECK(clSetKernelArg(kernel,  11, sizeof(cl_long),  &KW));
+    CL_CHECK(clSetKernelArg(kernel,  12, sizeof(cl_long),  &KH));
+    CL_CHECK(clSetKernelArg(kernel,  13, sizeof(cl_long),  &pelements));
+    CL_CHECK(clSetKernelArg(kernel,  14, sizeof(cl_long),  &CHW));
+    CL_CHECK(clSetKernelArg(kernel,  15, sizeof(int),      &s0));
+    CL_CHECK(clSetKernelArg(kernel,  16, sizeof(int),      &s1));
+    CL_CHECK(clSetKernelArg(kernel,  17, sizeof(int),      &p0));
+    CL_CHECK(clSetKernelArg(kernel,  18, sizeof(int),      &p1));
+    CL_CHECK(clSetKernelArg(kernel,  19, sizeof(int),      &d0));
+    CL_CHECK(clSetKernelArg(kernel,  20, sizeof(int),      &d1));
+
+    const int num_blocks = (pelements + 256 - 1) / 256;
+    size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC};
+    size_t local_work_size[] = {256, 1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+}
+
 //------------------------------------------------------------------------------
 // Op offloading
 //------------------------------------------------------------------------------
@@ -4049,6 +4334,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_gelu;
                     break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_quick;
+                    break;
                 case GGML_UNARY_OP_SILU:
                     if (!any_on_device) {
                         return false;
@@ -4121,6 +4412,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_rope;
             break;
+        case GGML_OP_IM2COL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_im2col;
+            break;
         default:
             return false;
     }