vulkan: fix GPU deduplication logic. (#19222)

* vulkan: fix GPU deduplication logic.

As reported in https://github.com/ggml-org/llama.cpp/issues/19221, the
(same uuid, same driver) logic is problematic for windows+intel igpu.

Let's just avoid filtering for MoltenVK which is apple-specific, and
keep the logic the  same as before 88d23ad5 - just dedup based on UUID.

Verified that MacOS + 4xVega still reports 4 GPUs with this version.

* vulkan: only skip dedup when both drivers are moltenVk

.github/workflows/build.yml

@@ -293,6 +293,7 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
@@ -303,6 +304,7 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
             -DGGML_OPENMP=OFF
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
@@ -1532,7 +1534,7 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
-          LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
   ggml-ci-arm64-cpu-high-perf:
     runs-on: ubuntu-22.04-arm
@@ -1558,7 +1560,7 @@ jobs:
       - name: Test
         id: ggml-ci
         run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
   ggml-ci-arm64-cpu-high-perf-sve:
     runs-on: ubuntu-22.04-arm

.github/workflows/server.yml

@@ -36,7 +36,7 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
@@ -45,7 +45,7 @@ jobs:
           - build_type: Release
             sanitizer: ""
             extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+      fail-fast: false
 
     steps:
       - name: Dependencies
@@ -72,7 +72,15 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
+          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
+            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
           cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
@@ -88,7 +96,7 @@ jobs:
 
       - name: Tests
         id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) && matrix.build_type == 'Release' }}
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
         run: |
           cd tools/server/tests
           export ${{ matrix.extra_args }}

CMakeLists.txt

@@ -164,29 +164,6 @@ llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
-if (NOT MSVC)
-    if (LLAMA_SANITIZE_THREAD)
-        message(STATUS "Using -fsanitize=thread")
-
-        add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
-    endif()
-
-    if (LLAMA_SANITIZE_ADDRESS)
-        message(STATUS "Using -fsanitize=address")
-
-        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
-    endif()
-
-    if (LLAMA_SANITIZE_UNDEFINED)
-        message(STATUS "Using -fsanitize=undefined")
-
-        add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
-    endif()
-endif()
-
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")
 

CODEOWNERS

@@ -27,6 +27,7 @@
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/debug/                        @danbev @pwilkin
 /examples/deprecation-warning/          @ggerganov
 /examples/diffusion/                    @am17an
 /examples/embedding/                    @ggerganov

ci/run.sh

@@ -635,6 +635,29 @@ function gg_check_build_requirements {
     fi
 }
 
+function gg_run_test_backend_ops_cpu {
+    cd ${SRC}
+
+    cd build-ci-release
+
+    set -e
+
+    (time ./bin/test-backend-ops -b CPU ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops-cpu.log
+
+    set +e
+}
+
+function gg_sum_test_backend_ops_cpu {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test-backend-ops for CPU backend\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops-cpu.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
 ## main
 
 export LLAMA_LOG_PREFIX=1
@@ -663,6 +686,10 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
+if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
+    test $ret -eq 0 && gg_run test_backend_ops_cpu
+fi
+
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
     test $ret -eq 0 && gg_run rerank_tiny

cmake/common.cmake

@@ -32,4 +32,27 @@ function(llama_add_compile_flags)
             set(CXX_FLAGS "" PARENT_SCOPE)
         endif()
     endif()
+
+    if (NOT MSVC)
+        if (LLAMA_SANITIZE_THREAD)
+            message(STATUS "Using -fsanitize=thread")
+
+            add_compile_options(-fsanitize=thread)
+            link_libraries     (-fsanitize=thread)
+        endif()
+
+        if (LLAMA_SANITIZE_ADDRESS)
+            message(STATUS "Using -fsanitize=address")
+
+            add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+            link_libraries     (-fsanitize=address)
+        endif()
+
+        if (LLAMA_SANITIZE_UNDEFINED)
+            message(STATUS "Using -fsanitize=undefined")
+
+            add_compile_options(-fsanitize=undefined)
+            link_libraries     (-fsanitize=undefined)
+        endif()
+    endif()
 endfunction()

common/debug.cpp

@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
     return v;
 }
 
+#define INDENT "    "
+
 template <bool abort>
 void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
     GGML_ASSERT(n > 0);
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         }
     }
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG_ERR("                                     [\n");
+        LOG(INDENT "[\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2 * n) {
-                LOG_ERR("                                      ..., \n");
+                LOG(INDENT INDENT "..., \n");
                 i2 = ne[2] - n;
             }
-            LOG_ERR("                                      [\n");
+            LOG(INDENT INDENT "[\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2 * n) {
-                    LOG_ERR("                                       ..., \n");
+                    LOG(INDENT INDENT INDENT "..., \n");
                     i1 = ne[1] - n;
                 }
-                LOG_ERR("                                       [");
+                LOG(INDENT INDENT INDENT "[");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2 * n) {
-                        LOG_ERR("..., ");
+                        LOG("   ..., ");
                         i0 = ne[0] - n;
                     }
                     const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
-                    LOG_ERR("%12.4f", v);
+                    LOG("%12.4f", v);
                     if (i0 < ne[0] - 1) {
-                        LOG_ERR(", ");
+                        LOG(", ");
                     }
                 }
-                LOG_ERR("],\n");
+                LOG("  ],\n");
             }
-            LOG_ERR("                                      ],\n");
+            LOG(INDENT INDENT "],\n");
         }
-        LOG_ERR("                                     ]\n");
-        LOG_ERR("                                     sum = %f\n", sum);
+        LOG(INDENT "]\n");
+        LOG(INDENT "sum = %f\n", sum);
     }
 
     if constexpr (abort) {
         if (std::isnan(sum)) {
-            LOG_ERR("encountered NaN - aborting\n");
+            LOG("encountered NaN - aborting\n");
             exit(0);
         }
     }
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
     }
 
     if (matches_filter) {
-        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
-                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
-                common_ggml_ne_string(t).c_str());
+        LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+            ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+            common_ggml_ne_string(t).c_str());
     }
 
     const bool is_host = ggml_backend_buffer_is_host(t->buffer);

common/jinja/runtime.cpp

@@ -144,6 +144,13 @@ value binary_expression::execute_impl(context & ctx) {
         return false;
     };
 
+    auto test_is_in = [&]() -> bool {
+        func_args args(ctx);
+        args.push_back(left_val);
+        args.push_back(right_val);
+        return global_builtins().at("test_is_in")(args)->as_bool();
+    };
+
     // Handle undefined and null values
     if (is_val<value_undefined>(left_val) || is_val<value_undefined>(right_val)) {
         if (is_val<value_undefined>(right_val) && (op.value == "in" || op.value == "not in")) {
@@ -223,19 +230,11 @@ value binary_expression::execute_impl(context & ctx) {
             return result;
         }
     } else if (is_val<value_array>(right_val)) {
-        auto & arr = right_val->as_array();
-        bool member = false;
-        for (const auto & item : arr) {
-            if (*left_val == *item) {
-                member = true;
-                break;
-            }
-        }
+        // case: 1 in [0, 1, 2]
+        bool member = test_is_in();
         if (op.value == "in") {
-            JJ_DEBUG("Checking membership: %s in Array is %d", left_val->type().c_str(), member);
             return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            JJ_DEBUG("Checking non-membership: %s not in Array is %d", left_val->type().c_str(), !member);
             return mk_val<value_bool>(!member);
         }
     }
@@ -252,22 +251,23 @@ value binary_expression::execute_impl(context & ctx) {
 
     // String membership
     if (is_val<value_string>(left_val) && is_val<value_string>(right_val)) {
-        auto left_str = left_val->as_string().str();
-        auto right_str = right_val->as_string().str();
+        // case: "a" in "abc"
+        bool member = test_is_in();
         if (op.value == "in") {
-            return mk_val<value_bool>(right_str.find(left_str) != std::string::npos);
+            return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            return mk_val<value_bool>(right_str.find(left_str) == std::string::npos);
+            return mk_val<value_bool>(!member);
         }
     }
 
     // Value key in object
     if (is_val<value_object>(right_val)) {
-        bool has_key = right_val->has_key(left_val);
+        // case: key in {key: value}
+        bool member = test_is_in();
         if (op.value == "in") {
-            return mk_val<value_bool>(has_key);
+            return mk_val<value_bool>(member);
         } else if (op.value == "not in") {
-            return mk_val<value_bool>(!has_key);
+            return mk_val<value_bool>(!member);
         }
     }
 

common/jinja/value.cpp

@@ -393,6 +393,33 @@ const func_builtins & global_builtins() {
         {"test_is_lt", test_compare_fn<value_compare_op::lt>},
         {"test_is_lessthan", test_compare_fn<value_compare_op::lt>},
         {"test_is_ne", test_compare_fn<value_compare_op::ne>},
+        {"test_is_in", [](const func_args & args) -> value {
+            args.ensure_count(2);
+            auto needle   = args.get_pos(0);
+            auto haystack = args.get_pos(1);
+            if (is_val<value_undefined>(haystack)) {
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_array>(haystack)) {
+                for (const auto & item : haystack->as_array()) {
+                    if (*needle == *item) {
+                        return mk_val<value_bool>(true);
+                    }
+                }
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_string>(haystack)) {
+                if (!is_val<value_string>(needle)) {
+                    throw raised_exception("'in' test expects args[1] as string when args[0] is string, got args[1] as " + needle->type());
+                }
+                return mk_val<value_bool>(
+                    haystack->as_string().str().find(needle->as_string().str()) != std::string::npos);
+            }
+            if (is_val<value_object>(haystack)) {
+                return mk_val<value_bool>(haystack->has_key(needle));
+            }
+            throw raised_exception("'in' test expects iterable as first argument, got " + haystack->type());
+        }},
         {"test_is_test", [](const func_args & args) -> value {
             args.ensure_vals<value_string>();
             auto & builtins = global_builtins();

common/ngram-map.cpp

@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
  * @return Vector of draft tokens, empty if no matching pattern is found
  */
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled) {
 
     // Simple implementation of self-speculative decoding without a draft model.
     //
     const size_t cur_len = tokens.size();
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (state.idx_last_check + state.config.check_rate > cur_len) {
-        llama_tokens draft_tokens;
-        return draft_tokens;
-    }
 
-    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
-    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+    const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
+    const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
 
     // vector for tokens we want to verify.
     // return empty vector if there is no match.
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
     }
     pattern.push_back(sampled); // add the last token to the pattern
 
-    // We do a search in the token history.
-    state.idx_last_check = cur_len;
-
     size_t match_pos = 0; // we ignore position 0, position 0 == no match
                           // search backwards, but skip the current match (we are currently there)
     for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {

common/ngram-map.h

@@ -27,23 +27,9 @@ struct common_ngram_simple_config {
     uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
 };
 
-// current state (and config) of n-gram simple.
-struct common_ngram_simple_state {
-    common_ngram_simple_config config;
-
-    size_t idx_last_check = 0; // index of last check in context history (mutable)
-
-    common_ngram_simple_state(const common_ngram_simple_config & config)
-        : config(config) {}
-};
-
 // Searches for a n-gram in the history and checks whether a draft sequence should be generated.
-// state:              the ngram simple state to search in.
-// inp:                the tokens generated so far.
-// sampled:            the token that was just sampled.
-// draft:              vector to store the draft tokens, initially empty.
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled);
 
 

common/speculative.cpp

@@ -463,12 +463,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
 
 // state of self-speculation (simple implementation, not ngram-map)
 struct common_speculative_state_ngram_simple : public common_speculative_state {
-    common_ngram_simple_state state;
+    common_ngram_simple_config config;
+
+    uint16_t check_id = 0; // used to control the frequency of generating drafts
 
     common_speculative_state_ngram_simple(
             enum common_speculative_type type,
-            common_ngram_simple_state state)
-        : common_speculative_state(type), state(state) {}
+            common_ngram_simple_config config)
+        : common_speculative_state(type), config(config) {}
 
     void begin(const llama_tokens & prompt) override {
         GGML_UNUSED(prompt);
@@ -479,7 +481,13 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+        ++check_id;
+        if (check_id < config.check_rate) {
+            return;
+        }
+        check_id = 0;
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
         GGML_UNUSED(params);
     }
 
@@ -889,14 +897,14 @@ common_speculative * common_speculative_init(
                 uint16_t mgram_size_value = ngram_map.size_value;
                 uint16_t check_rate       = ngram_map.check_rate;
 
-                auto config_simple = common_ngram_simple_config{
+                auto config_simple = common_ngram_simple_config {
                     /* .size_ngram      = */ ngram_size_key,
                     /* .size_mgram      = */ mgram_size_value,
                     /* .check_rate      = */ check_rate
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
                     /* .type            = */ config.type,
-                    /* .state           = */ common_ngram_simple_state(config_simple)
+                    /* .state           = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;
@@ -951,12 +959,8 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
     }
 
     for (auto & impl : spec->impls) {
-        const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+        common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
         impl->begin(prompt);
-
-        const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-        impl->t_begin_us += t_now_us - t_start_us; // accumulate duration for this refresh
     }
 }
 
@@ -971,14 +975,9 @@ llama_tokens common_speculative_draft(
 
     for (auto & impl : spec->impls) {
         {
-            const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+            common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
             impl->draft(params, prompt_tgt, id_last, result);
-
-            const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-
             impl->drafts_call_count++;
-            impl->t_draft_us += t_now_us - t_start_us; // accumulate duration for this implementation
         }
 
         if (!result.empty()) {
@@ -1006,15 +1005,15 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
 
     GGML_ASSERT(impl);
 
-    const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-    if (n_accepted > 0) {
-        impl->drafts_accepted_count++;
-        impl->drafts_accepted_tokens += n_accepted;
-    }
+    {
+        common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
+        if (n_accepted > 0) {
+            impl->drafts_accepted_count++;
+            impl->drafts_accepted_tokens += n_accepted;
+        }
 
-    impl->accept(n_accepted);
-    const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-    impl->t_accept_us += t_now_us - t_start_us; // accumulate duration for this acculumulation
+        impl->accept(n_accepted);
+    }
 }
 
 void common_speculative_print_stats(const common_speculative * spec) {

docs/backend/SYCL.md

@@ -22,12 +22,11 @@
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
-- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
 ### Llama.cpp + SYCL
 
 The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
-SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
+SYCL cross-platform capabilities enable support for other vendor GPUs as well.
 
 ## Recommended Release
 
@@ -42,6 +41,9 @@ The following releases are verified and recommended:
 
 ## News
 
+- 2026.02
+  - Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. User can't build up the software for Nvidia & AMD GPU.
+
 - 2025.11
   - Support malloc memory on device more than 4GB.
 
@@ -111,8 +113,8 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc A-Series              | Support | Arc A770, Arc A730M, Arc A750         |
-| Intel Arc B-Series              | Support | Arc B580         |
+| Intel Arc A-Series            | Support | Arc A770, Arc A730M, Arc A750         |
+| Intel Arc B-Series            | Support | Arc B580                              |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
 | Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
 
@@ -127,20 +129,7 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 
 ### Other Vendor GPU
 
-**Verified devices**
-
-| Nvidia GPU               | Status    | Verified Model |
-|--------------------------|-----------|----------------|
-| Ampere Series            | Supported | A100, A4000    |
-| Ampere Series *(Mobile)* | Supported | RTX 40 Series  |
-
-| AMD GPU                  | Status       | Verified Model |
-|--------------------------|--------------|----------------|
-| Radeon Pro               | Experimental | W6800          |
-| Radeon RX                | Experimental | 6700 XT        |
-
-Note: AMD GPU support is highly experimental and is incompatible with F16.
-Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
+NA
 
 ## Docker
 
@@ -149,11 +138,11 @@ The docker build option is currently limited to *Intel GPU* targets.
 ### Build image
 
 ```sh
-# Using FP16
-docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
-
 # Using FP32
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=OFF" --target light -f .devops/intel.Dockerfile .
+
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
 ```
 
 *Notes*:
@@ -212,14 +201,6 @@ Platform #0: Intel(R) OpenCL HD Graphics
  `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```
 
-- **Nvidia GPU**
-
-In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
-
-- **AMD GPU**
-
-To target AMD GPUs with SYCL, the ROCm stack must be installed first.
-
 2. **Install Intel® oneAPI Base toolkit**
 
 SYCL backend depends on:
@@ -248,23 +229,6 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 |2025.1|
 |2024.1|
 
-- **Adding support to Nvidia GPUs**
-
-**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
-
-**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.
-
 3. **Verify installation and environment**
 
 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -285,25 +249,6 @@ When targeting an intel GPU, the user should expect one or more devices among th
 [opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) UHD Graphics 730 OpenCL 3.0 NEO  [24.39.31294]
 ```
 
-- **Nvidia GPU**
-
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
-
-```
-[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
-[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
-[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
-```
-
-- **AMD GPU**
-
-For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
-
-```
-[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
-[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
-```
-
 ### II. Build llama.cpp
 
 #### Intel GPU
@@ -332,47 +277,6 @@ It is possible to come across some precision issues when running tests that stem
 instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS`
 as `-cl-fp32-correctly-rounded-divide-sqrt`
 
-#### Nvidia GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with Nvidia BLAS acceleration through SYCL
-# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
-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=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
-It is possible to come across some precision issues when running tests that stem from using faster
-instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler.
-
-#### AMD GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with rocBLAS acceleration through SYCL
-
-## AMD
-# Use FP32, FP16 is not supported
-# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
-GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
 ### III. Run the inference
 
 #### Retrieve and prepare model
@@ -766,15 +670,15 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
-| 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_TARGET   | INTEL *(default)*                     | Set the SYCL target device type.            |
+| GGML_SYCL_DEVICE_ARCH | Optional                           | Set the SYCL device architecture. 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. (1.) |
-| 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). |
+| GGML_SYCL_GRAPH    | OFF *(default)* \|ON *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | 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. |
 
-1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.
 
 #### Runtime
 
@@ -782,7 +686,7 @@ 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 for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
-| 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. |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | 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 |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|

docs/build.md

@@ -252,9 +252,7 @@ CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.ggu
 
 The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.
 
-**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
-
-See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
+Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
 
 ### Unified Memory
 

examples/model-conversion/scripts/utils/tensor-info.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+from safetensors import safe_open
+
+
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
+
+
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX
+
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})
+
+    return None
+
+
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return list(weight_map.keys())
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(model_path: Path, unique: bool = False):
+    tensor_names = get_all_tensor_names(model_path)
+
+    if unique:
+        seen = set()
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen.add(normalized)
+                print(normalized)
+    else:
+        for tensor_name in sorted(tensor_names):
+            print(tensor_name)
+
+
+def print_tensor_info(model_path: Path, tensor_name: str):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",  # optional (if --list is used for example)
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list",
+        action="store_true",
+        help="List unique tensor patterns in the model (layer numbers replaced with #)"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list:
+        list_all_tensors(model_path, unique=True)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name)
+
+
+if __name__ == "__main__":
+    main()

ggml/include/ggml-cpu.h

@@ -19,6 +19,9 @@ extern "C" {
         // abort ggml_graph_compute when true
         ggml_abort_callback abort_callback;
         void *              abort_callback_data;
+
+        // use only reference implementations
+        bool use_ref;
     };
 
     // numa strategies
@@ -132,6 +135,8 @@ extern "C" {
     GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
     GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
+    GGML_BACKEND_API void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref);
+
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);

ggml/include/ggml-virtgpu.h

@@ -7,8 +7,6 @@
 extern "C" {
 #endif
 
-#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
 
 #ifdef  __cplusplus

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -268,9 +268,9 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
                            _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 
-static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
-    return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -782,6 +782,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accum1 = _mm256_setzero_ps();
     __m256 accum2 = _mm256_setzero_ps();
+
     for (; ib + 1 < nb; ib += 2) {
         const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
         const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
@@ -795,10 +796,10 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
         const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
-                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
-                _mm256_cvtepi32_ps(p_2), accum2);
+        const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
+        const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
+        accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
     }
 
     sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
@@ -830,7 +831,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
         int sumi1 = 0;
         int sumi2 = 0;
         for (int j = 0; j < QK_MXFP4/2; ++j) {
@@ -3817,4 +3818,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -24,6 +24,9 @@ struct ggml_compute_params {
     void * wdata;
 
     struct ggml_threadpool * threadpool;
+
+    // use reference implementation
+    bool use_ref;
 };
 
 

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

@@ -5,7 +5,6 @@
 #include "ggml-backend.h"
 #include "traits.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "quants.h"
 #include "ggml-threading.h"
@@ -76,6 +75,9 @@
 // precomputed f32 table for f16 (256 KB) (simd-mappings.h)
 float ggml_table_f32_f16[1 << 16];
 
+// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
+float ggml_table_f32_e8m0_half[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int sve_cnt;
@@ -2867,12 +2869,20 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
+                        const int64_t neq2 = node->src[0]->ne[2]; // number of query heads
                         const int64_t DK = node->src[1]->ne[0];
                         const int64_t DV = node->src[2]->ne[0];
 
                         // Tiled flash attention scratch (tile sizes defined in common.h)
                         // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
-                        cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+
+                        // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
+                        // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
+                        size_t n_chunks = n_tasks;
+                        size_t decode   = sizeof(float)*(neq2*n_chunks*(2+DV) + n_tasks*(DK + 2*DV));
+
+                        cur += MAX(prefill, decode);
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {
@@ -2929,11 +2939,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     set_numa_thread_affinity(state->ith);
 
     struct ggml_compute_params params = {
-        /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
-        /*.wsize     =*/ cplan->work_size,
-        /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ tp,
+        /*.ith        =*/ state->ith,
+        /*.nth        =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
+        /*.wsize      =*/ cplan->work_size,
+        /*.wdata      =*/ cplan->work_data,
+        /*.threadpool =*/ tp,
+        /*.use_ref    =*/ cplan->use_ref,
     };
 
     GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
@@ -3673,6 +3684,11 @@ void ggml_cpu_init(void) {
                 ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
+            // initialize E8M0 half table (256 entries)
+            for (int i = 0; i < (1 << 8); ++i) {
+                ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
+            }
+
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
 
             GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);

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

@@ -105,6 +105,8 @@ struct ggml_backend_cpu_context {
 
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
+
+    bool                use_ref;  // use reference implementation
 };
 
 static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
@@ -143,6 +145,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
 
     cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
     cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cpu_plan->cplan.use_ref             = cpu_ctx->use_ref;
 
     return cpu_plan;
 }
@@ -182,6 +185,7 @@ static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, s
 
     cplan.abort_callback      = cpu_ctx->abort_callback;
     cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cplan.use_ref             = cpu_ctx->use_ref;
 
     return ggml_graph_compute(cgraph, &cplan);
 }
@@ -223,6 +227,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     ctx->work_size           = 0;
     ctx->abort_callback      = NULL;
     ctx->abort_callback_data = NULL;
+    ctx->use_ref             = false;
 
     ggml_backend_t cpu_backend = new ggml_backend {
         /* .guid    = */ ggml_backend_cpu_guid(),
@@ -270,6 +275,13 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_
     ctx->abort_callback_data = abort_callback_data;
 }
 
+void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->use_ref = use_ref;
+}
+
 // CPU backend - device
 
 struct ggml_backend_cpu_device_context {
@@ -646,6 +658,9 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
     if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
         return (void *)ggml_is_numa;
     }
+    if (strcmp(name, "ggml_backend_cpu_set_use_ref") == 0) {
+        return (void *)ggml_backend_cpu_set_use_ref;
+    }
 
     // threadpool - TODO:  move to ggml-base
     if (strcmp(name, "ggml_threadpool_new") == 0) {

ggml/src/ggml-cpu/ops.cpp

@@ -8042,12 +8042,14 @@ void ggml_compute_forward_top_k(
     }
 }
 
-// ggml_compute_forward_flash_attn_ext
-
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
         ggml_tensor * dst,
-        int ir0, int ir1) {
+        int ir0, int ir1,
+        int64_t ic_start, int64_t ic_end,
+        float * partials, int64_t partial_stride) {
+
+    const bool write_partials = (partials != nullptr);
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -8124,7 +8126,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
     int ith = params->ith;
 
-    // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
         const int iq3 = ir/(neq2*neq1);
@@ -8165,7 +8166,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
 
-        for (int64_t ic = 0; ic < nek1; ++ic) {
+        for (int64_t ic = ic_start; ic < ic_end; ++ic) {
             const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
@@ -8238,8 +8239,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             }
         }
 
-        // sinks
-        if (sinks) {
+        // sinks - apply only on the first kv-chunk
+        if (sinks && ic_start == 0) {
             const float s = ((float *)((char *) sinks->data))[h];
 
             float ms = 1.0f;
@@ -8247,6 +8248,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
             if (s > M) {
                 ms = expf(M - s);
+                M = s;
                 ggml_vec_scale_f32(DV, VKQ32, ms);
             } else {
                 vs = expf(s - M);
@@ -8255,20 +8257,26 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             S = S*ms + vs;
         }
 
-        // V /= S
-        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
-        ggml_vec_scale_f32(DV, VKQ32, S_inv);
+        if (write_partials) {
+            // Write M, S, VKQ to partials for later reduction
+            // partials layout: [M, S, VKQ[DV]] per query head
+            float * partial = partials + ir * partial_stride;
+            partial[0] = M;
+            partial[1] = S;
+            memcpy(partial + 2, VKQ32, DV * sizeof(float));
+        } else {
+            // V /= S
+            const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
+            ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
-        // dst indices
-        const int i1 = iq1;
-        const int i2 = iq2;
-        const int i3 = iq3;
+            // dst indices
+            const int i1 = iq1;
+            const int i2 = iq2;
+            const int i3 = iq3;
 
-        // original
-        //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
-
-        // permute(0, 2, 1, 3)
-        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+        }
     }
 }
 
@@ -8546,6 +8554,78 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
     }
 }
 
+// Reduction function: combines partial results across KV chunks
+// Partials layout in wdata: [n_q_heads][n_chunks][2 + DV]
+static void ggml_flash_attn_ext_reduce_partials(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        const int64_t n_chunks,
+        const int64_t chunk_size) {
+
+    const ggml_tensor * q = dst->src[0];
+    const ggml_tensor * k = dst->src[1];
+    const ggml_tensor * v = dst->src[2];
+
+    const int64_t DK        = k->ne[0];
+    const int64_t DV        = v->ne[0];
+    const int64_t nek1      = k->ne[1];
+    const int64_t n_q_heads = q->ne[2];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t wdata_per_thread = DK + 2*DV + CACHE_LINE_SIZE_F32;
+    float *       thread_wdata     = (float *) params->wdata + ith * wdata_per_thread;
+
+    const int64_t partials_offset  = nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
+    const int64_t partial_size     = 2 + DV;
+    const float * partials_base    = (const float *) params->wdata + partials_offset;
+
+    // Output layout
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const size_t  nb1 = dst->nb[1];
+
+    // Each thread reduces a subset of query heads
+    for (int64_t q_head = ith; q_head < n_q_heads; q_head += nth) {
+        float   M_final   = -INFINITY;
+        float   S_final   = 0.0f;
+        float * VKQ_final = thread_wdata;
+        memset(VKQ_final, 0, DV * sizeof(float));
+
+        // Combine partials from all chunks
+        for (int64_t chunk_idx = 0; chunk_idx < n_chunks; ++chunk_idx) {
+            const int64_t ic_start = chunk_idx * chunk_size;
+            if (ic_start >= nek1) continue;
+
+            const float * partial   = partials_base + (q_head * n_chunks + chunk_idx) * partial_size;
+            const float   M_chunk   = partial[0];
+            const float   S_chunk   = partial[1];
+            const float * VKQ_chunk = partial + 2;
+
+            if (S_chunk == 0.0f) continue;
+
+            const float M_new     = fmaxf(M_final, M_chunk);
+            const float scale_old = expf(M_final - M_new);
+            const float scale_new = expf(M_chunk - M_new);
+
+            for (int64_t d = 0; d < DV; ++d) {
+                VKQ_final[d] = VKQ_final[d] * scale_old + VKQ_chunk[d] * scale_new;
+            }
+            S_final = S_final * scale_old + S_chunk * scale_new;
+            M_final = M_new;
+        }
+
+        // Normalize and write to output
+        if (S_final != 0.0f) {
+            const float S_inv = 1.0f / S_final;
+            ggml_vec_scale_f32(DV, VKQ_final, S_inv);
+        }
+        // iq1=0, iq3=0 for decode
+        memcpy((char *) dst->data + (0*ne2*ne1 + q_head + 0*ne1)*nb1, VKQ_final, nb1);
+    }
+}
+
 static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8567,6 +8647,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const int64_t DV = nev0;
     const int64_t N  = neq1;
 
+
     GGML_ASSERT(ne0 == DV);
     GGML_ASSERT(ne2 == N);
 
@@ -8587,60 +8668,92 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int64_t nr = neq1*neq2*neq3;
-
-    // rows per thread
     const int ith = params->ith;
     const int nth = params->nth;
 
-    // disable for NUMA
-    const bool disable_chunking = ggml_is_numa();
+    // When use_ref is set, force the vec-only reference implementation (no tiling, no KV-chunking)
+    const bool use_ref = params->use_ref;
 
-    // 4x chunks per thread
-    int nth_scaled = nth * 4;
-    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
-    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
-
-    if (nth == 1 || nchunk < nth || disable_chunking) {
-        nchunk = nth;
-    }
-
-    if (ith == 0) {
-        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-        ggml_threadpool_chunk_set(params->threadpool, nth);
-    }
-
-    ggml_barrier(params->threadpool);
-
-    // The number of elements in each chunk
-    const int64_t dr = (nr + nchunk - 1) / nchunk;
-
-    static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
-    static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
     const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
-    const bool use_tiled = (q->type == GGML_TYPE_F32 &&
-                            kv_is_f32_or_f16 &&
-                            k->type == v->type &&
-                            nek1 % KV_TILE_SZ == 0 &&
-                            neq1 >= Q_TILE_SZ);  // Only use tiled for batch >= tile size
+    const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;
 
-    // The first chunk comes from our thread_id, the rest will get auto-assigned.
-    int current_chunk = ith;
+    if (use_split_kv_path) {
+        const int64_t chunk_size = (nek1 + nth - 1) / nth;
 
-    while (current_chunk < nchunk) {
-        const int64_t ir0 = dr * current_chunk;
-        const int64_t ir1 = MIN(ir0 + dr, nr);
+        // Partials buffer layout: [q_head][kv_chunk][M, S, VKQ]
+        const int64_t partial_size  = 2 + DV;
+        float *       partials_base = (float *) params->wdata + nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
 
-        if (use_tiled) {
-            ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+        const int64_t ic_start = ith * chunk_size;
+        const int64_t ic_end   = std::min(ic_start + chunk_size, nek1);
+
+        const int64_t partial_stride = nth * partial_size;
+        float *       chunk_partials = partials_base + ith * partial_size;
+
+        if (ic_start < nek1) {
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(
+                    params, dst, q_head, q_head + 1, ic_start, ic_end,
+                    chunk_partials, partial_stride);
+            }
         } else {
-            ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                float * q_partials = chunk_partials + q_head * partial_stride;
+                q_partials[0] = -INFINITY;  // M
+                q_partials[1] = 0.0f;       // S
+            }
         }
 
-        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        ggml_barrier(params->threadpool);
+        ggml_flash_attn_ext_reduce_partials(params, dst, nth, chunk_size);
+    } else {
+
+        // total rows in q
+        const int64_t nr = neq1*neq2*neq3;
+
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        if (ith == 0) {
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+        static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
+        static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+        const bool use_tiled = !use_ref &&
+                               (q->type == GGML_TYPE_F32 &&
+                                kv_is_f32_or_f16 &&
+                                k->type == v->type &&
+                                nek1 % KV_TILE_SZ == 0 &&
+                                neq1 >= Q_TILE_SZ);
+
+        int current_chunk = ith;
+
+        while (current_chunk < nchunk) {
+            const int64_t ir0 = dr * current_chunk;
+            const int64_t ir1 = MIN(ir0 + dr, nr);
+
+            if (use_tiled) {
+                ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+            } else {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, nek1, nullptr, 0);
+            }
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
     }
 }
 

ggml/src/ggml-cpu/simd-mappings.h

@@ -116,6 +116,17 @@ extern "C" {
 // defined in ggml-cpu.c, initialized in ggml_cpu_init()
 extern float ggml_table_f32_f16[1 << 16];
 
+// precomputed f32 table for e8m0 half (1 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_e8m0_half[1 << 8];
+
+// Use lookup table for E8M0 on x86 (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
+#else
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
+#endif
+
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.

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

@@ -2279,13 +2279,19 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
     if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        if (ne2 == 1) {
+        static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
+        if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                if (ne2 <= 4) {
+                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                    return;
+                }
             } else {
-                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                if (GGML_CUDA_CC_IS_AMD(cc)) {
+                    ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                    return;
+                }
             }
-            return;
         }
 
         if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
@@ -5049,16 +5055,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
-            // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
-            // PR: https://github.com/ggml-org/llama.cpp/pull/19042
-            if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
-#ifdef _WIN32
-                _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
-#else
-                setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
-#endif // _WIN32
-            }
-
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
             const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 

ggml/src/ggml-cuda/mmq.cuh

@@ -3697,13 +3697,20 @@ static __global__ void mul_mat_q(
          tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
-
 template <ggml_type type, int mmq_x, bool need_check>
-static __global__ void mul_mat_q_stream_k_fixup(
-        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
-        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
-        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst,
-        const int ncols_max) {
+static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
+                                                const int32_t * expert_bounds,
+                                                float * __restrict__ dst,
+                                                const float * __restrict__ tmp_last_tile,
+                                                const int    ncols_x,
+                                                const int    nrows_x,
+                                                const int    ncols_dst,
+                                                const size_t stride_col_dst,
+                                                const int    nchannels_y,
+                                                const size_t stride_channel_dst,
+                                                const int    nsamples_y,
+                                                const size_t stride_sample_dst,
+                                                const int    ncols_max) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     ITER_K          = get_iter_k(type);

ggml/src/ggml-cuda/mmvf.cu

@@ -4,26 +4,48 @@
 #include "mmvf.cuh"
 #include "convert.cuh"
 
-template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false, bool is_multi_token_id = false>
 static __global__ void mul_mat_vec_f(
         const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
-        const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
+        const int ncols2, const uint3 nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
         const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const int ids_stride) {
     const int row         = blockIdx.x;
+    // for MUL_MAT_ID - blockIdx.y = n_expert_used, blockIdx.z = ncols_dst (tokens)
     const int channel_dst = blockIdx.y;
-    const int channel_x   = ids ? ids[channel_dst]          : fastdiv((uint32_t) channel_dst, channel_ratio);
-    const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
-    const int sample_dst  = blockIdx.z;
+    const int tid         = threadIdx.x;
+
+    int token_idx;
+    int channel_x;
+    int channel_y;
+    int sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        token_idx  = ids ? blockIdx.z                                          : 0;
+        channel_x  = ids ? ids[blockIdx.y + token_idx * ids_stride]            : fastdiv((uint32_t) channel_dst, channel_ratio);
+        channel_y  = ids ? fastmodulo(blockIdx.y, nchannels_y)                 : channel_dst;
+        sample_dst = ids ? 0                                                   : blockIdx.z;
+    }
+
     const int sample_x    = fastdiv((uint32_t) sample_dst, sample_ratio);
     const int sample_y    = sample_dst;
-    const int tid         = threadIdx.x;
 
     constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
     x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
     y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
     dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
+    if constexpr (is_multi_token_id) {
+        y   += token_idx*stride_col_y2*2;
+        dst += token_idx*stride_col_dst;
+    }
 
     bool use_gate = false;
     bool use_bias = false;
@@ -56,8 +78,10 @@ static __global__ void mul_mat_vec_f(
     if (use_gate) {
         gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
     }
+
+    const int channel_bias = ids ? channel_x : channel_dst;
+
     if constexpr (has_fusion) {
-        const int channel_bias = ids ? channel_x : channel_dst;
         if (use_bias) {
             x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
         }
@@ -349,36 +373,36 @@ static __global__ void mul_mat_vec_f(
     }
 }
 
-template<typename T, typename type_acc, int ncols_dst, int block_size>
+template<typename T, typename type_acc, int ncols_dst, int block_size, bool is_multi_token_id = false>
 static void mul_mat_vec_f_switch_fusion(
         const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
-        const int64_t ncols, const int64_t nrows,
+        const int64_t ncols, const uint3 nchannels_y,
         const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
         const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const int ids_stride, const cudaStream_t stream) {
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
        }
     }
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 
 }
 
-template <typename T, typename type_acc, int ncols_dst>
+template <typename T, typename type_acc, int ncols_dst, bool is_multi_token_id = false>
 void launch_mul_mat_vec_f_cuda(
         const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const int64_t ncols, const int64_t nrows,
@@ -386,12 +410,13 @@ void launch_mul_mat_vec_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        const int64_t nsamples_or_ntokens, const int64_t ids_stride, cudaStream_t stream) {
     GGML_ASSERT(ncols        % 2 == 0);
     GGML_ASSERT(stride_row   % 2 == 0);
     GGML_ASSERT(stride_col_y % 2 == 0);
     GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
     GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const uint3 nchannels_y_fd   = ids ? init_fastdiv_values(nchannels_y) : make_uint3(0, 0, 0);
     const uint3 channel_ratio_fd = ids ? make_uint3(0, 0, 0) : init_fastdiv_values(nchannels_dst / nchannels_x);
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
@@ -415,56 +440,56 @@ void launch_mul_mat_vec_f_cuda(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
 
     const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
-    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_or_ntokens);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case   64: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case   96: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  128: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  160: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  192: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  224: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         case  256: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -480,55 +505,88 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        cudaStream_t stream) {
+        const int64_t ids_stride, cudaStream_t stream) {
+
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst, true>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
+    if (has_ids) {
+        // Single-token MUL_MAT_ID path
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
     switch (ncols_dst) {
         case 1:
             launch_mul_mat_vec_f_cuda<T, type_acc, 1>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 2:
             launch_mul_mat_vec_f_cuda<T, type_acc, 2>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 3:
             launch_mul_mat_vec_f_cuda<T, type_acc, 3>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 4:
             launch_mul_mat_vec_f_cuda<T, type_acc, 4>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 5:
             launch_mul_mat_vec_f_cuda<T, type_acc, 5>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 6:
             launch_mul_mat_vec_f_cuda<T, type_acc, 6>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 7:
             launch_mul_mat_vec_f_cuda<T, type_acc, 7>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         case 8:
             launch_mul_mat_vec_f_cuda<T, type_acc, 8>
                 (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -544,21 +602,21 @@ static void mul_mat_vec_f_cuda(
         const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
-        enum ggml_prec prec, cudaStream_t stream) {
+        const int64_t ids_stride, enum ggml_prec prec, cudaStream_t stream) {
 
     if constexpr(std::is_same_v<T, half>) {
         if (prec == GGML_PREC_DEFAULT) {
             mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
                 (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             return;
         }
     }
     mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
         (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
 }
 
 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
@@ -573,7 +631,7 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const size_t ts_src1 = ggml_type_size(src1->type);
     const size_t ts_dst  = ggml_type_size(dst->type);
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVF_MAX_BATCH_SIZE);
     GGML_ASSERT(ne13 == ne3);
 
     GGML_ASSERT(        nb00       == ts_src0);
@@ -626,29 +684,31 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
     const int64_t ncols_dst          = ids ? ne2  : ne1;
     const int64_t nchannels_y        = ids ? ne11 : ne12;
     const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -695,19 +755,19 @@ void ggml_cuda_op_mul_mat_vec_f(
             const float * src0_d = (const float *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
             mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));

ggml/src/ggml-cuda/mmvf.cuh

@@ -1,5 +1,7 @@
 #include "common.cuh"
 
+#define MMVF_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVF kernels.
+
 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
     const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
 

ggml/src/ggml-cuda/mmvq.cu

@@ -137,15 +137,15 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-// tell the compiler to use as many registers as it wants, see nwarps definition below
-template <ggml_type type, int ncols_dst, bool has_fusion>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
-        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst) {
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const uint32_t ids_stride) {
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
@@ -162,11 +162,25 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
     const uint32_t channel_dst = blockIdx.y;
-    const uint32_t channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-    const uint32_t channel_y   = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-    const uint32_t sample_dst  = blockIdx.z;
+
+    uint32_t token_idx = 0;
+    uint32_t channel_x;
+    uint32_t channel_y;
+    uint32_t sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+        sample_dst = blockIdx.z;
+    }
+
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
@@ -188,11 +202,11 @@ static __global__ void mul_mat_vec_q(
         active_glu    = fusion.glu_op;
     }
 
-    const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
     float x_biases[ncols_dst]    = { 0.0f };
     float gate_biases[ncols_dst] = { 0.0f };
     if constexpr (has_fusion) {
+        const uint32_t channel_bias = ids ? channel_x : channel_dst;
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
             // 1. Hide latency by prefetching bias and gate here
@@ -222,6 +236,9 @@ static __global__ void mul_mat_vec_q(
     float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    if constexpr (is_multi_token_id) {
+        y += token_idx*stride_col_y;
+    }
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -275,6 +292,10 @@ static __global__ void mul_mat_vec_q(
 
     dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
 
+    if constexpr (is_multi_token_id) {
+        dst += token_idx*stride_col_dst;
+    }
+
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
@@ -335,40 +356,41 @@ static __global__ void mul_mat_vec_q(
 }
 
 static std::pair<dim3, dim3> calc_launch_params(
-        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
         const int warp_size, const mmvq_parameter_table_id table_id) {
     const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
-    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
     const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst>
+template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
         const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
+        const uint32_t ids_stride, cudaStream_t stream) {
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
         }
     }
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }
 
 template <ggml_type type>
@@ -379,7 +401,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
     GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
@@ -393,8 +415,19 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
+        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+             dims.first, dims.second, 0, ids_stride, stream);
+        return;
+    }
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
@@ -402,7 +435,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
@@ -410,7 +443,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
@@ -418,7 +451,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
@@ -426,7 +459,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
@@ -434,7 +467,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
@@ -442,7 +475,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
@@ -450,7 +483,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
@@ -458,7 +491,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -474,127 +507,127 @@ static void mul_mat_vec_q_switch_type(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
     switch (type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -622,7 +655,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVQ_MAX_BATCH_SIZE);
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -693,11 +726,13 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
+
     mul_mat_vec_q_switch_type(
         src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,               stream);
+        ne03,              ne3,           s03, s13,              s3,               ids_stride, stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -726,7 +761,7 @@ void ggml_cuda_op_mul_mat_vec_q(
     ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
         src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
-        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -534,6 +534,36 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    const int nsg = 8;
+    const int n   = op->src[1]->ne[1];
+    const int k   = op->src[1]->ne[0];
+
+    snprintf(base, 256, "kernel_solve_tri_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_nsg=%d_n=%d_k=%d", base, nsg, n, k);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, nsg, FC_SOLVE_TRI + 0);
+        ggml_metal_cv_set_int16(cv, n,   FC_SOLVE_TRI + 1);
+        ggml_metal_cv_set_int16(cv, k,   FC_SOLVE_TRI + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.nsg  = nsg;
+    res.smem = GGML_PAD(GGML_PAD(n, 32)*nsg*sizeof(float), 16);
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
     char base[256];
     char name[256];

ggml/src/ggml-metal/ggml-metal-device.h

@@ -121,6 +121,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1152,6 +1152,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return has_simdgroup_reduction;
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
+        case GGML_OP_SOLVE_TRI:
             return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:

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

@@ -78,13 +78,14 @@
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
 #define FC_SSM_CONV                    900
-#define FC_COUNT_EQUAL                 1000
+#define FC_SOLVE_TRI                   1000
+#define FC_COUNT_EQUAL                 1100
 
 // op-specific constants
-#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NQPSG 8
 #define OP_FLASH_ATTN_EXT_NCPSG 64
 
-#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
 // kernel argument structs
@@ -733,6 +734,33 @@ typedef struct {
     uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
     int32_t  ne00t;
     int32_t  ne00;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -341,6 +341,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_rwkv(ctx, idx);
             } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                n_fuse = ggml_metal_op_solve_tri(ctx, idx);
+            } break;
         case GGML_OP_MUL_MAT:
             {
                 n_fuse = ggml_metal_op_mul_mat(ctx, idx);
@@ -1557,6 +1561,63 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_kargs_solve_tri args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, (ne10 + nsg - 1)/nsg, ne02, ne03, 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -2295,7 +2356,7 @@ size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
     //    return res;
     //}
 
-    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPSG : OP_FLASH_ATTN_EXT_NQPSG;
     const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
 
     const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
@@ -2411,7 +2472,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
 
         GGML_ASSERT(nqptg <= 32);
@@ -2578,9 +2639,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int nkpsg = 1*ncpsg;
+        const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 1  == 0);
@@ -2632,6 +2693,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_op_concurrency_reset(ctx);
         }
 
+        // note: for simplicity assume the K is larger or equal than V
+        GGML_ASSERT(ne10 >= ne20);
+
         // ne00 + 2*ncpsg*(nsg)
         // for each query, we load it as f16 in shared memory (ne00)
         // and store the soft_max values and the mask
@@ -2639,28 +2703,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // ne20*(nsg)
         // each simdgroup has a full f32 head vector in shared mem to accumulate results
         //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*GGML_PAD(ne20, 128)*(nsg))*(sizeof(float)/2), 16))
-
-        int64_t nsgmax = 2;
-        while (true) {
-            const size_t smem = FATTN_SMEM(nsgmax);
-            // avoid using more than half of the threadgroup memory - can cause slow downs especially for large head sizes
-            if (smem > props_dev->max_theadgroup_memory_size/2) {
-                break;
-            }
-            nsgmax *= 2;
-        }
-        nsgmax /= 2;
-
-        // simdgroups per threadgroup (a.k.a. warps)
-        //const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
-        const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) 1024/32)));
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
-        while (nsg <= nsgt) {
-            nsg *= 2;
-        }
-        nsg /= 2;
 
         // workgroups
         // each workgroup handles nsg*nkpsg cache values
@@ -2673,7 +2718,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         } else {
             nwg = 32;
             nsg = 1;
-            while (2*nwg*nsg*nkpsg < ne11 && nsg < 4) {
+            while (2*nwg*nsg*ncpsg < ne11 && nsg < 4) {
                 nsg *= 2;
             }
         }
@@ -2739,7 +2784,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
         } else {
             // sanity checks
             assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
@@ -2752,7 +2797,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
 
             // sync the 2 kernels
             ggml_metal_op_concurrency_reset(ctx);

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -60,6 +60,7 @@ int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rwkv              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal.cpp

@@ -7,6 +7,9 @@
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
 
+#include <mutex>
+#include <string>
+
 #define GGML_METAL_NAME "MTL"
 #define GGML_METAL_MAX_DEVICES 16
 

ggml/src/ggml-metal/ggml-metal.metal

@@ -2737,6 +2737,83 @@ kernel void kernel_rwkv_wkv7_f32(
     }
 }
 
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t ne02 = args.ne02;
+    const int32_t ne03 = args.ne03;
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}
+
 kernel void kernel_argmax_f32(
         constant ggml_metal_kargs_argmax & args,
         device   const char * src0,
@@ -5931,7 +6008,7 @@ template<
     void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
     short DK,         // K head size
     short DV,         // V head size
-    short Q  = OP_FLASH_ATTN_EXT_NQPTG, // queries per threadgroup
+    short Q  = OP_FLASH_ATTN_EXT_NQPSG, // queries per threadgroup
     short C  = OP_FLASH_ATTN_EXT_NCPSG> // cache items per threadgroup
 kernel void kernel_flash_attn_ext(
         constant ggml_metal_kargs_flash_attn_ext & args,
@@ -6141,11 +6218,10 @@ template<
     void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
     short DK,       // K head size
     short DV,       // V head size
-    short NE,       // head elements per thread
-    short Q,        // queries per threadgroup
-    short C,        // cache items per threadgroup
-    short NSG>      // number of simd groups
-void kernel_flash_attn_ext_vec_impl(
+    short NE = 4,   // head elements per thread
+    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPSG,  // queries per threadgroup
+    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
+kernel void kernel_flash_attn_ext_vec(
         constant ggml_metal_kargs_flash_attn_ext_vec & args,
         device const char * q,
         device const char * k,
@@ -6162,6 +6238,7 @@ void kernel_flash_attn_ext_vec_impl(
     static_assert(DV % 32 == 0, "DV must be divisible by 32");
 
 #define NWG  (FC_flash_attn_ext_vec_nwg)
+#define NSG  (FC_flash_attn_ext_vec_nsg)
 
 #define NS10 (FC_flash_attn_ext_vec_ns10)
 #define NS20 (FC_flash_attn_ext_vec_ns20)
@@ -6190,12 +6267,12 @@ void kernel_flash_attn_ext_vec_impl(
 
     const short T = PK + NSG*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*PK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*PK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*PK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*PK); // same as above but in s4_t
-    threadgroup half  * sm  = (threadgroup half  *) (shmem_f16 +   sgitg*SH + 2*C + Q*PK); // scratch buffer for mask
-    threadgroup o4_t  * so4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*PV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                      0*PK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                      0*PK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + NSG*PK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + NSG*PK); // same as above but in s4_t
+    threadgroup half  * sm  = (threadgroup half  *) (shmem_f16 +   sgitg*SH + 2*C + NSG*PK); // scratch buffer for mask
+    threadgroup o4_t  * so4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*PV       + NSG*PK + NSG*SH); // scratch buffer for the results
 
     // store the result for all queries in shared memory (the O matrix from the paper)
     so4 += tiisg;
@@ -6213,11 +6290,13 @@ void kernel_flash_attn_ext_vec_impl(
     // load heads from Q to shared memory
     device const float4 * q4 = (device const float4 *) ((device const char *) q);
 
-    for (short i = tiisg; i < PK4; i += NW) {
-        if (iq1 < args.ne01 && i < DK4) {
-            sq4[i] = (q4_t) q4[i];
-        } else {
-            sq4[i] = (q4_t) 0.0f;
+    if (iq1 < args.ne01) {
+        for (short i = tiisg; i < PK4; i += NW) {
+            if (i < DK4) {
+                sq4[i] = (q4_t) q4[i];
+            } else {
+                sq4[i] = (q4_t) 0.0f;
+            }
         }
     }
 
@@ -6295,7 +6374,7 @@ void kernel_flash_attn_ext_vec_impl(
             }
 
             // skip -INF blocks
-            if (simd_max(sm[tiisg]) == -INFINITY) {
+            if (simd_max(sm[tiisg]) <= -MAXHALF) {
                 continue;
             }
 
@@ -6569,57 +6648,11 @@ void kernel_flash_attn_ext_vec_impl(
     }
 
 #undef NWG
+#undef NSG
 #undef NS10
 #undef NS20
 }
 
-template<
-    typename q4_t,  // query types in shared memory
-    typename k4_t,  // key types in shared memory
-    typename v4_t,  // value types in shared memory
-    typename qk_t,  // Q*K types
-    typename s_t,   // soft-max types
-    typename s4_t,
-    typename o4_t,  // attention accumulation types
-    typename kd4_t, // key type in device memory
-    short nl_k,
-    void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &),
-    typename vd4_t, // value type in device memory
-    short nl_v,
-    void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
-    short DK,       // K head size
-    short DV,       // V head size
-    short NE = 4,   // head elements per thread
-    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPTG,  // queries per threadgroup
-    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
-kernel void kernel_flash_attn_ext_vec(
-        constant ggml_metal_kargs_flash_attn_ext_vec & args,
-        device const char * q,
-        device const char * k,
-        device const char * v,
-        device const char * mask,
-        device const char * sinks,
-        device const char * pad,
-        device       char * dst,
-        threadgroup  half * shmem_f16 [[threadgroup(0)]],
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort  tiisg[[thread_index_in_simdgroup]],
-        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
-#define FWD_TMPL q4_t, k4_t, v4_t, qk_t, s_t, s4_t, o4_t, kd4_t, nl_k, deq_k_t4, vd4_t, nl_v, deq_v_t4, DK, DV, NE, Q, C
-#define FWD_ARGS args, q, k, v, mask, sinks, pad, dst, shmem_f16, tgpig, tiisg, sgitg
-    switch (FC_flash_attn_ext_vec_nsg) {
-      // note: disabled cases to reduce library load time
-        case 1:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  1>(FWD_ARGS); break;
-        case 2:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  2>(FWD_ARGS); break;
-        case 4:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  4>(FWD_ARGS); break;
-      //case 8:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  8>(FWD_ARGS); break;
-      //case 16: kernel_flash_attn_ext_vec_impl<FWD_TMPL, 16>(FWD_ARGS); break;
-      //case 32: kernel_flash_attn_ext_vec_impl<FWD_TMPL, 32>(FWD_ARGS); break;
-    }
-#undef FWD_TMPL
-#undef FWD_ARGS
-}
-
 // note: I think the s_t can be half instead of float, because the Q*K scaling is done before storing to shared mem
 //       in the other (non-vec) kernel, we need s_t to also be float because we scale during the soft_max
 //

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

@@ -453,7 +453,6 @@ struct ggml_backend_opencl_context {
     cl_program program_rms_norm;
     cl_program program_group_norm;
     cl_program program_rope;
-    cl_program program_scale;
     cl_program program_silu;
     cl_program program_sigmoid;
     cl_program program_softmax_f32;
@@ -462,11 +461,8 @@ struct ggml_backend_opencl_context {
     cl_program program_softmax_4_f16;
     cl_program program_argsort_f32_i32;
     cl_program program_sum_rows_f32;
-    cl_program program_repeat;
     cl_program program_pad;
-    cl_program program_tanh;
     cl_program program_upscale;
-    cl_program program_concat;
     cl_program program_conv_2d_f16;
     cl_program program_conv_2d_f32;
     cl_program program_conv_2d_f16_f32;
@@ -485,7 +481,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
     cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
     cl_kernel kernel_add_id;
-    cl_kernel kernel_scale;
+    cl_kernel kernel_scale_f32, kernel_scale_f32_4;
     cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
     cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
     cl_kernel kernel_mean_f32;
@@ -544,18 +540,17 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
-    cl_kernel kernel_repeat;
+    cl_kernel kernel_repeat_f32;
     cl_kernel kernel_pad;
-    cl_kernel kernel_tanh_f32_nd;
-    cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_tanh_f32, kernel_tanh_f32_4, kernel_tanh_f32_nc;
+    cl_kernel kernel_tanh_f16, kernel_tanh_f16_4, kernel_tanh_f16_nc;
     cl_kernel kernel_expm1_f32_nd;
     cl_kernel kernel_expm1_f16_nd;
     cl_kernel kernel_softplus_f32_nd;
     cl_kernel kernel_softplus_f16_nd;
     cl_kernel kernel_upscale;
     cl_kernel kernel_upscale_bilinear;
-    cl_kernel kernel_concat_f32_contiguous;
-    cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_concat_f32;
     cl_kernel kernel_conv_2d_f16;
     cl_kernel kernel_conv_2d_f32;
     cl_kernel kernel_conv_2d_f16_f32;
@@ -1483,10 +1478,12 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("scale.cl");
 #endif
-        backend_ctx->program_scale =
+        cl_program prog =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
 
-        CL_CHECK((backend_ctx->kernel_scale = clCreateKernel(backend_ctx->program_scale, "kernel_scale", &err), err));
+        CL_CHECK((backend_ctx->kernel_scale_f32   = clCreateKernel(prog, "kernel_scale_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_scale_f32_4 = clCreateKernel(prog, "kernel_scale_f32_4", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
         GGML_LOG_CONT(".");
     }
 
@@ -1814,16 +1811,11 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("repeat.cl");
 #endif
-        if (!kernel_src.empty()) {
-            backend_ctx->program_repeat =
-                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
-            CL_CHECK((backend_ctx->kernel_repeat = clCreateKernel(backend_ctx->program_repeat, "kernel_repeat", &err), err));
-            GGML_LOG_CONT(".");
-        } else {
-            GGML_LOG_WARN("ggml_opencl: repeat kernel source not found or empty. Repeat operations will not be available.\n");
-            backend_ctx->program_repeat = nullptr;
-            backend_ctx->kernel_repeat = nullptr;
-        }
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_repeat_f32 = clCreateKernel(prog, "kernel_repeat_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
     }
 
     // pad
@@ -1856,18 +1848,16 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 #else
         const std::string kernel_src = read_file("tanh.cl");
 #endif
-        if (!kernel_src.empty()) {
-            backend_ctx->program_tanh =
-                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
-            CL_CHECK((backend_ctx->kernel_tanh_f32_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f32_nd", &err), err));
-            CL_CHECK((backend_ctx->kernel_tanh_f16_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f16_nd", &err), err));
-            GGML_LOG_CONT(".");
-        } else {
-            GGML_LOG_WARN("ggml_opencl: tanh kernel source not found or empty. Tanh operation will not be available.\n");
-            backend_ctx->program_tanh = nullptr;
-            backend_ctx->kernel_tanh_f32_nd = nullptr;
-            backend_ctx->kernel_tanh_f16_nd = nullptr;
-        }
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_tanh_f32    = clCreateKernel(prog, "kernel_tanh_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f32_4  = clCreateKernel(prog, "kernel_tanh_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f32_nc = clCreateKernel(prog, "kernel_tanh_f32_nc", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16    = clCreateKernel(prog, "kernel_tanh_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16_4  = clCreateKernel(prog, "kernel_tanh_f16_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16_nc = clCreateKernel(prog, "kernel_tanh_f16_nc", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
     }
 
     // expm1
@@ -1959,22 +1949,13 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
             #include "concat.cl.h"
         };
 #else
-
         const std::string kernel_src = read_file("concat.cl");
 #endif
-        if (!kernel_src.empty()) {
-            backend_ctx->program_concat =
-                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
-
-            CL_CHECK((backend_ctx->kernel_concat_f32_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_contiguous", &err), err));
-            CL_CHECK((backend_ctx->kernel_concat_f32_non_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_non_contiguous", &err), err));
-            GGML_LOG_CONT(".");
-        } else {
-            GGML_LOG_WARN("ggml_opencl: concat kernel source not found or empty. Concat operations will not be available.\n");
-            backend_ctx->program_concat = nullptr;
-            backend_ctx->kernel_concat_f32_contiguous = nullptr;
-            backend_ctx->kernel_concat_f32_non_contiguous = nullptr;
-        }
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_concat_f32 = clCreateKernel(prog, "kernel_concat_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
     }
 
     // timestep_embedding
@@ -3318,8 +3299,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_SIGMOID:
                     return ggml_is_contiguous(op->src[0]);
                 case GGML_UNARY_OP_TANH:
-                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
-                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                   return op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16;
                 case GGML_UNARY_OP_EXPM1:
                    return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
                           (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
@@ -7029,79 +7009,87 @@ static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const
     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_abs = extra0->offset + src0->view_offs;
-    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
 
     cl_kernel kernel;
-    if (dst->type == GGML_TYPE_F32) {
-        kernel = backend_ctx->kernel_tanh_f32_nd;
-    } else if (dst->type == GGML_TYPE_F16) {
-        kernel = backend_ctx->kernel_tanh_f16_nd;
-    } else {
-        GGML_ASSERT(false && "Unsupported type for ggml_cl_tanh");
-    }
-    GGML_ASSERT(kernel != nullptr);
 
-    const int ne00 = src0->ne[0]; const int ne01 = src0->ne[1]; const int ne02 = src0->ne[2]; const int ne03 = src0->ne[3];
-    const cl_ulong nb00 = src0->nb[0]; const cl_ulong nb01 = src0->nb[1]; const cl_ulong nb02 = src0->nb[2]; const cl_ulong nb03 = src0->nb[3];
-
-    const int ne10 = dst->ne[0]; const int ne11 = dst->ne[1]; const int ne12 = dst->ne[2]; const int ne13 = dst->ne[3];
-    const cl_ulong nb10 = dst->nb[0]; const cl_ulong nb11 = dst->nb[1]; const cl_ulong nb12 = dst->nb[2]; const cl_ulong nb13 = dst->nb[3];
-
-    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
-    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
-    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
-    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
-
-    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
-    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
-    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
-    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
-    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
-    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
-
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
-    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
-    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
-    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
-    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
-    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
-    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
-
-    size_t global_work_size[3];
-    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
-        return;
-    }
-    global_work_size[0] = (size_t)ne10;
-    global_work_size[1] = (size_t)ne11;
-    global_work_size[2] = (size_t)ne12;
-
-    size_t lws0 = 16, lws1 = 4, lws2 = 1;
-    if (ne10 < 16) lws0 = ne10;
-    if (ne11 < 4) lws1 = ne11;
-    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
-
-    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
-    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
-    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
-
-
-    size_t local_work_size[] = {lws0, lws1, lws2};
-
-    size_t* local_work_size_ptr = local_work_size;
-    if (!backend_ctx->non_uniform_workgroups) {
-        if (global_work_size[0] % local_work_size[0] != 0 ||
-            global_work_size[1] % local_work_size[1] != 0 ||
-            global_work_size[2] % local_work_size[2] != 0) {
-            local_work_size_ptr = NULL;
+    if (ggml_is_contiguous(src0)) {
+        // Handle contiguous input
+        int n = ggml_nelements(dst);
+        if (n % 4 == 0) {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_tanh_f32_4;
+            } else {
+                kernel = backend_ctx->kernel_tanh_f16_4;
+            }
+            n /= 4;
+        } else {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_tanh_f32;
+            } else {
+                kernel = backend_ctx->kernel_tanh_f16;
+            }
         }
-    }
-    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
 
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+        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};
+
+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;
+        }
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+    } else {
+        // Handle non-contiguous input
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_tanh_f32_nc;
+        } else {
+            kernel = backend_ctx->kernel_tanh_f16_nc;
+        }
+
+        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));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb3));
+
+        int nth = 64;
+
+        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};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
 }
 
 static void ggml_cl_expm1(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -7319,53 +7307,58 @@ static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, con
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
-    if (backend_ctx->kernel_repeat == nullptr) {
-        GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__);
-        return;
-    }
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad  = (ggml_tensor_extra_cl *)dst->extra;
 
-    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
-    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd  = extrad->offset + dst->view_offs;
 
-    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
-    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const int src0_ne0 = src0->ne[0]; const int src0_ne1 = src0->ne[1]; const int src0_ne2 = src0->ne[2]; const int src0_ne3 = src0->ne[3];
-    const cl_ulong src0_nb0 = src0->nb[0]; const cl_ulong src0_nb1 = src0->nb[1]; const cl_ulong src0_nb2 = src0->nb[2]; const cl_ulong src0_nb3 = src0->nb[3];
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
-    const int dst_ne0 = dst->ne[0]; const int dst_ne1 = dst->ne[1]; const int dst_ne2 = dst->ne[2]; const int dst_ne3 = dst->ne[3];
-    const cl_ulong dst_nb0 = dst->nb[0]; const cl_ulong dst_nb1 = dst->nb[1]; const cl_ulong dst_nb2 = dst->nb[2]; const cl_ulong dst_nb3 = dst->nb[3];
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
 
-    cl_kernel kernel = backend_ctx->kernel_repeat;
+    const cl_ulong nb0 = dst->nb[0];
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
 
-    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
-    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),    &extra_dst->data_device));
-    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong),  &off_src0));
-    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
-    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &src0_ne0));
-    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &src0_ne1));
-    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &src0_ne2));
-    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &src0_ne3));
-    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong),  &src0_nb0));
-    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong),  &src0_nb1));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &src0_nb2));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &src0_nb3));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &dst_ne0));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &dst_ne1));
-    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &dst_ne2));
-    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dst_ne3));
-    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &dst_nb0));
-    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &dst_nb1));
-    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &dst_nb2));
-    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &dst_nb3));
+    cl_kernel kernel = backend_ctx->kernel_repeat_f32;
 
-    size_t gws0 = dst_ne1 > 0 ? (size_t)dst_ne1 : 1;
-    size_t gws1 = dst_ne2 > 0 ? (size_t)dst_ne2 : 1;
-    size_t gws2 = dst_ne3 > 0 ? (size_t)dst_ne3 : 1;
+    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));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb3));
 
-    size_t global_work_size[] = { gws0, gws1, gws2 };
+    int nth = 64;
 
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
+    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -7589,121 +7582,76 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
-    cl_command_queue queue = backend_ctx->queue;
 
-    if (backend_ctx->kernel_concat_f32_contiguous == nullptr || backend_ctx->kernel_concat_f32_non_contiguous == nullptr) {
-        GGML_LOG_WARN("%s: concat kernels not available, skipping OpenCL execution.\n", __func__);
-        return;
-    }
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
-    ggml_tensor_extra_cl * extra0_cl = (ggml_tensor_extra_cl *)src0->extra;
-    ggml_tensor_extra_cl * extra1_cl = (ggml_tensor_extra_cl *)src1->extra;
-    ggml_tensor_extra_cl * extrad_cl = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd  = extrad->offset + dst->view_offs;
 
-    cl_ulong off_src0 = extra0_cl->offset + src0->view_offs;
-    cl_ulong off_src1 = extra1_cl->offset + src1->view_offs;
-    cl_ulong off_dst  = extrad_cl->offset + dst->view_offs;
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
-    const int32_t dim = ((const int32_t *) dst->op_params)[0];
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
+
+    const cl_ulong nb0 = dst->nb[0];
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    const cl_int dim = ((const int32_t *) dst->op_params)[0];
     GGML_ASSERT(dim >= 0 && dim <= 3);
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-        if (dim == 3) {
+    int nth = MIN(64, ne0);
 
-            size_t nbytes_src0 = ggml_nbytes(src0);
-            size_t nbytes_src1 = ggml_nbytes(src1);
+    cl_kernel kernel = backend_ctx->kernel_concat_f32;
 
-            CL_CHECK(clEnqueueCopyBuffer(queue, extra0_cl->data_device, extrad_cl->data_device,
-                                         off_src0, off_dst, nbytes_src0, 0, NULL, NULL));
-            CL_CHECK(clEnqueueCopyBuffer(queue, extra1_cl->data_device, extrad_cl->data_device,
-                                         off_src1, off_dst + nbytes_src0, nbytes_src1, 0, NULL, NULL));
-        } else {
+    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),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_int),   &dim));
 
-            cl_kernel kernel = backend_ctx->kernel_concat_f32_contiguous;
-            size_t global_work_size[3];
+    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
 
-            for (int i3 = 0; i3 < dst->ne[3]; ++i3) {
-                cl_ulong current_off_src0 = off_src0 + (i3 * src0->nb[3]);
-                cl_ulong current_off_src1 = off_src1 + (i3 * src1->nb[3]);
-                cl_ulong current_off_dst  = off_dst  + (i3 * dst->nb[3]);
-
-                int d_ne00 = src0->ne[0]; int d_ne01 = src0->ne[1]; int d_ne02 = src0->ne[2];
-                int d_ne10 = src1->ne[0]; int d_ne11 = src1->ne[1]; int d_ne12 = src1->ne[2];
-                int d_ne0  = dst->ne[0];  int d_ne1  = dst->ne[1];  int d_ne2  = dst->ne[2];
-
-                CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
-                CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &current_off_src0));
-                CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
-                CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &current_off_src1));
-                CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
-                CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &current_off_dst));
-                CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &d_ne00));
-                CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne01));
-                CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne02));
-                CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne10));
-                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &d_ne11));
-                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &d_ne12));
-                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &d_ne0));
-                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &d_ne1));
-                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &d_ne2));
-                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dim));
-
-                global_work_size[0] = d_ne0;
-                global_work_size[1] = d_ne1;
-                global_work_size[2] = d_ne2;
-
-                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
-            }
-        }
-    } else {
-        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
-
-        cl_long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
-        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
-
-        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
-
-        cl_long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
-        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
-
-
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_src1));
-        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
-
-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_long),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_long),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_long),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_long),      &ne03));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
-        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
-        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),    &nb03));
-
-        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),    &nb10));
-        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),    &nb11));
-        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
-        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
-
-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_long),     &d_ne0));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_long),     &d_ne1));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_long),     &d_ne2));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_long),     &d_ne3));
-        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
-        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
-        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
-        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong),    &d_nb3));
-        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &dim));
-
-        size_t global_work_size_nc[] = { d_ne1 > 0 ? (size_t)d_ne1 : 1,
-                                         d_ne2 > 0 ? (size_t)d_ne2 : 1,
-                                         d_ne3 > 0 ? (size_t)d_ne3 : 1 };
-
-        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size_nc, NULL, dst);
-    }
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -8394,6 +8342,7 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
     CL_CHECK(clReleaseMemObject(D_sub_buffer));
     CL_CHECK(clReleaseMemObject(D_image1d));
 #else
+    GGML_UNUSED(backend);
     GGML_UNUSED(src0);
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);
@@ -9913,7 +9862,16 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
-    cl_kernel kernel = backend_ctx->kernel_scale;
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_scale_f32_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_scale_f32;
+    }
 
     CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
     CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
@@ -9922,8 +9880,6 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &scale));
     CL_CHECK(clSetKernelArg(kernel, 5, sizeof(float),    &bias));
 
-    int n = ggml_nelements(dst)/4;
-
     size_t global_work_size[] = {(size_t)n, 1, 1};
     size_t local_work_size[] = {64, 1, 1};
 

ggml/src/ggml-opencl/kernels/concat.cl

@@ -1,109 +1,51 @@
-kernel void kernel_concat_f32_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
-    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
-    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
-    int dim
+kernel void kernel_concat_f32(
+    global  const char * src0,
+    ulong                offset0,
+    global  const char * src1,
+    ulong                offset1,
+    global        char * dst,
+    ulong                offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb00,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim
 ) {
-    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
-    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
-    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
 
-    int i0 = get_global_id(0); // Index along dst's 0th dimension
-    int i1 = get_global_id(1); // Index along dst's 1st dimension
-    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
 
-    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
-        return;
-    }
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
-    ulong src_idx;
+    global const float * x;
 
-    if (dim == 0) {
-        if (i0 < d_ne00) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
-            dst[dst_idx] = src1[src_idx];
-        }
-    } else if (dim == 1) {
-        if (i1 < d_ne01) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
-        }
-    } else if (dim == 2) {
-        if (i2 < d_ne02) { // Data from src0
-            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
-            dst[dst_idx] = src0[src_idx];
-        } else { // Data from src1
-
-            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
-            dst[dst_idx] = src1[src_idx];
-        }
-    }
-}
-
-kernel void kernel_concat_f32_non_contiguous(
-    global const char * p_src0, ulong off_src0,
-    global const char * p_src1, ulong off_src1,
-    global char * p_dst, ulong off_dst,
-
-    long ne00, long ne01, long ne02, long ne03,
-    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
-
-    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
-
-    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
-    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
-    int dim
-) {
-    global const char * src0_base = p_src0 + off_src0;
-    global const char * src1_base = p_src1 + off_src1;
-    global char * dst_base        = p_dst + off_dst;
-
-    long current_i1 = get_global_id(0); // Index for dst_dim_1
-    long current_i2 = get_global_id(1); // Index for dst_dim_2
-    long current_i3 = get_global_id(2); // Index for dst_dim_3
-
-    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
-        return;
-    }
-
-    global const float * x_val_ptr;
-    global float * y_val_ptr;
-
-    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
-        bool use_src0;
-        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
-
-        if (dim == 0) {
-            use_src0 = (current_i0 < ne00);
-            if (!use_src0) { s_i0 = current_i0 - ne00; }
-        } else if (dim == 1) {
-            use_src0 = (current_i1 < ne01);
-            if (!use_src0) { s_i1 = current_i1 - ne01; }
-        } else if (dim == 2) {
-            use_src0 = (current_i2 < ne02);
-            if (!use_src0) { s_i2 = current_i2 - ne02; }
-        } else { // dim == 3
-            use_src0 = (current_i3 < ne03);
-            if (!use_src0) { s_i3 = current_i3 - ne03; }
-        }
-
-        if (use_src0) {
-            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
         } else {
-            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
         }
 
-        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
-        *y_val_ptr = *x_val_ptr;
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
     }
 }

ggml/src/ggml-opencl/kernels/repeat.cl

@@ -1,39 +1,38 @@
-kernel void kernel_repeat(
-    global const char * src0_data_in,
-    global       char * dst_data_in,
-    ulong src0_offset,
-    ulong dst_offset,
-    int src0_ne0, int src0_ne1, int src0_ne2, int src0_ne3,
-    ulong src0_nb0, ulong src0_nb1, ulong src0_nb2, ulong src0_nb3,
-    int dst_ne0, int dst_ne1, int dst_ne2, int dst_ne3,
-    ulong dst_nb0, ulong dst_nb1, ulong dst_nb2, ulong dst_nb3
+kernel void kernel_repeat_f32(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int     ne00,
+        int     ne01,
+        int     ne02,
+        int     ne03,
+        ulong   nb00,
+        ulong   nb01,
+        ulong   nb02,
+        ulong   nb03,
+        int     ne0,
+        ulong   nb0,
+        ulong   nb1,
+        ulong   nb2,
+        ulong   nb3
 ) {
-    global const char * src0_data = src0_data_in + src0_offset;
-    global       char * dst_data  = dst_data_in + dst_offset;
+    src0 = src0 + offset0;
+    dst  = dst  + offsetd;
 
-    const int d3 = get_global_id(2);
-    const int d2 = get_global_id(1);
-    const int d1 = get_global_id(0);
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
 
-    if (d3 >= dst_ne3 || d2 >= dst_ne2 || d1 >= dst_ne1) {
-        return;
-    }
+    const int i03 = i3%ne03;
+    const int i02 = i2%ne02;
+    const int i01 = i1%ne01;
 
-    const int s3 = d3 % src0_ne3;
-    const int s2 = d2 % src0_ne2;
-    const int s1 = d1 % src0_ne1;
+    global const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1;
 
-    const global char * p_src0_slice = src0_data + (ulong)s3*src0_nb3 + (ulong)s2*src0_nb2 + (ulong)s1*src0_nb1;
-    global char * p_dst_slice  = dst_data  + (ulong)d3*dst_nb3 + (ulong)d2*dst_nb2 + (ulong)d1*dst_nb1;
-
-    for (int d0 = 0; d0 < dst_ne0; ++d0) {
-        // Determine source index for dimension 0 based on tiling/broadcasting.
-        const int s0 = d0 % src0_ne0;
-
-        const global char * restrict current_src_el_ptr = p_src0_slice + (ulong)s0*src0_nb0;
-        global char * restrict current_dst_el_ptr  = p_dst_slice  + (ulong)d0*dst_nb0;
-        for (int k = 0; k < src0_nb0; ++k) {
-            current_dst_el_ptr[k] = current_src_el_ptr[k];
-        }
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i00 = i0%ne00;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i00*nb00));
     }
 }

ggml/src/ggml-opencl/kernels/scale.cl

@@ -1,9 +1,19 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
-//------------------------------------------------------------------------------
-// scale
-//------------------------------------------------------------------------------
-kernel void kernel_scale(
+kernel void kernel_scale_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float scale,
+        float bias
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
+}
+
+kernel void kernel_scale_f32_4(
         global float4 * src0,
         ulong offset0,
         global float4 * dst,

ggml/src/ggml-opencl/kernels/tanh.cl

@@ -1,63 +1,109 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 
-#ifdef cl_intel_required_subgroup_size
-#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
-#define INTEL_GPU 1
-#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
-#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
-#elif defined(cl_qcom_reqd_sub_group_size)
-#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
-#define ADRENO_GPU 1
-#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
-#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#endif
-
-kernel void kernel_tanh_f32_nd(
-    global void * p_src0_base, ulong off_src0_abs,
-    global void * p_dst_base,  ulong off_dst_abs,
-    int ne00, int ne01, int ne02, int ne03,
-    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
-    int ne10, int ne11, int ne12, int ne13,
-    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+kernel void kernel_tanh_f32(
+        global const float * src0,
+        ulong                offset0,
+        global       float * dst,
+        ulong                offsetd
 ) {
-    int i0 = get_global_id(0);
-    int i1 = get_global_id(1);
-    int i2 = get_global_id(2);
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
 
-    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
-        for (int i3 = 0; i3 < ne13; ++i3) {
-            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
-            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
 
-            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
-            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+kernel void kernel_tanh_f32_4(
+        global const float4 * src0,
+        ulong                 offset0,
+        global       float4 * dst,
+        ulong                 offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
 
-            *dst_val_ptr = tanh(*src_val_ptr);
-        }
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f16(
+        global const half * src0,
+        ulong               offset0,
+        global       half * dst,
+        ulong               offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f16_4(
+        global const half4 * src0,
+        ulong                offset0,
+        global       half4 * dst,
+        ulong                offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f32_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const float * x = (global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       float * y = (global       float *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = tanh(*x);
     }
 }
 
-kernel void kernel_tanh_f16_nd(
-    global void * p_src0_base, ulong off_src0_abs,
-    global void * p_dst_base,  ulong off_dst_abs,
-    int ne00, int ne01, int ne02, int ne03,
-    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
-    int ne10, int ne11, int ne12, int ne13,
-    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+kernel void kernel_tanh_f16_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
 ) {
-    int i0 = get_global_id(0);
-    int i1 = get_global_id(1);
-    int i2 = get_global_id(2);
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
 
-    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
-        for (int i3 = 0; i3 < ne13; ++i3) {
-            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
-            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
 
-            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
-            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const half * x = (global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       half * y = (global       half *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-            *dst_val_ptr = tanh(*src_val_ptr);
-        }
+        *y = tanh(*x);
     }
 }

ggml/src/ggml-sycl/CMakeLists.txt

@@ -1,7 +1,7 @@
 message(STATUS  "GGML_SYCL_TARGET=${GGML_SYCL_TARGET}")
 
-if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
-    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
+if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL)$")
+    message(FATAL_ERROR "GGML_SYCL_TARGET: Invalid target, the supported options are [INTEL]")
 endif()
 
 check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
@@ -125,25 +125,22 @@ endif()
 target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
 
 if (GGML_SYCL_F16)
-    if (GGML_SYCL_TARGET STREQUAL "AMD")
-        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
-    endif()
     add_compile_definitions(GGML_SYCL_F16)
 endif()
 
 if (GGML_SYCL_TARGET STREQUAL "INTEL")
     add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
     target_link_options(ggml-sycl PRIVATE  -Xs   -ze-intel-greater-than-4GB-buffer-required)
-elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-    # INFO: Allowed Sub_group_sizes are not consistent through all
-    # hip targets. For example, 64 is used for certain models, but the backend
-    # does not support it.
-    # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
-    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+
+    # Link against Intel oneMKL
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+        set(SYCL_COMPILER ON)
+    endif()
+    find_package(MKL REQUIRED)
+    target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
 else()
     # default for other target
+    message(FATAL_ERROR "GGML_SYCL_TARGET is not supported")
     add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
 endif()
 
@@ -151,82 +148,6 @@ if (GGML_SYCL_GRAPH)
     target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
 endif()
 
-# Link against Intel oneMKL or oneMath
-if (GGML_SYCL_TARGET STREQUAL "INTEL")
-    # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically
-    # See https://github.com/uxlfoundation/oneMath/issues/654
-    if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-        set(SYCL_COMPILER ON)
-    endif()
-    find_package(MKL REQUIRED)
-    target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
-    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL)
-else()
-    find_package(oneMath QUIET)
-    if (NOT oneMath_FOUND)
-        message(STATUS "oneMath not found: oneMath will be automatically downloaded")
-        # Use FetchContent to automatically pull and build oneMath
-        include(FetchContent)
-        set(BUILD_FUNCTIONAL_TESTS False)
-        set(BUILD_EXAMPLES False)
-        set(TARGET_DOMAINS blas)
-        if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-            set(ENABLE_MKLCPU_BACKEND False)
-            set(ENABLE_MKLGPU_BACKEND False)
-            set(ENABLE_CUBLAS_BACKEND True)
-        elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-            set(ENABLE_MKLCPU_BACKEND False)
-            set(ENABLE_MKLGPU_BACKEND False)
-            set(ENABLE_ROCBLAS_BACKEND True)
-            # Ensure setting a string variable here is not overriden by oneMath CACHE variables
-            cmake_policy(SET CMP0126 NEW)
-            # Setting the device architecture is only needed and useful for AMD devices in oneMath
-            set(HIP_TARGETS ${GGML_SYCL_DEVICE_ARCH} CACHE STRING "oneMath HIP target" FORCE)
-        endif()
-        FetchContent_Declare(
-            ONEMATH
-            GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
-        )
-        FetchContent_MakeAvailable(ONEMATH)
-        # Create alias to match with find_package targets name
-        function(onemath_alias target)
-            if (TARGET ${target}_obj)
-                # Silence verbose warnings from external libraries
-                target_compile_options(${target}_obj PRIVATE -w)
-            endif()
-            if (TARGET ${target})
-                add_library(ONEMATH::${target} ALIAS ${target})
-            endif()
-        endfunction()
-        onemath_alias(onemath)
-        onemath_alias(onemath_blas_mklcpu)
-        onemath_alias(onemath_blas_mklgpu)
-        onemath_alias(onemath_blas_cublas)
-        onemath_alias(onemath_blas_rocblas)
-    endif()
-
-    # Below oneMath compile-time dispatching is used for better performance
-    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_cublas)
-        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
-        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda")
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
-    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        if (NOT GGML_SYCL_DEVICE_ARCH)
-            message(FATAL_ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
-        endif()
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
-        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
-        target_link_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_AMD)
-    else()
-        # Fallback to oneMath runtime dispatcher
-        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath)
-        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GENERIC)
-    endif()
-endif()
-
 if (GGML_SYCL_DEVICE_ARCH)
     target_compile_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
     target_link_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -15,17 +15,9 @@
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
-#include <map>
-
-#ifdef GGML_SYCL_USE_INTEL_ONEMKL
 #include <oneapi/mkl.hpp>
-// Allow to use the same namespace for Intel oneMKL and oneMath
-namespace oneapi {
-    namespace math = mkl;
-}
-#else
-#include <oneapi/math.hpp>
-#endif
+
+#include <map>
 
 #include "ggml.h"
 
@@ -91,32 +83,13 @@ inline std::string get_device_backend_and_type(const sycl::device &device) {
 }
 
 template <typename Ts> struct matrix_info_t {
-    oneapi::math::transpose transpose_info[2];
+    oneapi::mkl::transpose transpose_info[2];
     Ts                     value_info[2];
     std::int64_t           size_info[3];
     std::int64_t           ld_info[3];
     std::int64_t           groupsize_info;
 };
 
-inline auto get_onemath_backend(sycl::queue& queue)
-#if defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
-  -> sycl::queue&
-#endif
-{
-// If the backend is known at compile-time, use oneMath backend_selector to use
-// compile-time dispatching and avoid the need to dlopen libraries. Otherwise
-// fallback to runtime dispatching.
-#if defined(GGML_SYCL_NVIDIA)
-    return oneapi::math::backend_selector<oneapi::math::backend::cublas>{ queue };
-#elif defined(GGML_SYCL_AMD)
-    return oneapi::math::backend_selector<oneapi::math::backend::rocblas>{ queue };
-#elif defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
-    return queue;
-#else
-    static_assert(false, "Unsupported backend");
-#endif
-}
-
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;
@@ -1734,7 +1707,7 @@ namespace dpct
     namespace detail
     {
     template <class Ta, class Tb, class Tc, class Ts>
-    inline void gemm_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+    inline void gemm_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
                           int n, int k, const void * alpha, const void * a, int lda, const void * b, int ldb,
                           const void * beta, void * c, int ldc) {
         Ts   alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
@@ -1742,7 +1715,7 @@ namespace dpct
         auto data_a      = get_memory<const Ta>(a);
         auto data_b      = get_memory<const Tb>(b);
         auto data_c      = get_memory<Tc>(c);
-        oneapi::math::blas::column_major::gemm(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value, data_a,
+        oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a,
                                                lda, data_b, ldb, beta_value, data_c, ldc);
     }
 
@@ -1774,7 +1747,7 @@ namespace dpct
         };
 
         template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
+        inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans,
                                     int m, int n, int k, const void * alpha, const void ** a, int lda, const void ** b,
                                     int ldb, const void * beta, void ** c, int ldc, int batch_size,
                                     matrix_info_t<float> * matrix_info) {
@@ -1793,8 +1766,8 @@ namespace dpct
             matrix_info->ld_info[2] = ldc;
             matrix_info->groupsize_info = batch_size;
 
-            sycl::event e = oneapi::math::blas::column_major::gemm_batch(
-                get_onemath_backend(q), matrix_info->transpose_info, matrix_info->transpose_info + 1,
+            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
+                q, matrix_info->transpose_info, matrix_info->transpose_info + 1,
                 matrix_info->size_info, matrix_info->size_info + 1, matrix_info->size_info + 2,
                 reinterpret_cast<Ts *>(matrix_info->value_info), reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
                 reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
@@ -1803,7 +1776,7 @@ namespace dpct
         }
 
         template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
+        inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans,
                                     int m, int n, int k, const void * alpha, const void * a, int lda,
                                     long long int stride_a, const void * b, int ldb, long long int stride_b,
                                     const void * beta, void * c, int ldc, long long int stride_c, int batch_size) {
@@ -1812,7 +1785,7 @@ namespace dpct
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
-            oneapi::math::blas::column_major::gemm_batch(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value,
+            oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value,
                                                          data_a, lda, stride_a, data_b, ldb, stride_b, beta_value,
                                                          data_c, ldc, stride_c, batch_size);
         }
@@ -2299,7 +2272,7 @@ namespace dpct
                            sycl::range<3>(x, y, 1), direction);
     }
 
-    inline void gemm(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n,
+    inline void gemm(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m, int n,
                      int k, const void * alpha, const void * a, library_data_t a_type, int lda, const void * b,
                      library_data_t b_type, int ldb, const void * beta, void * c, library_data_t c_type, int ldc,
                      library_data_t scaling_type) {
@@ -2366,7 +2339,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
+            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
             break;
         }
@@ -2405,7 +2378,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
+            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
             break;
         }
@@ -2447,7 +2420,7 @@ namespace dpct
     /// \param [in] ldc Leading dimension of C.
     /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
     /// \param [in] scaling_type Data type of the scaling factors.
-    inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+    inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
                            int n, int k, const void * alpha, const void * a[], library_data_t a_type, int lda,
                            const void * b[], library_data_t b_type, int ldb, const void * beta, void * c[],
                            library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type,
@@ -2485,7 +2458,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
             break;
         }
@@ -2493,7 +2466,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
             break;
         }
@@ -2569,7 +2542,7 @@ namespace dpct
     /// \param [in] stride_c Stride between the different C matrices.
     /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
     /// \param [in] scaling_type Data type of the scaling factors.
-    inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+    inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
                            int n, int k, const void * alpha, const void * a, library_data_t a_type, int lda,
                            long long int stride_a, const void * b, library_data_t b_type, int ldb,
                            long long int stride_b, const void * beta, void * c, library_data_t c_type, int ldc,
@@ -2642,7 +2615,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
                 batch_size);
             break;
@@ -2651,7 +2624,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
                 batch_size);
             break;

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -2167,8 +2167,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
             const sycl::half alpha_f16 = 1.0f;
             const sycl::half beta_f16  = 0.0f;
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
-                *stream, oneapi::math::transpose::trans,
-                oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
+                *stream, oneapi::mkl::transpose::trans,
+                oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
                 &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
                 src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
                 dst_f16.get(), dpct::library_data_t::real_half, ldc,
@@ -2211,8 +2211,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
         {
             const float alpha = 1.0f;
             const float beta  = 0.0f;
-            SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
-                get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
+            SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+                *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff,
                 src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
                 dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
         }
@@ -3165,8 +3165,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             const int64_t smb = ne12 == 1 ? s13       : s12;
 
             // there is no broadcast and src0, src1 are contiguous across dims 2, 3
-            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
-                                                        oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::mkl::transpose::trans,
+                                                        oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
                                                         src0_f16, dpct::library_data_t::real_half, nb01 / nb00, sma,
                                                         src1_f16, dpct::library_data_t::real_half, s11, smb, beta, dst_ddf,
                                                         mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
@@ -3190,7 +3190,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             });
 
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
-                *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                *queue, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
                 (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
                 (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, s11, beta,
                 (void **) (ptrs_dst.get() + 0 * ne23), mkl_data_type, ne0, ne23, mkl_compute_type, matrix_info.get())));
@@ -3524,12 +3524,11 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // SYCL_USE_XMX
 
-    // mmvq path is faster in the CUDA backend.
-    if (!g_ggml_sycl_prioritize_dmmv && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda
-        // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization
-        // is enabled takes precedence over DMMV, the current if-else implementation
-        // requires disabling DMMV if both conditions are met
-        || (should_reorder_tensor(ctx, dst) && ggml_sycl_supports_reorder_mmvq(src0->type)))) {
+    // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization
+    // is enabled takes precedence over DMMV, the current if-else implementation
+    // requires disabling DMMV if both conditions are met
+    if (!g_ggml_sycl_prioritize_dmmv && ((should_reorder_tensor(ctx, dst) &&
+                                          ggml_sycl_supports_reorder_mmvq(src0->type)))) {
         use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
     }
 
@@ -4189,16 +4188,6 @@ void ggml_backend_sycl_get_device_memory(int device, size_t *free,
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_memory\n");
     ggml_sycl_set_device(device);
 
-    /*
-    DPCT1009:218: SYCL uses exceptions to report errors and does not use the
-    error codes. The original code was commented out and a warning string was
-    inserted. You need to rewrite this code.
-    */
-    /*
-    DPCT1106:217: 'cudaMemGetInfo' was migrated with the Intel extensions for
-    device information which may not be supported by all compilers or runtimes.
-    You may need to adjust the code.
-    */
     SYCL_CHECK(CHECK_TRY_ERROR(
         dpct::dev_mgr::instance().get_device(device).get_memory_info(*free, *total)));
 }

ggml/src/ggml-sycl/outprod.cpp

@@ -32,12 +32,12 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Handle transposition of src1
     const bool src1_T = ggml_is_transposed(src1);
-    const oneapi::math::transpose src1_op = src1_T ? oneapi::math::transpose::nontrans : oneapi::math::transpose::trans;
+    const oneapi::mkl::transpose src1_op = src1_T ? oneapi::mkl::transpose::nontrans : oneapi::mkl::transpose::trans;
     const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float);
 
     try {
-        // Perform matrix multiplication using oneMath GEMM
-        oneapi::math::blas::column_major::gemm(get_onemath_backend(*stream), oneapi::math::transpose::nontrans, src1_op,
+        // Perform matrix multiplication using oneMKL GEMM
+        oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op,
                                                ne0, ne1, ne01, alpha, src0_d, ne00, src1_d, ldb, beta, dst_d, ne0);
     }
     catch (sycl::exception const& exc) {

ggml/src/ggml-sycl/rope.cpp

@@ -207,7 +207,6 @@ static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, cons
         const int p = sector;
         theta_base  = pos[channel_x] * sycl::pow(theta_scale, (float) p);
     } else {
-        // Simplified from CUDA backend code: if (sector >= sections.v[0] && sector < sec_w) which is just sector >= sections.v[0]
         const int p = sector - sections.v[0];
         theta_base  = pos[channel_x + ne2] * sycl::pow(theta_scale, (float) p);
     }

ggml/src/ggml-sycl/wkv.cpp

@@ -1,7 +1,7 @@
 #include <sycl/sycl.hpp>
 #include "wkv.hpp"
 
-constexpr int WKV_BLOCK_SIZE = 64;  // Matching CUDA_WKV_BLOCK_SIZE
+constexpr int WKV_BLOCK_SIZE = 64;
 
 // Helper function for the main kernel
 template <int block_size>

ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp

@@ -36,7 +36,7 @@ apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) {
     result.data      = reinterpret_cast<uint64_t>(tensor->data);
     if (tensor->data) {
         if (!tensor->buffer) {
-            GGML_ABORT("tensor has data but not buffer");
+            GGML_ABORT("%s: tensor has data but not buffer", __func__);
         }
         // tensor->data is serialized as an offset to the buffer base address
         result.data -= reinterpret_cast<uint64_t>(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base);

ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp

@@ -27,7 +27,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
 
     const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
     if (!shmem_data) {
-        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
         apir_decoder_set_fatal(dec);
         return 1;
     }
@@ -45,7 +45,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
         if (dev->iface.supports_op(dev, op)) {
             continue;
         }
-        GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
 
         status = GGML_STATUS_ABORTED;
         apir_encode_ggml_status(enc, &status);

ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp

@@ -36,18 +36,22 @@ uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec
     ggml_backend_buffer_type_t buft;
     buft = apir_decode_ggml_buffer_type(dec);
 
-    size_t value = buft->iface.get_max_size(buft);
+    size_t value = SIZE_MAX;
+    if (buft->iface.get_max_size) {
+        value = buft->iface.get_max_size(buft);
+    }
+
     apir_encode_size_t(enc, &value);
 
     return 0;
 }
 
+/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
 uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
     GGML_UNUSED(ctx);
-    ggml_backend_buffer_type_t buft;
-    buft = apir_decode_ggml_buffer_type(dec);
+    GGML_UNUSED(dec);
+    const bool is_host = false;
 
-    bool is_host = buft->iface.is_host(buft);
     apir_encode_bool_t(enc, &is_host);
 
     return 0;

ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp

@@ -40,7 +40,7 @@ uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl
     void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
 
     if (!shmem_data) {
-        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
         return 1;
     }
 
@@ -71,7 +71,7 @@ uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl
 
     void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
     if (!shmem_data) {
-        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
         return 1;
     }
 
@@ -121,7 +121,7 @@ uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virg
     buffer = apir_decode_ggml_buffer(dec);
 
     if (!apir_untrack_backend_buffer(buffer)) {
-        GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer);
+        GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: unknown buffer %p\n", __func__, (void *) buffer);
         return 1;
     }
 

ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp

@@ -124,7 +124,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
 
     void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
     if (!shmem_ptr) {
-        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
         apir_decoder_set_fatal(dec);
         return 1;
     }

ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp

@@ -17,26 +17,26 @@ uint64_t timer_count = 0;
 
 uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
     if (reg != NULL) {
-        GGML_LOG_WARN("%s: already initialized\n", __func__);
+        GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: already initialized\n", __func__);
         return APIR_BACKEND_INITIALIZE_ALREADY_INITED;
     }
     ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p;
 
     reg = ggml_backend_reg_fct();
     if (reg == NULL) {
-        GGML_LOG_ERROR("%s: backend registration failed\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend registration failed\n", __func__);
         return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
     }
 
     if (!reg->iface.get_device_count(reg)) {
-        GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device found\n", __func__);
         return APIR_BACKEND_INITIALIZE_NO_DEVICE;
     }
 
     dev = reg->iface.get_device(reg, 0);
 
     if (!dev) {
-        GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device received\n", __func__);
         return APIR_BACKEND_INITIALIZE_NO_DEVICE;
     }
 

ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h

@@ -16,6 +16,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
 uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
 uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
 uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
 uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
 uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
 uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
@@ -62,7 +63,7 @@ static inline const char * backend_dispatch_command_name(ApirBackendCommandType
         case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
             return "backend_buffer_type_get_max_size";
         case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
-            return "backend_buffer_type_is_host";
+            return "backend_buffer_type_is_host (DEPRECATED)";
         case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
             return "backend_buffer_type_alloc_buffer";
         case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
@@ -110,7 +111,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
     /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME  = */ backend_buffer_type_get_name,
     /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT  = */ backend_buffer_type_get_alignment,
     /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE  = */ backend_buffer_type_get_max_size,
-    /* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST  = */ backend_buffer_type_is_host,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST  = */ backend_buffer_type_is_host /* DEPRECATED */,
     /* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER  = */ backend_buffer_type_alloc_buffer,
     /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE  = */ backend_buffer_type_get_alloc_size,
 

ggml/src/ggml-virtgpu/backend/backend-dispatched.h

@@ -11,6 +11,8 @@
 #include "shared/apir_cs.h"
 #include "shared/apir_cs_ggml.h"
 
+#define GGML_VIRTGPU_BCK "ggml-virtgpu-backend: "
+
 struct virgl_apir_context {
     uint32_t               ctx_id;
     virgl_apir_callbacks * iface;

ggml/src/ggml-virtgpu/backend/backend.cpp

@@ -35,14 +35,8 @@ void apir_backend_deinit(uint32_t virgl_ctx_id) {
         buffer->iface.free_buffer(buffer);
     }
 
-    if (dev) {
-        size_t free, total;
-        dev->iface.get_memory(dev, &free, &total);
-        GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024);
-    }
-
     if (backend_library_handle) {
-        GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__);
+        GGML_LOG_INFO(GGML_VIRTGPU_BCK "The GGML backend library was loaded. Unloading it.\n");
         dlclose(backend_library_handle);
         backend_library_handle = NULL;
     }
@@ -65,7 +59,7 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
         if (apir_logfile) {
             ggml_log_set(log_to_file_callback, apir_logfile);
         } else {
-            GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file);
+            GGML_LOG_INFO(GGML_VIRTGPU_BCK "Could not open the log file at '%s'\n", apir_log_to_file);
         }
     }
 
@@ -74,7 +68,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
     const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
 
     if (!library_name) {
-        GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
+                       "%s: cannot open the GGML library: env var '%s' not defined\n",
+                       __func__, APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
+
 
         return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
     }
@@ -82,13 +79,16 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
     backend_library_handle = dlopen(library_name, RTLD_LAZY);
 
     if (!backend_library_handle) {
-        GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror());
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
+                       "%s: cannot open the GGML library: %s\n", __func__, dlerror());
 
         return APIR_LOAD_LIBRARY_CANNOT_OPEN;
     }
 
     if (!library_reg) {
-        GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
+                       "%s: cannot register the GGML library: env var '%s' not defined\n",
+                       __func__, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
 
         return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
     }
@@ -96,8 +96,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
     void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
     dlsym_error                 = dlerror();
     if (dlsym_error) {
-        GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg,
-              APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
+                       "%s: cannot find the GGML backend registration symbol '%s' (from %s): %s\n",
+                       __func__, library_reg, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
+
 
         return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
     }
@@ -134,7 +136,9 @@ uint32_t apir_backend_dispatcher(uint32_t               virgl_ctx_id,
     };
 
     if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
-        GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
+        GGML_LOG_ERROR(GGML_VIRTGPU_BCK
+                       "%s: Received an invalid dispatch index (%d >= %d)\n",
+                        __func__, cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
         return APIR_BACKEND_FORWARD_INDEX_INVALID;
     }
 

ggml/src/ggml-virtgpu/backend/shared/apir_cs.h

@@ -86,7 +86,7 @@ static inline bool apir_decoder_peek_internal(apir_decoder * dec,
     assert(val_size <= size);
 
     if (unlikely(size > (size_t) (dec->end - dec->cur))) {
-        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
         apir_decoder_set_fatal(dec);
         memset(val, 0, val_size);
         return false;
@@ -103,7 +103,7 @@ static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val
 
 static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) {
     if (unlikely(size > (size_t) (dec->end - dec->cur))) {
-        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
         apir_decoder_set_fatal(dec);
         return NULL;
     }
@@ -221,7 +221,7 @@ static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expec
     uint64_t size;
     apir_decode_uint64_t(dec, &size);
     if (size != expected_size) {
-        GGML_LOG_ERROR("Couldn't decode array from the decoder\n");
+        GGML_LOG_ERROR("%s: Couldn't decode array from the decoder\n", __func__);
         apir_decoder_set_fatal(dec);
         size = 0;
     }
@@ -322,7 +322,7 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
     if (size) {
         val[size - 1] = '\0';
     } else {
-        GGML_LOG_ERROR("Couldn't decode the blog array\n");
+        GGML_LOG_ERROR("%s: Couldn't decode the blog array\n", __func__);
         apir_decoder_set_fatal(dec);
     }
 }
@@ -332,7 +332,8 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
 static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
     size_t alloc_size;
     if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
-        GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count);
+        GGML_LOG_ERROR("%s: overflow in array allocation of %zu * %zu bytes\n",
+                       __func__, size, count);
         return NULL;
     }
 

ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h

@@ -39,11 +39,17 @@ static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor
 
 static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
     const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
+
+    if (!apir_rpc_tensor) {
+        return NULL;
+    }
+
     ggml_init_params params{
         /*.mem_size   =*/ ggml_tensor_overhead(),
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
+
     ggml_context * ctx = ggml_init(params);
 
     const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
@@ -71,6 +77,10 @@ static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decod
     return (ggml_backend_buffer_type_t) handle;
 }
 
+static inline void apir_encode_apir_buffer_type_host_handle(apir_encoder * enc, apir_buffer_type_host_handle_t handle) {
+    apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
+}
+
 static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
     apir_buffer_type_host_handle_t handle;
 
@@ -154,13 +164,13 @@ static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml
     size_t tensor_size = sizeof(*tensor);
 
     if (tensor->extra) {
-        GGML_ABORT("Cannot pass tensors with extra");
+        GGML_ABORT("%s: Cannot pass tensors with extra", __func__);
     }
 
     if (tensor->src[0] && tensor->buffer) {
         static int first = 1;
         if (first) {
-            GGML_LOG_WARN("Cannot pass tensors with src and buffer\n");
+            GGML_LOG_WARN("%s: Cannot pass tensors with src and buffer\n", __func__);
             first = 0;
         }
     }

ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp

@@ -6,7 +6,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
 
     ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
     if (!context) {
-        GGML_ABORT("Couldn't allocate the buffer context ...");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
     }
 
     context->gpu = gpu;
@@ -20,7 +20,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
         context->base         = context->apir_context.shmem.mmap_ptr;
         context->is_from_ptr  = true;
     } else {
-        context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
+        context->apir_context = apir_buffer_type_alloc_buffer(gpu, gpu->cached_buffer_type.host_handle, size);
         context->is_from_ptr  = false;
         context->base         = NULL;
     }
@@ -34,36 +34,19 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
 static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     virtgpu * gpu = BUFT_TO_GPU(buft);
 
-    return apir_buffer_type_get_name(gpu, buft);
+    return gpu->cached_buffer_type.name;
 }
 
 static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     virtgpu * gpu = BUFT_TO_GPU(buft);
 
-    static size_t align = 0;
-
-    if (align == 0) {
-        align = apir_buffer_type_get_alignment(gpu, buft);
-    }
-
-    return align;
+    return gpu->cached_buffer_type.alignment;
 }
 
 static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
     virtgpu * gpu = BUFT_TO_GPU(buft);
 
-    static size_t max_size = 0;
-    if (max_size == 0) {
-        max_size = apir_buffer_type_get_max_size(gpu, buft);
-    }
-
-    return max_size;
-}
-
-static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    virtgpu * gpu = BUFT_TO_GPU(buft);
-
-    return apir_buffer_type_is_host(gpu, buft);
+    return gpu->cached_buffer_type.max_size;
 }
 
 static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
@@ -76,7 +59,7 @@ static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buff
         return ggml_nbytes(tensor);
     }
 
-    return apir_buffer_type_get_alloc_size(gpu, buft, tensor);
+    return apir_buffer_type_get_alloc_size(gpu, gpu->cached_buffer_type.host_handle, tensor);
 }
 
 const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {

ggml/src/ggml-virtgpu/ggml-backend-device.cpp

@@ -3,32 +3,27 @@
 static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    return apir_device_get_name(gpu);
+    return gpu->cached_device_info.name;
 }
 
 static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    return apir_device_get_description(gpu);
+    // Return the pre-cached description from the virtgpu structure
+    return gpu->cached_device_info.description;
 }
 
 static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    static enum ggml_backend_dev_type type;
-    static bool                       has_type = false;
-    if (!has_type) {
-        has_type = true;
-        type     = (enum ggml_backend_dev_type) apir_device_get_type(gpu);
-    }
-
-    return type;
+    return (enum ggml_backend_dev_type) gpu->cached_device_info.type;
 }
 
 static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    return apir_device_get_memory(gpu, free, total);
+    *free = gpu->cached_device_info.memory_free;
+    *total = gpu->cached_device_info.memory_total;
 }
 
 static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
@@ -77,13 +72,22 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
 ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
+    static std::atomic<bool> initialized = false;
+    static ggml_backend_buffer_type buft;
 
-    static ggml_backend_buffer_type buft{
-        /* .iface    = */ ggml_backend_remoting_buffer_type_interface,
-        /* .device   = */ dev,
-        /* .context  = */ (void *) ctx,
-    };
+    if (!initialized) {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (!initialized) {
+            buft = {
+                /* .iface    = */ ggml_backend_remoting_buffer_type_interface,
+                /* .device   = */ dev,
+                /* .context  = */ (void *) gpu->cached_buffer_type.host_handle,
+            };
+            initialized = true;
+        }
+    }
 
     return &buft;
 }
@@ -91,13 +95,22 @@ ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_bac
 static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
     virtgpu * gpu = DEV_TO_GPU(dev);
 
-    apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
+    static std::atomic<bool> initialized = false;
+    static ggml_backend_buffer_type buft;
 
-    static ggml_backend_buffer_type buft{
-        /* .iface    = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
-        /* .device   = */ dev,
-        /* .context  = */ (void *) ctx,
-    };
+    if (!initialized) {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (!initialized) {
+            buft = {
+                /* .iface    = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
+                /* .device   = */ dev,
+                /* .context  = */ (void *) gpu->cached_buffer_type.host_handle,
+            };
+            initialized = true;
+        }
+    }
 
     return &buft;
 }
@@ -110,7 +123,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_b
 
     ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
     if (!context) {
-        GGML_ABORT("Couldn't allocate the buffer context ...");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
     }
 
     context->gpu          = gpu;

ggml/src/ggml-virtgpu/ggml-backend-reg.cpp

@@ -4,37 +4,70 @@
 #include <iostream>
 #include <mutex>
 
+void ggml_virtgpu_cleanup(virtgpu * gpu);
+
 static virtgpu * apir_initialize() {
-    static virtgpu * apir_gpu_instance = NULL;
-    static bool      apir_initialized  = false;
+    static virtgpu *         gpu          = NULL;
+    static std::atomic<bool> initialized  = false;
+
+    if (initialized) {
+        // fast track
+        return gpu;
+    }
 
     {
         static std::mutex           mutex;
         std::lock_guard<std::mutex> lock(mutex);
 
-        if (apir_initialized) {
-            return apir_gpu_instance;
+        if (initialized) {
+            // thread safe
+            return gpu;
         }
 
-        apir_gpu_instance = create_virtgpu();
-        if (!apir_gpu_instance) {
-            GGML_ABORT("failed to initialize the virtgpu");
+        gpu = create_virtgpu();
+        if (!gpu) {
+            initialized = true;
+            return NULL;
         }
 
-        apir_initialized = true;
+        // Pre-fetch and cache all device information, it will not change
+        gpu->cached_device_info.description  = apir_device_get_description(gpu);
+        if (!gpu->cached_device_info.description) {
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device description", __func__);
+        }
+        gpu->cached_device_info.name         = apir_device_get_name(gpu);
+        if (!gpu->cached_device_info.name) {
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device name", __func__);
+        }
+        gpu->cached_device_info.device_count = apir_device_get_count(gpu);
+        gpu->cached_device_info.type         = apir_device_get_type(gpu);
+
+        apir_device_get_memory(gpu,
+                              &gpu->cached_device_info.memory_free,
+                              &gpu->cached_device_info.memory_total);
+
+        apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
+        gpu->cached_buffer_type.host_handle             = buft_host_handle;
+        gpu->cached_buffer_type.name                    = apir_buffer_type_get_name(gpu, buft_host_handle);
+        if (!gpu->cached_buffer_type.name) {
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu buffer type name", __func__);
+        }
+        gpu->cached_buffer_type.alignment               = apir_buffer_type_get_alignment(gpu, buft_host_handle);
+        gpu->cached_buffer_type.max_size                = apir_buffer_type_get_max_size(gpu, buft_host_handle);
+
+        initialized = true;
     }
 
-    return apir_gpu_instance;
+    return gpu;
 }
 
 static int ggml_backend_remoting_get_device_count() {
     virtgpu * gpu = apir_initialize();
     if (!gpu) {
-        GGML_LOG_WARN("apir_initialize failed\n");
         return 0;
     }
 
-    return apir_device_get_count(gpu);
+    return gpu->cached_device_info.device_count;
 }
 
 static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
@@ -52,17 +85,21 @@ ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) {
 
 static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
     if (devices.size() > 0) {
-        GGML_LOG_INFO("%s: already initialized\n", __func__);
+        GGML_LOG_INFO(GGML_VIRTGPU "%s: already initialized\n", __func__);
         return;
     }
 
     virtgpu * gpu = apir_initialize();
     if (!gpu) {
-        GGML_LOG_ERROR("apir_initialize failed\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: apir_initialize failed\n", __func__);
         return;
     }
 
-    static bool initialized = false;
+    static std::atomic<bool> initialized = false;
+
+    if (initialized) {
+        return; // fast track
+    }
 
     {
         static std::mutex           mutex;
@@ -70,10 +107,10 @@ static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
         if (!initialized) {
             for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) {
                 ggml_backend_remoting_device_context * ctx       = new ggml_backend_remoting_device_context;
-                char                                   desc[256] = "API Remoting device";
+                char                                   desc[256] = "ggml-virtgpu API Remoting device";
 
                 ctx->device      = i;
-                ctx->name        = GGML_REMOTING_FRONTEND_NAME + std::to_string(i);
+                ctx->name        = GGML_VIRTGPU_NAME + std::to_string(i);
                 ctx->description = desc;
                 ctx->gpu         = gpu;
 
@@ -98,7 +135,7 @@ static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_
 static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) {
     UNUSED(reg);
 
-    return GGML_REMOTING_FRONTEND_NAME;
+    return GGML_VIRTGPU_NAME;
 }
 
 static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
@@ -111,8 +148,7 @@ static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
 ggml_backend_reg_t ggml_backend_virtgpu_reg() {
     virtgpu * gpu = apir_initialize();
     if (!gpu) {
-        GGML_LOG_ERROR("virtgpu_apir_initialize failed\n");
-        return NULL;
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_apir_initialize failed\n", __func__);
     }
 
     static ggml_backend_reg reg = {
@@ -129,9 +165,25 @@ ggml_backend_reg_t ggml_backend_virtgpu_reg() {
 
     ggml_backend_remoting_reg_init_devices(&reg);
 
-    GGML_LOG_INFO("%s: initialized\n", __func__);
-
     return &reg;
 }
 
+// public function, not exposed in the GGML interface at the moment
+void ggml_virtgpu_cleanup(virtgpu * gpu) {
+    if (gpu->cached_device_info.name) {
+        free(gpu->cached_device_info.name);
+        gpu->cached_device_info.name = NULL;
+    }
+    if (gpu->cached_device_info.description) {
+        free(gpu->cached_device_info.description);
+        gpu->cached_device_info.description = NULL;
+    }
+    if (gpu->cached_buffer_type.name) {
+        free(gpu->cached_buffer_type.name);
+        gpu->cached_buffer_type.name = NULL;
+    }
+
+    mtx_destroy(&gpu->data_shmem_mutex);
+}
+
 GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)

ggml/src/ggml-virtgpu/ggml-remoting.h

@@ -8,6 +8,9 @@
 #include <memory>
 #include <string>
 
+#define GGML_VIRTGPU_NAME "ggml-virtgpu"
+#define GGML_VIRTGPU "ggml-virtgpu: "
+
 // USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes
 
 #define USE_ALWAYS_TRUE_SUPPORTS_OP 1
@@ -62,7 +65,7 @@ static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggm
 
 static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
     if (!buffer->context) {
-        GGML_ABORT("%s: no context available :/", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: no context available :/", __func__);
     }
     return BUFFER_TO_HOST_HANDLE(buffer);
 }

ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml

@@ -24,10 +24,10 @@ functions:
         frontend_return: "int"
 
       get_name:
-        frontend_return: "const char *"
+        frontend_return: "char *"
 
       get_description:
-        frontend_return: "const char *"
+        frontend_return: "char *"
 
       get_type:
         frontend_return: "uint32_t"
@@ -64,35 +64,33 @@ functions:
     group_description: "buffer-type"
     functions:
       get_name:
-        frontend_return: "const char *"
+        frontend_return: "char *"
         frontend_extra_params:
-        - "ggml_backend_buffer_type_t buft"
+        - "apir_buffer_type_host_handle_t host_handle"
 
       get_alignment:
         frontend_return: "size_t"
         frontend_extra_params:
-        - "ggml_backend_buffer_type_t buft"
+        - "apir_buffer_type_host_handle_t host_handle"
 
       get_max_size:
         frontend_return: "size_t"
         frontend_extra_params:
-        - "ggml_backend_buffer_type_t buft"
+        - "apir_buffer_type_host_handle_t host_handle"
 
       is_host:
-        frontend_return: "bool"
-        frontend_extra_params:
-        - "ggml_backend_buffer_type_t buft"
+        deprecated: true
 
       alloc_buffer:
         frontend_return: "apir_buffer_context_t"
         frontend_extra_params:
-        - "ggml_backend_buffer_type_t buffer_buft"
+        - "apir_buffer_type_host_handle_t host_handle"
         - "size_t size"
 
       get_alloc_size:
         frontend_return: "size_t"
         frontend_extra_params:
-        - "ggml_backend_buffer_type_t buft"
+        - "apir_buffer_type_host_handle_t host_handle"
         - "const ggml_tensor *op"
 
   buffer:

ggml/src/ggml-virtgpu/regenerate_remoting.py

@@ -116,7 +116,7 @@ class RemotingCodebaseGenerator:
                         'frontend_return': func_metadata.get('frontend_return', 'void'),
                         'frontend_extra_params': func_metadata.get('frontend_extra_params', []),
                         'group_description': group_description,
-                        'newly_added': func_metadata.get('newly_added', False)
+                        'deprecated': func_metadata.get('deprecated', False),
                     })
                     enum_value += 1
 
@@ -165,6 +165,9 @@ class RemotingCodebaseGenerator:
 
             signature = "uint32_t"
             params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx"
+            if func['deprecated']:
+                decl_lines.append(f"/* {func['enum_name']} is deprecated. Keeping the handler for backward compatibility. */")
+
             decl_lines.append(f"{signature} {func['backend_function']}({params});")
 
         # Switch cases
@@ -176,7 +179,9 @@ class RemotingCodebaseGenerator:
                 switch_lines.append(f"  /* {func['group_description']} */")
                 current_group = func['group_name']
 
-            switch_lines.append(f"  case {func['enum_name']}: return \"{func['backend_function']}\";")
+            deprecated = " (DEPRECATED)" if func['deprecated'] else ""
+
+            switch_lines.append(f"  case {func['enum_name']}: return \"{func['backend_function']}{deprecated}\";")
 
         # Dispatch table
         table_lines = []
@@ -188,7 +193,8 @@ class RemotingCodebaseGenerator:
                 table_lines.append("")
                 current_group = func['group_name']
 
-            table_lines.append(f"  /* {func['enum_name']}  = */ {func['backend_function']},")
+            deprecated = " /* DEPRECATED */" if func['deprecated'] else ""
+            table_lines.append(f"  /* {func['enum_name']}  = */ {func['backend_function']}{deprecated},")
 
         header_content = f'''\
 #pragma once
@@ -225,6 +231,10 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
                 decl_lines.append(f"/* {func['group_description']} */")
                 current_group = func['group_name']
 
+            if func['deprecated']:
+                decl_lines.append(f"/* {func['frontend_function']} is deprecated. */")
+                continue
+
             # Build parameter list
             params = [self.naming_patterns['frontend_base_param']]
             params.extend(func['frontend_extra_params'])
@@ -287,7 +297,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
         generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path]
 
         if not self.clang_format_available:
-            logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted."
+            logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted.\n"
                             "   Install clang-format to enable automatic code formatting.")
         else:
             logging.info("\n🎨 Formatting files with clang-format...")

ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp

@@ -18,12 +18,17 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
     virtgpu_shmem * shmem = &temp_shmem;
+    bool using_shared_shmem = false;
 
     if (cgraph_size <= gpu->data_shmem.mmap_size) {
-        // prefer the init-time allocated page, if large enough
+        // Lock mutex before using shared data_shmem buffer
+        if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
+            GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
+        }
+        using_shared_shmem = true;
         shmem = &gpu->data_shmem;
     } else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
-        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
     }
 
     apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
@@ -42,7 +47,10 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
 
     remote_call_finish(gpu, encoder, decoder);
 
-    if (shmem != &gpu->data_shmem) {
+    // Unlock mutex before cleanup
+    if (using_shared_shmem) {
+        mtx_unlock(&gpu->data_shmem_mutex);
+    } else {
         virtgpu_shmem_destroy(gpu, shmem);
     }
 

ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp

@@ -1,20 +1,20 @@
 #include "virtgpu-forward-impl.h"
 
-const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+char * apir_buffer_type_get_name(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
 
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME);
 
-    apir_encode_ggml_buffer_type(encoder, buft);
+    apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
 
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
     const size_t string_size = apir_decode_array_size_unchecked(decoder);
     char *       string      = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
     if (!string) {
-        GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
         apir_decoder_set_fatal(decoder);
     }
     apir_decode_char_array(decoder, string, string_size);
@@ -24,14 +24,14 @@ const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t
     return string;
 }
 
-size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+size_t apir_buffer_type_get_alignment(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
 
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT);
 
-    apir_encode_ggml_buffer_type(encoder, buft);
+    apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
 
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
@@ -43,14 +43,14 @@ size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t
     return alignment;
 }
 
-size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+size_t apir_buffer_type_get_max_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
 
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE);
 
-    apir_encode_ggml_buffer_type(encoder, buft);
+    apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
 
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
@@ -62,26 +62,7 @@ size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t b
     return max_size;
 }
 
-bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
-    apir_encoder *        encoder;
-    apir_decoder *        decoder;
-    ApirForwardReturnCode ret;
-
-    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST);
-
-    apir_encode_ggml_buffer_type(encoder, buft);
-
-    REMOTE_CALL(gpu, encoder, decoder, ret);
-
-    bool is_host;
-    apir_decode_bool_t(decoder, &is_host);
-
-    remote_call_finish(gpu, encoder, decoder);
-
-    return is_host;
-}
-
-apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) {
+apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, size_t size) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -90,7 +71,7 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
 
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER);
 
-    apir_encode_ggml_buffer_type(encoder, buft);
+    apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
 
     apir_encode_size_t(encoder, &size);
 
@@ -103,14 +84,14 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
     return buffer_context;
 }
 
-size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) {
+size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, const ggml_tensor * op) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
 
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE);
 
-    apir_encode_ggml_buffer_type(encoder, buft);
+    apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
 
     apir_encode_ggml_tensor_inline(encoder, op);
 

ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp

@@ -36,13 +36,18 @@ void apir_buffer_set_tensor(virtgpu *               gpu,
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
     virtgpu_shmem * shmem = &temp_shmem;
+    bool using_shared_shmem = false;
 
     if (size <= gpu->data_shmem.mmap_size) {
-        // prefer the init-time allocated page, if large enough
+        // Lock mutex before using shared data_shmem buffer
+        if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
+            GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
+        }
+        using_shared_shmem = true;
         shmem = &gpu->data_shmem;
 
     } else if (virtgpu_shmem_create(gpu, size, shmem)) {
-        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
     }
 
     memcpy(shmem->mmap_ptr, data, size);
@@ -55,7 +60,10 @@ void apir_buffer_set_tensor(virtgpu *               gpu,
 
     remote_call_finish(gpu, encoder, decoder);
 
-    if (shmem != &gpu->data_shmem) {
+    // Unlock mutex before cleanup
+    if (using_shared_shmem) {
+        mtx_unlock(&gpu->data_shmem_mutex);
+    } else {
         virtgpu_shmem_destroy(gpu, shmem);
     }
 
@@ -79,13 +87,18 @@ void apir_buffer_get_tensor(virtgpu *               gpu,
 
     virtgpu_shmem   temp_shmem;  // Local storage for large buffers
     virtgpu_shmem * shmem = &temp_shmem;
+    bool using_shared_shmem = false;
 
     if (size <= gpu->data_shmem.mmap_size) {
-        // prefer the init-time allocated page, if large enough
+        // Lock mutex before using shared data_shmem buffer
+        if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
+            GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
+        }
+        using_shared_shmem = true;
         shmem = &gpu->data_shmem;
 
     } else if (virtgpu_shmem_create(gpu, size, shmem)) {
-        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+        GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
     }
 
     apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
@@ -98,7 +111,10 @@ void apir_buffer_get_tensor(virtgpu *               gpu,
 
     remote_call_finish(gpu, encoder, decoder);
 
-    if (shmem != &gpu->data_shmem) {
+    // Unlock mutex before cleanup
+    if (using_shared_shmem) {
+        mtx_unlock(&gpu->data_shmem_mutex);
+    } else {
         virtgpu_shmem_destroy(gpu, shmem);
     }
 }

ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp

@@ -2,11 +2,6 @@
 #include "virtgpu-shm.h"
 
 int apir_device_get_count(virtgpu * gpu) {
-    static int32_t dev_count = -1;
-    if (dev_count != -1) {
-        return dev_count;
-    }
-
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -14,6 +9,7 @@ int apir_device_get_count(virtgpu * gpu) {
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT);
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
+    int32_t dev_count = -1;
     apir_decode_int32_t(decoder, &dev_count);
 
     remote_call_finish(gpu, encoder, decoder);
@@ -21,11 +17,7 @@ int apir_device_get_count(virtgpu * gpu) {
     return dev_count;
 }
 
-const char * apir_device_get_name(virtgpu * gpu) {
-    static char * string = nullptr;
-    if (string) {
-        return string;
-    }
+char * apir_device_get_name(virtgpu * gpu) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -34,9 +26,9 @@ const char * apir_device_get_name(virtgpu * gpu) {
     REMOTE_CALL(gpu, encoder, decoder, ret);
 
     const size_t string_size = apir_decode_array_size_unchecked(decoder);
-    string                   = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
+    char            * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
     if (!string) {
-        GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
         return NULL;
     }
     apir_decode_char_array(decoder, string, string_size);
@@ -46,7 +38,7 @@ const char * apir_device_get_name(virtgpu * gpu) {
     return string;
 }
 
-const char * apir_device_get_description(virtgpu * gpu) {
+char * apir_device_get_description(virtgpu * gpu) {
     apir_encoder *        encoder;
     apir_decoder *        decoder;
     ApirForwardReturnCode ret;
@@ -58,7 +50,7 @@ const char * apir_device_get_description(virtgpu * gpu) {
     const size_t string_size = apir_decode_array_size_unchecked(decoder);
     char *       string      = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
     if (!string) {
-        GGML_LOG_ERROR("%s: Could not allocate the device description buffer\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device description buffer\n", __func__);
 
         return NULL;
     }
@@ -181,7 +173,7 @@ apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, si
     REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR);
 
     if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) {
-        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+        GGML_ABORT(GGML_VIRTGPU "Couldn't allocate the guest-host shared buffer");
     }
 
     apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id);

ggml/src/ggml-virtgpu/virtgpu-forward-impl.h

@@ -11,7 +11,7 @@
         int32_t forward_flag = (int32_t) apir_command_type__;                                              \
         encoder_name         = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \
         if (!encoder_name) {                                                                               \
-            GGML_ABORT("%s: failed to prepare the remote call encoder", __func__);                       \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__);                       \
         }                                                                                                  \
     } while (0)
 
@@ -19,10 +19,10 @@
     do {                                                                                                          \
         ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL);       \
         if (!decoder_name) {                                                                                      \
-            GGML_ABORT("%s: failed to kick the remote call", __func__);                                         \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to kick the remote call", __func__);                                         \
         }                                                                                                         \
         if (ret_name < APIR_FORWARD_BASE_INDEX) {                                                                 \
-            GGML_ABORT("%s: failed to forward the API call: %s: code %d", __func__,                             \
+            GGML_ABORT(GGML_VIRTGPU "%s: failed to forward the API call: %s: code %d", __func__,                             \
                        apir_forward_error(ret_name), ret_name);                                                   \
         }                                                                                                         \
         ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX);                                  \

ggml/src/ggml-virtgpu/virtgpu-forward.gen.h

@@ -3,8 +3,8 @@
 /* device */
 void                           apir_device_get_device_count(struct virtgpu * gpu);
 int                            apir_device_get_count(struct virtgpu * gpu);
-const char *                   apir_device_get_name(struct virtgpu * gpu);
-const char *                   apir_device_get_description(struct virtgpu * gpu);
+char *                         apir_device_get_name(struct virtgpu * gpu);
+char *                         apir_device_get_description(struct virtgpu * gpu);
 uint32_t                       apir_device_get_type(struct virtgpu * gpu);
 void                           apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total);
 bool                           apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op);
@@ -17,14 +17,15 @@ void                           apir_device_get_props(struct virtgpu * gpu,
 apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
 
 /* buffer-type */
-const char *          apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
-size_t                apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
-size_t                apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
-bool                  apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
-apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu *           gpu,
-                                                    ggml_backend_buffer_type_t buffer_buft,
-                                                    size_t                     size);
-size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op);
+char *                apir_buffer_type_get_name(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
+size_t                apir_buffer_type_get_alignment(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
+size_t                apir_buffer_type_get_max_size(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
+apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu *               gpu,
+                                                    apir_buffer_type_host_handle_t host_handle,
+                                                    size_t                         size);
+size_t                apir_buffer_type_get_alloc_size(struct virtgpu *               gpu,
+                                                      apir_buffer_type_host_handle_t host_handle,
+                                                      const ggml_tensor *            op);
 
 /* buffer */
 void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);

ggml/src/ggml-virtgpu/virtgpu-shm.cpp

@@ -85,8 +85,7 @@ int virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem) {
     void * ptr = virtgpu_ioctl_map(gpu, gem_handle, size);
     if (!ptr) {
         virtgpu_ioctl_gem_close(gpu, gem_handle);
-        GGML_LOG_ERROR("virtgpu_ioctl_map FAILED\n");
-        exit(1);
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_ioctl_map failed\n", __func__);
         return 1;
     }
 

ggml/src/ggml-virtgpu/virtgpu.cpp

@@ -33,7 +33,7 @@ static int virtgpu_handshake(virtgpu * gpu) {
 
     encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_HANDSHAKE, 0);
     if (!encoder) {
-        GGML_ABORT("%s: failed to prepare the remote call encoder", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__);
         return 1;
     }
 
@@ -52,7 +52,7 @@ static int virtgpu_handshake(virtgpu * gpu) {
     log_call_duration(call_duration_ns, "API Remoting handshake");
 
     if (!decoder) {
-        GGML_ABORT(
+        GGML_ABORT(GGML_VIRTGPU
             "%s: failed to initiate the communication with the virglrenderer library. "
             "Most likely, the wrong virglrenderer library was loaded in the hypervisor.",
             __func__);
@@ -65,7 +65,8 @@ static int virtgpu_handshake(virtgpu * gpu) {
     uint32_t host_minor;
 
     if (ret_magic != APIR_HANDSHAKE_MAGIC) {
-        GGML_ABORT("%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
                    apir_backend_initialize_error(ret_magic));
     } else {
         apir_decode_uint32_t(decoder, &host_major);
@@ -78,13 +79,13 @@ static int virtgpu_handshake(virtgpu * gpu) {
         return 1;
     }
 
-    GGML_LOG_INFO("%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor);
-    GGML_LOG_INFO("%s: Host is running with %u.%u\n", __func__, host_major, host_minor);
+    GGML_LOG_INFO(GGML_VIRTGPU "%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor);
+    GGML_LOG_INFO(GGML_VIRTGPU "%s: Host is running with %u.%u\n", __func__, host_major, host_minor);
 
     if (guest_major != host_major) {
-        GGML_LOG_ERROR("Host major (%d) and guest major (%d) version differ\n", host_major, guest_major);
+        GGML_LOG_ERROR(GGML_VIRTGPU "Host major (%d) and guest major (%d) version differ\n", host_major, guest_major);
     } else if (guest_minor != host_minor) {
-        GGML_LOG_WARN("Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor);
+        GGML_LOG_WARN(GGML_VIRTGPU "Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor);
     }
 
     return 0;
@@ -97,7 +98,7 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
 
     encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_LOADLIBRARY, 0);
     if (!encoder) {
-        GGML_ABORT("%s: hypercall error: failed to prepare the remote call encoder", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: hypercall error: failed to prepare the API Remoting command encoder", __func__);
         return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
     }
 
@@ -108,36 +109,67 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
     log_call_duration(call_duration_ns, "API Remoting LoadLibrary");
 
     if (!decoder) {
-        GGML_ABORT("%s: hypercall error: failed to kick the API remoting hypercall.\n", __func__);
+        GGML_ABORT(GGML_VIRTGPU "%s: hypercall error: failed to trigger the API Remoting hypercall.\n", __func__);
         return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
     }
 
     remote_call_finish(gpu, encoder, decoder);
 
     if (ret == APIR_LOAD_LIBRARY_SUCCESS) {
-        GGML_LOG_INFO("%s: The API Remoting backend was successfully loaded and initialized\n", __func__);
+        GGML_LOG_INFO(GGML_VIRTGPU "The API Remoting backend was successfully loaded and initialized\n");
 
         return ret;
     }
 
     // something wrong happened, find out what.
-
     if (ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
-        GGML_ABORT("%s: virglrenderer could not load the API Remoting backend library: %s (code %d)", __func__,
-                   apir_load_library_error(ret), ret);
+        if (ret == APIR_LOAD_LIBRARY_ENV_VAR_MISSING) {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: virglrenderer could not open the API Remoting backend library, "
+                       "some environment variables are missing. "
+                       "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
+                       __func__, apir_load_library_error(ret));
+        } else if (ret == APIR_LOAD_LIBRARY_CANNOT_OPEN) {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: virglrenderer could not open the API Remoting backend library. "
+                       "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
+                       __func__, apir_load_library_error(ret));
+        } else if (ret == APIR_LOAD_LIBRARY_ENV_VAR_MISSING) {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: could not load the backend library, some symbols are missing. "
+                       "Make sure virglrenderer is correctly configured by the hypervisor. (%s) ",
+                       __func__, apir_load_library_error(ret));
+        } else {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: virglrenderer could not load the API Remoting backend library. (%s - code %d)", __func__,
+                       apir_load_library_error(ret), ret);
+        }
         return ret;
     }
 
-    GGML_LOG_INFO("%s: virglrenderer successfully loaded the API Remoting backend library", __func__);
+    GGML_LOG_INFO(GGML_VIRTGPU
+                  "%s: virglrenderer successfully loaded the API Remoting backend library.\n", __func__);
 
     ApirLoadLibraryReturnCode apir_ret = (ApirLoadLibraryReturnCode) (ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
 
-    if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
-        GGML_ABORT("%s: the API Remoting backend library couldn't load the backend library: apir code=%d | %s)",
+    if (apir_ret == APIR_LOAD_LIBRARY_CANNOT_OPEN) {
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: the API Remoting backend library couldn't load the GGML backend library. "
+                   "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
+                   __func__, apir_load_library_error(apir_ret));
+    } else if (apir_ret == APIR_LOAD_LIBRARY_SYMBOL_MISSING) {
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: the API Remoting backend library couldn't load the GGML backend library, some symbols are missing. "
+                   "Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
+                   __func__, apir_load_library_error(apir_ret));
+    } else if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: the API Remoting backend library couldn't load the GGML backend library: apir code=%d | %s)",
                    __func__, apir_ret, apir_load_library_error(apir_ret));
     } else {
         uint32_t lib_ret = apir_ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX;
-        GGML_ABORT("%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
                    lib_ret);
     }
     return ret;
@@ -149,38 +181,58 @@ virtgpu * create_virtgpu() {
     gpu->use_apir_capset = getenv("GGML_REMOTING_USE_APIR_CAPSET") != nullptr;
     util_sparse_array_init(&gpu->shmem_array, sizeof(virtgpu_shmem), 1024);
 
+    // Initialize mutex to protect shared data_shmem buffer
+    if (mtx_init(&gpu->data_shmem_mutex, mtx_plain) != thrd_success) {
+        delete gpu;
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to initialize data_shmem mutex", __func__);
+        return NULL;
+    }
+
     if (virtgpu_open(gpu) != APIR_SUCCESS) {
-        GGML_ABORT("%s: failed to open the virtgpu device", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU
+                       "%s: failed to open the virtgpu device\n", __func__);
         return NULL;
     }
 
     if (virtgpu_init_capset(gpu) != APIR_SUCCESS) {
-        GGML_ABORT("%s: failed to initialize the GPU capset", __func__);
+        if (gpu->use_apir_capset) {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: failed to initialize the virtgpu APIR capset. Make sure that the virglrenderer library supports it.", __func__);
+        } else {
+            GGML_ABORT(GGML_VIRTGPU
+                       "%s: failed to initialize the virtgpu Venus capset", __func__);
+        }
         return NULL;
     }
 
     if (virtgpu_init_context(gpu) != APIR_SUCCESS) {
-        GGML_ABORT("%s: failed to initialize the GPU context", __func__);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to initialize the GPU context", __func__);
         return NULL;
     }
 
     if (virtgpu_shmem_create(gpu, SHMEM_REPLY_SIZE, &gpu->reply_shmem)) {
-        GGML_ABORT("%s: failed to create the shared reply memory pages", __func__);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to create the shared reply memory pages", __func__);
         return NULL;
     }
 
     if (virtgpu_shmem_create(gpu, SHMEM_DATA_SIZE, &gpu->data_shmem)) {
-        GGML_ABORT("%s: failed to create the shared data memory pages", __func__);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to create the shared data memory pages", __func__);
         return NULL;
     }
 
     if (virtgpu_handshake(gpu)) {
-        GGML_ABORT("%s: failed to handshake with the virglrenderer library", __func__);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to handshake with the virglrenderer library", __func__);
         return NULL;
     }
 
     if (virtgpu_load_library(gpu) != APIR_LOAD_LIBRARY_SUCCESS) {
-        GGML_ABORT("%s: failed to load the backend library", __func__);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to load the backend library", __func__);
         return NULL;
     }
 
@@ -191,7 +243,8 @@ static virt_gpu_result_t virtgpu_open(virtgpu * gpu) {
     drmDevicePtr devs[8];
     int          count = drmGetDevices2(0, devs, ARRAY_SIZE(devs));
     if (count < 0) {
-        GGML_LOG_ERROR("%s: failed to enumerate DRM devices\n", __func__);
+        GGML_LOG_ERROR(GGML_VIRTGPU
+                       "%s: failed to enumerate DRM devices\n", __func__);
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
@@ -213,16 +266,19 @@ static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr d
 
     int fd = open(node_path, O_RDWR | O_CLOEXEC);
     if (fd < 0) {
-        GGML_ABORT("failed to open %s", node_path);
+        GGML_ABORT(GGML_VIRTGPU
+                   "%s: failed to open %s", __func__, node_path);
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
     drmVersionPtr version = drmGetVersion(fd);
     if (!version || strcmp(version->name, "virtio_gpu") || version->version_major != 0) {
         if (version) {
-            GGML_ABORT("unknown DRM driver %s version %d", version->name, version->version_major);
+            GGML_LOG_ERROR(GGML_VIRTGPU
+                           "%s: unknown DRM driver %s version %d\n", __func__, version->name, version->version_major);
         } else {
-            GGML_ABORT("failed to get DRM driver version");
+            GGML_LOG_ERROR(GGML_VIRTGPU
+                           "%s: failed to get DRM driver version\n", __func__);
         }
 
         if (version) {
@@ -236,7 +292,7 @@ static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr d
 
     drmFreeVersion(version);
 
-    GGML_LOG_INFO("using DRM device %s\n", node_path);
+    GGML_LOG_INFO(GGML_VIRTGPU "using DRM device %s\n", node_path);
 
     return APIR_SUCCESS;
 }
@@ -245,7 +301,7 @@ static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) {
     assert(!gpu->capset.version);
     const int ret = virtgpu_ioctl_context_init(gpu, gpu->capset.id);
     if (ret) {
-        GGML_LOG_INFO("failed to initialize context: %s\n", strerror(errno));
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to initialize context: %s\n", __func__, strerror(errno));
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
@@ -254,10 +310,10 @@ static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) {
 
 static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
     if (gpu->use_apir_capset) {
-        GGML_LOG_INFO("Using the APIR capset\n");
+        GGML_LOG_INFO(GGML_VIRTGPU "Using the APIR capset\n");
         gpu->capset.id = VIRTGPU_DRM_CAPSET_APIR;
     } else {
-        GGML_LOG_INFO("Using the Venus capset\n");
+        GGML_LOG_INFO(GGML_VIRTGPU "Using the Venus capset\n");
         gpu->capset.id = VIRTGPU_DRM_CAPSET_VENUS;
     }
     gpu->capset.version = 0;
@@ -266,7 +322,9 @@ static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
         virtgpu_ioctl_get_caps(gpu, gpu->capset.id, gpu->capset.version, &gpu->capset.data, sizeof(gpu->capset.data));
 
     if (ret) {
-        GGML_LOG_INFO("failed to get APIR v%d capset: %s\n", gpu->capset.version, strerror(errno));
+        GGML_LOG_ERROR(GGML_VIRTGPU
+                       "%s: failed to get APIR v%d capset: %s\n",
+                       __func__, gpu->capset.version, strerror(errno));
         return APIR_ERROR_INITIALIZATION_FAILED;
     }
 
@@ -333,9 +391,9 @@ apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type,
      * Prepare the command encoder and its buffer
      */
 
-    static char encoder_buffer[4096];
+    thread_local char encoder_buffer[4096];
 
-    static apir_encoder enc;
+    thread_local apir_encoder enc;
     enc = {
         .cur   = encoder_buffer,
         .start = encoder_buffer,
@@ -369,19 +427,19 @@ void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec) {
     UNUSED(gpu);
 
     if (!enc) {
-        GGML_LOG_ERROR("Invalid (null) encoder\n");
+        GGML_ABORT(GGML_VIRTGPU "%s: Invalid (null) encoder", __func__);
     }
 
     if (!dec) {
-        GGML_LOG_ERROR("Invalid (null) decoder\n");
+        GGML_ABORT(GGML_VIRTGPU "%s: Invalid (null) decoder", __func__);
     }
 
     if (apir_encoder_get_fatal(enc)) {
-        GGML_LOG_ERROR("Failed to encode the output parameters.\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: Failed to encode the output parameters.", __func__);
     }
 
     if (apir_decoder_get_fatal(dec)) {
-        GGML_LOG_ERROR("Failed to decode the input parameters.\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: Failed to decode the input parameters.", __func__);
     }
 }
 
@@ -423,7 +481,7 @@ uint32_t remote_call(virtgpu *       gpu,
     int ret = drmIoctl(gpu->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &args);
 
     if (ret != 0) {
-        GGML_ABORT("%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret);
+        GGML_ABORT(GGML_VIRTGPU "%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret);
     }
 
     /*
@@ -467,7 +525,7 @@ uint32_t remote_call(virtgpu *       gpu,
     }
 
     if (max_wait_ms && timedout) {
-        GGML_LOG_ERROR("timed out waiting for the host answer...\n");
+        GGML_LOG_ERROR(GGML_VIRTGPU "%s: timed out waiting for the host answer...\n", __func__);
         return APIR_FORWARD_TIMEOUT;
     }
 
@@ -489,10 +547,13 @@ static void log_call_duration(long long call_duration_ns, const char * name) {
     double call_duration_s  = (double) call_duration_ns / 1e9;  // 1 second = 1e9 nanoseconds
 
     if (call_duration_s > 1) {
-        GGML_LOG_INFO("%s: waited %.2fs for the %s host reply...\n", __func__, call_duration_s, name);
+        GGML_LOG_INFO(GGML_VIRTGPU
+                      "waited %.2fs for the %s host reply...\n", call_duration_s, name);
     } else if (call_duration_ms > 1) {
-        GGML_LOG_INFO("%s: waited %.2fms for the %s host reply...\n", __func__, call_duration_ms, name);
+        GGML_LOG_INFO(GGML_VIRTGPU
+                      "waited %.2fms for the %s host reply...\n", call_duration_ms, name);
     } else {
-        GGML_LOG_INFO("%s: waited %lldns for the %s host reply...\n", __func__, call_duration_ns, name);
+        GGML_LOG_INFO(GGML_VIRTGPU
+                      "waited %lldns for the %s host reply...\n", call_duration_ns, name);
     }
 }

ggml/src/ggml-virtgpu/virtgpu.h

@@ -17,6 +17,8 @@
 
 #include <cstring>
 
+#include "ggml-remoting.h"
+
 #define VIRGL_RENDERER_UNSTABLE_APIS 1
 #include "apir_hw.h"
 #include <drm/virtgpu_drm.h>
@@ -73,6 +75,27 @@ struct virtgpu {
     /* APIR communication pages */
     virtgpu_shmem reply_shmem;
     virtgpu_shmem data_shmem;
+
+    /* Mutex to protect shared data_shmem buffer from concurrent access */
+    mtx_t data_shmem_mutex;
+
+    /* Cached device information to prevent memory leaks and race conditions */
+    struct {
+        char *   description;
+        char *   name;
+        int32_t  device_count;
+        uint32_t type;
+        size_t   memory_free;
+        size_t   memory_total;
+    } cached_device_info;
+
+    /* Cached buffer type information to prevent memory leaks and race conditions */
+    struct {
+        apir_buffer_type_host_handle_t host_handle;
+        char *                         name;
+        size_t                         alignment;
+        size_t                         max_size;
+    } cached_buffer_type;
 };
 
 static inline int virtgpu_ioctl(virtgpu * gpu, unsigned long request, void * args) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -254,6 +254,7 @@ enum vk_device_architecture {
     AMD_RDNA3,
     INTEL_XE2,
     NVIDIA_PRE_TURING,
+    NVIDIA_TURING,
 };
 
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
@@ -336,18 +337,34 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
         const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
 
         bool cooperative_matrix = false;
+        bool sm_builtins = false;
 
         // Detect "pre-turing" based on lack of coopmat support.
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0) {
                 cooperative_matrix = true;
-                break;
+            } else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) {
+                sm_builtins = true;
             }
         }
 
         if (!cooperative_matrix) {
             return vk_device_architecture::NVIDIA_PRE_TURING;
         }
+
+        if (sm_builtins) {
+            vk::PhysicalDeviceProperties2 props2;
+            vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props;
+
+            props2.pNext = &sm_props;
+
+            device.getProperties2(&props2);
+
+            // Turing has 32, following architectures have 48
+            if (sm_props.shaderWarpsPerSM == 32) {
+                return vk_device_architecture::NVIDIA_TURING;
+            }
+        }
     }
     return vk_device_architecture::OTHER;
 }
@@ -1246,25 +1263,30 @@ struct vk_op_diag_mask_push_constants {
 
 struct vk_op_rope_push_constants {
     uint32_t rope_mode;
-    uint32_t ncols;
     uint32_t nrows;
     uint32_t n_dims;
     float freq_scale;
-    uint32_t p_delta_rows;
     float freq_base;
     float ext_factor;
     float attn_factor;
     float corr_dims[2];
     float theta_scale;
     uint32_t has_ff;
-    uint32_t ne02;
-    uint32_t s1;
-    uint32_t s2;
     int32_t sections[4];
     uint32_t is_imrope;
     uint32_t is_back;
     uint32_t set_rows_stride;
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t ne02;
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
 };
+static_assert(sizeof(vk_op_rope_push_constants) <= 128, "sizeof(vk_op_rope_push_constants) must be <= 128");
 
 // For fused rms_norm+mul+rope(+view+set_rows)
 struct vk_op_rms_norm_mul_rope_push_constants {
@@ -3182,9 +3204,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         const uint32_t D_lsb = D ^ (D & (D-1));
         uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
 
-        // Nvidia prefers shared memory use to load large tiles of K
+        // Nvidia prefers shared memory use to load large tiles of K.
+        // Switch to loading from global memory when it would use too much shared memory.
         // AMD prefers loading K directly from global memory
-        const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 1 : 0;
+        const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256 ? 1 : 0;
 
         return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem};
     };
@@ -5538,9 +5561,9 @@ static void ggml_vk_instance_init() {
                 // Check if there are two physical devices corresponding to the same GPU
                 // This handles the case where the same GPU appears with different drivers (e.g., RADV + AMDVLK on Linux),
                 // see https://github.com/ggml-org/llama.cpp/pull/7582 for original deduplication.
-                // However, for MoltenVK on macOS, multiple GPUs on the same card may report the same UUID,
-                // see https://github.com/KhronosGroup/MoltenVK/issues/2683. Until this is fixed, we'll only deduplicate
-                // when drivers differ (same driver + same UUID = likely different GPUs)
+                // MoltenVK on macOS may report the same UUID for distinct GPUs on multi-GPU cards,
+                // see https://github.com/KhronosGroup/MoltenVK/issues/2683. Skip when both old/new
+                // driver is MoltenVK
                 auto old_device = std::find_if(
                     vk_instance.device_indices.begin(),
                     vk_instance.device_indices.end(),
@@ -5557,11 +5580,9 @@ static void ggml_vk_instance_init() {
                             old_id.deviceLUIDValid && new_id.deviceLUIDValid &&
                             std::equal(std::begin(old_id.deviceLUID), std::end(old_id.deviceLUID), std::begin(new_id.deviceLUID))
                         );
+                        bool both_molten_vk = (new_driver.driverID == vk::DriverId::eMoltenvk && old_driver.driverID == vk::DriverId::eMoltenvk);
 
-                        // Only deduplicate if same UUID AND different drivers
-                        // (same driver + same UUID on MoltenVK = likely different GPUs on multi-GPU card)
-                        bool different_driver = (old_driver.driverID != new_driver.driverID);
-                        return same_uuid && different_driver;
+                        return same_uuid && !both_molten_vk;
                     }
                 );
                 if (old_device == vk_instance.device_indices.end()) {
@@ -8390,7 +8411,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA;
+    const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256;
     const uint32_t kshstride = (k_load_shmem ? hsk_pad : MatBr) / 4 + 2;
     const uint32_t vsh_stride = MatBc / 4 * row_split;
     const uint32_t ksh = ((kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)) * f16vec4;
@@ -8460,6 +8481,11 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     FaCodePath path = ctx->device->coopmat2 ? FA_COOPMAT2 :
                       ctx->device->coopmat1_fa_support ? FA_COOPMAT1 : FA_SCALAR;
 
+    if (path == FA_COOPMAT1 && ctx->device->architecture == vk_device_architecture::NVIDIA_TURING) {
+        // Nvidia compiler bug, see https://github.com/ggml-org/llama.cpp/pull/19075#issuecomment-3820716090
+        path = FA_SCALAR;
+    }
+
     if (path == FA_COOPMAT1) {
         const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
                                              (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
@@ -10383,12 +10409,22 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
 
     uint32_t nb01 = src0->nb[1] / ggml_type_size(src0->type);
     uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type);
+    uint32_t nb03 = src0->nb[3] / ggml_type_size(src0->type);
+
+    uint32_t nb11 = dst->nb[1] / ggml_type_size(dst->type);
+    uint32_t nb12 = dst->nb[2] / ggml_type_size(dst->type);
+    uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);
 
     vk_op_rope_push_constants rope {
-        (uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
-        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
-        has_ff, (uint32_t)src0->ne[2], nb01, nb02,
+        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
+        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
         { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
+
+        (uint32_t)src0->ne[0],
+        (uint32_t)src0->ne[1],
+        (uint32_t)src0->ne[2],
+        nb01, nb02, nb03,
+        nb11, nb12, nb13,
     };
 
     return rope;
@@ -14776,6 +14812,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_REPEAT_BACK:
             return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ROPE:
+            return ggml_is_contiguous_rows(op) && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ROPE_BACK:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:

ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp

@@ -112,12 +112,11 @@ void rms_norm(uint num_iters) {
 #if RMS_NORM_ROPE_FUSION
     barrier();
     rope_params rp = p.rope;
-    uint rope_row = (samp*nchannels + channel)*nrows + row;
     for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
         if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
-            rope_neox(t, rope_row, rp);
+            rope_neox(t, row, channel, samp, rp);
         } else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
-            rope_norm(t, rope_row, rp);
+            rope_norm(t, row, channel, samp, rp);
         }
     }
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl

@@ -4,12 +4,12 @@ float rope_yarn_ramp(const float low, const float high, const uint i0) {
     return 1.0f - min(1.0f, max(0.0f, y));
 }
 
-uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
+uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03, rope_params p) {
 #if RMS_NORM_ROPE_FUSION
     // Per-row offset in shared memory
     const uint ix = i0;
 #else
-    const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
+    const uint ix = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
 #endif
     return ix;
 }
@@ -34,26 +34,19 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out
     sin_theta = sin(theta) * mscale;
 }
 
-void rope_norm(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0;
-    const uint ix = rope_a_coord(i0, i01, i02, p);
+    uint idst = i0 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -63,7 +56,7 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -77,25 +70,19 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_neox(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -105,7 +92,7 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -120,26 +107,19 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
 }
 
 
-void rope_multi(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -156,26 +136,26 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (p.is_imrope != 0) {
         if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         } else {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     } else {
         if (sector < p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= p.sections[0] && sector < sec_w) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     }
 
@@ -191,20 +171,13 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_vision(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    const uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     const int sect_dims = p.sections[0] + p.sections[1];
     const int sec_w = p.sections[1] + p.sections[0];
@@ -213,11 +186,11 @@ void rope_vision(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (sector < p.sections[0]) {
         const uint p0 = sector;
-        theta_base = rope_data_pos[i02]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2]*pow(p.theta_scale, p0);
     }
     else if (sector >= p.sections[0] && sector < sec_w) {
         const uint p0 = sector - p.sections[0];
-        theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2 + p.ne02]*pow(p.theta_scale, p0);
     }
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;

ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_multi(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_multi(i0, i1, i2, i3, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_neox(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_neox(i0, i1, i2, i3, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_norm(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_norm(i0, i1, i2, i3, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl

@@ -5,24 +5,29 @@
 
 struct rope_params {
     uint rope_mode;
-    uint ncols;
     uint nrows;
     uint n_dims;
     float freq_scale;
-    uint p_delta_rows;
     float freq_base;
     float ext_factor;
     float attn_factor;
     float corr_dims[2];
     float theta_scale;
     uint has_ff;
-    uint ne02;
-    uint nb01;
-    uint nb02;
     int sections[4];
     uint is_imrope;
     uint is_back;
     uint set_rows_stride;
+
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
 };
 
 #endif // !defined(GGML_ROPE_PARAMS)

ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_vision(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_vision(i0, i1, i2, i3, pc);
 }

ggml/src/ggml.c

@@ -7517,8 +7517,11 @@ void ggml_quantize_free(void) {
 
     iq2xs_free_impl(GGML_TYPE_IQ2_XXS);
     iq2xs_free_impl(GGML_TYPE_IQ2_XS);
+    iq2xs_free_impl(GGML_TYPE_IQ2_S);
     iq2xs_free_impl(GGML_TYPE_IQ1_S);
+    iq2xs_free_impl(GGML_TYPE_IQ1_M);
     iq3xs_free_impl(256);
+    iq3xs_free_impl(512);
 
     ggml_critical_section_end();
 }

gguf-py/gguf/constants.py

@@ -284,6 +284,8 @@ class Keys:
     class ClipVision:
         PROJECTOR_TYPE      = "clip.vision.projector_type" # for mixed modality models
         IMAGE_SIZE          = "clip.vision.image_size"
+        IMAGE_MIN_PIXELS    = "clip.vision.image_min_pixels"
+        IMAGE_MAX_PIXELS    = "clip.vision.image_max_pixels"
         PREPROC_IMAGE_SIZE  = "clip.vision.preproc_image_size"
         PATCH_SIZE          = "clip.vision.patch_size"
         EMBEDDING_LENGTH    = "clip.vision.embedding_length"

gguf-py/gguf/gguf_writer.py

@@ -1113,6 +1113,12 @@ class GGUFWriter:
     def add_vision_image_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.IMAGE_SIZE, value)
 
+    def add_vision_max_pixels(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.IMAGE_MAX_PIXELS, value)
+
+    def add_vision_min_pixels(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.IMAGE_MIN_PIXELS, value)
+
     def add_vision_preproc_image_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.PREPROC_IMAGE_SIZE, value)
 

scripts/sync_vendor.py

@@ -12,8 +12,8 @@ vendor = {
     # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
     "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
 
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/httplib.h": "vendor/cpp-httplib/httplib.h",
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/LICENSE":   "vendor/cpp-httplib/LICENSE",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/httplib.h": "vendor/cpp-httplib/httplib.h",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/LICENSE":   "vendor/cpp-httplib/LICENSE",
 
     "https://raw.githubusercontent.com/sheredom/subprocess.h/b49c56e9fe214488493021017bf3954b91c7c1f5/subprocess.h": "vendor/sheredom/subprocess.h",
 }

src/llama-context.cpp

@@ -1027,11 +1027,7 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
         llama_sampler_chain_n(sampler) > 0;
 
     if (sampler && can_offload) {
-        ggml_backend_buffer_type_t buft = ggml_backend_dev_buffer_type(model.dev_output());
-        auto * host_buft = ggml_backend_dev_host_buffer_type(model.dev_output());
-        if (host_buft) {
-            buft = host_buft;
-        }
+        auto * buft = ggml_backend_dev_buffer_type(model.dev_output());
 
         sampler->iface->backend_init(sampler, buft);
 

src/llama-graph.cpp

@@ -2419,6 +2419,9 @@ void llm_graph_context::build_sampling() const {
         return;
     }
 
+    std::array<ggml_tensor *, 2> outs;
+    outs[0] = res->t_logits;
+
     auto inp_sampling = std::make_unique<llm_graph_input_sampling>(samplers);
     res->add_input(std::move(inp_sampling));
 
@@ -2439,14 +2442,14 @@ void llm_graph_context::build_sampling() const {
     // add a dummy row of logits
     // this trick makes the graph static, regardless of which samplers are activated
     // this is important in order to minimize graph reallocations
-    // TODO: use `ggml_build_forward_select()` when available (https://github.com/ggml-org/llama.cpp/pull/18550)
     ggml_tensor * logits_t = ggml_pad(ctx0, res->t_logits, 0, 1, 0, 0);
 
     for (const auto & [seq_id, sampler] : samplers) {
         const auto it = seq_to_logit_row.find(seq_id);
 
         // inactive samplers always work on the first row
-        const auto row_idx = seq_to_logit_row.find(seq_id) != seq_to_logit_row.end() ? it->second : 0;
+        const auto row_idx = it != seq_to_logit_row.end() ? it->second : 0;
+        const int i_out    = it != seq_to_logit_row.end() ? 1          : 0;
 
         ggml_tensor * logits_seq = ggml_view_1d(ctx0, logits_t, logits_t->ne[0], row_idx * logits_t->nb[1]);
         ggml_format_name(logits_seq, "logits_seq_%d", seq_id);
@@ -2463,22 +2466,26 @@ void llm_graph_context::build_sampling() const {
 
         if (data.sampled != nullptr) {
             res->t_sampled[seq_id] = data.sampled;
-            ggml_build_forward_expand(gf, data.sampled);
+            outs[1] = data.sampled;
+            ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
         }
 
         if (data.probs != nullptr) {
             res->t_sampled_probs[seq_id] = data.probs;
-            ggml_build_forward_expand(gf, data.probs);
+            outs[1] = data.probs;
+            ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
         }
 
         if (data.logits != nullptr) {
             res->t_sampled_logits[seq_id] = data.logits;
-            ggml_build_forward_expand(gf, data.logits);
+            outs[1] = data.logits;
+            ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
         }
 
         if (data.candidates != nullptr) {
             res->t_candidates[seq_id] = data.candidates;
-            ggml_build_forward_expand(gf, data.candidates);
+            outs[1] = data.candidates;
+            ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
         }
     }
 

src/llama-sampling.cpp

@@ -1025,11 +1025,7 @@ struct llama_sampler_dist : public llama_sampler_backend {
 
     std::mt19937 rng;
 
-    // backend input
-    struct ggml_tensor * inp_uniform;
-
-    ggml_context_ptr        inp_ctx;
-    ggml_backend_buffer_ptr inp_buf;
+    ggml_tensor * inp_uniform;
 };
 
 static const char * llama_sampler_dist_name(const struct llama_sampler * smpl) {
@@ -1138,37 +1134,10 @@ static bool llama_sampler_dist_backend_init(
         ggml_backend_buffer_type_t   buft) {
     auto * sctx = (llama_sampler_dist *) smpl->ctx;
 
-    // allocate inputs
-    {
-        ggml_init_params params = {
-            /*.mem_size   =*/ ggml_tensor_overhead(),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-
-        sctx->inp_ctx.reset(ggml_init(params));
-
-        // Create the uniform random scalar input tensor. This will be set by
-        // llama_sampler_dist_backend_set_input after this graph is built.
-        sctx->inp_uniform = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1);
-        ggml_set_name (sctx->inp_uniform, "uniform");
-        ggml_set_input(sctx->inp_uniform);
-
-        // Allocate all tensors from our context to the backend
-        sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
-
-        ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
-    }
-
     const bool res = llama_sampler_backend_support(smpl, buft);
 
     sctx->init(res);
 
-    if (!res) {
-        sctx->inp_ctx.reset(nullptr);
-        sctx->inp_buf.reset(nullptr);
-    }
-
     return res;
 }
 
@@ -1178,8 +1147,13 @@ static void llama_sampler_dist_backend_apply(
         struct ggml_cgraph        * gf,
         struct llama_sampler_data * data) {
     GGML_UNUSED(gf);
+
     auto * sctx = (llama_sampler_dist *) smpl->ctx;
 
+    sctx->inp_uniform = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+    ggml_set_name (sctx->inp_uniform, "uniform");
+    ggml_set_input(sctx->inp_uniform);
+
     struct ggml_tensor * probs = ggml_soft_max(ctx, data->logits);
     ggml_set_name(probs, "dist_probs");
 
@@ -1226,6 +1200,7 @@ static void llama_sampler_dist_backend_apply(
 
 static void llama_sampler_dist_backend_set_input(struct llama_sampler * smpl) {
     auto * sctx = (llama_sampler_dist *) smpl->ctx;
+
     GGML_ASSERT(sctx->inp_uniform != nullptr);
 
     // We sample in double precision and cast to float to match rnd numbers of
@@ -1262,8 +1237,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
             /* .seed_cur    = */ seed_cur,
             /* .rng         = */ std::mt19937(seed_cur),
             /* .inp_uniform = */ nullptr,
-            /* .inp_ctx     = */ nullptr,
-            /* .inp_buf     = */ nullptr,
         }
     );
 }
@@ -3461,9 +3434,6 @@ struct llama_sampler_logit_bias : public llama_sampler_backend {
 
     struct ggml_tensor * inp_logit_bias;
     struct ggml_tensor * inp_logit_idxs;
-
-    ggml_context_ptr        inp_ctx;
-    ggml_backend_buffer_ptr inp_buf;
 };
 
 static const char * llama_sampler_logit_bias_name(const struct llama_sampler * smpl) {
@@ -3526,6 +3496,16 @@ static void llama_sampler_logit_bias_backend_apply(
         return;
     }
 
+    const size_t n = sctx->logit_bias.size();
+
+    sctx->inp_logit_bias = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n);
+    ggml_set_name(sctx->inp_logit_bias, "logit_bias");
+    ggml_set_input(sctx->inp_logit_bias);
+
+    sctx->inp_logit_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n);
+    ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
+    ggml_set_input(sctx->inp_logit_idxs);
+
     ggml_tensor * cur = ggml_fill(ctx, data->logits, 0.0f);
 
     cur = ggml_reshape_2d(ctx, cur, 1, ggml_nelements(cur));
@@ -3562,6 +3542,8 @@ static void llama_sampler_logit_bias_backend_set_input(struct llama_sampler * sm
 static bool llama_sampler_logit_bias_backend_init(
         struct llama_sampler       * smpl,
         ggml_backend_buffer_type_t   buft) {
+    GGML_UNUSED(buft);
+
     auto * sctx = (llama_sampler_logit_bias *) smpl->ctx;
 
     sctx->init(true);
@@ -3570,29 +3552,6 @@ static bool llama_sampler_logit_bias_backend_init(
         return true;
     }
 
-    ggml_init_params params = {
-        /*.mem_size   =*/ 2*ggml_tensor_overhead(),
-        /*.mem_buffer =*/ nullptr,
-        /*.no_alloc   =*/ true,
-    };
-
-    sctx->inp_ctx.reset(ggml_init(params));
-
-    const size_t n = sctx->logit_bias.size();
-
-    sctx->inp_logit_bias = ggml_new_tensor_2d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1, n);
-    ggml_set_name(sctx->inp_logit_bias, "logit_bias");
-    ggml_set_input(sctx->inp_logit_bias);
-
-    sctx->inp_logit_idxs = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_I32, n);
-    ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
-    ggml_set_input(sctx->inp_logit_idxs);
-
-    // Allocate all tensors from our context to the backend
-    sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
-
-    ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
-
     return true;
 }
 
@@ -3628,8 +3587,6 @@ struct llama_sampler * llama_sampler_init_logit_bias(
             /* .to_search      = */ {},
             /* .inp_logit_bias = */ nullptr,
             /* .inp_logit_idxs = */ nullptr,
-            /* .inp_ctx        = */ nullptr,
-            /* .inp_buf        = */ nullptr,
         }
     );
 }

src/llama-vocab.cpp

@@ -2262,6 +2262,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
                         || t.first == "<|code_prefix|>" // GLM-4.5
+                        || t.first == "<|prefix|>"      // Falcon-H1-Tiny-Coder
                         ) {
                     special_fim_pre_id = t.second;
                     if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2282,6 +2283,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
                         || t.first == "<|code_suffix|>" // GLM-4.5
+                        || t.first == "<|suffix|>"      // Falcon-H1-Tiny-Coder
                         ) {
                     special_fim_suf_id = t.second;
                     if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2302,6 +2304,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
                         || t.first == "<|code_middle|>" // GLM-4.5
+                        || t.first == "<|middle|>"      // Falcon-H1-Tiny-Coder
                         ) {
                     special_fim_mid_id = t.second;
                     if ((attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

src/models/openelm.cpp

@@ -43,7 +43,7 @@ llm_build_openelm::llm_build_openelm(const llama_model & model, const llm_graph_
             ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head);
             cb(Kcur, "Kcur", il);
 
-            ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv)));
+            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv));
             cb(Vcur, "Vcur", il);
 
             Qcur = build_norm(Qcur,

src/models/qwen3next.cpp

@@ -265,9 +265,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
     cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
 
     ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp);
+    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
+                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
+
+    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
     cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
 
+    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
+    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
+
 
     // state to be updated per chunk
     ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
@@ -322,9 +328,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
             : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
 
         // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk));
+        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
         //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff)));
+        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
 
         // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
         ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));

tests/test-backend-ops.cpp

@@ -8032,6 +8032,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_IMROPE, GGML_ROPE_TYPE_VISION}) {
             for (bool ff : {false, true}) {
                 test_cases.emplace_back(new test_rope(type, {128,  32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 0, true, true));
+                test_cases.emplace_back(new test_rope(type, {128,  32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
+                test_cases.emplace_back(new test_rope(type, {128,  32, 2, 3}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
             }
         }
     }
@@ -8591,6 +8593,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
             output_printer->print_operation(info);
             return false;
         }
+        // Use reference implementation on the CPU backend for comparison
+        using ggml_backend_cpu_set_use_ref_t = void (*)(ggml_backend_t, bool);
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
+        auto * set_use_ref = (ggml_backend_cpu_set_use_ref_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_use_ref");
+        if (set_use_ref) {
+            set_use_ref(backend_cpu, true);
+        }
 
         size_t n_ok = 0;
         size_t                   tests_run = 0;

tests/test-jinja.cpp

@@ -189,12 +189,24 @@ static void test_conditionals(testing & t) {
         "negated"
     );
 
-    test_template(t, "in operator",
+    test_template(t, "in operator (element in array)",
         "{% if 'x' in items %}found{% endif %}",
         {{"items", json::array({"x", "y"})}},
         "found"
     );
 
+    test_template(t, "in operator (substring)",
+        "{% if 'bc' in 'abcd' %}found{% endif %}",
+        json::object(),
+        "found"
+    );
+
+    test_template(t, "in operator (object key)",
+        "{% if 'key' in obj %}found{% endif %}",
+        {{"obj", {{"key", 1}, {"other", 2}}}},
+        "found"
+    );
+
     test_template(t, "is defined",
         "{% if x is defined %}yes{% else %}no{% endif %}",
         {{"x", 1}},
@@ -1036,6 +1048,42 @@ static void test_tests(testing & t) {
         json::object(),
         "yes"
     );
+
+    test_template(t, "is in (array, true)",
+        "{{ 'yes' if 2 is in([1, 2, 3]) }}",
+        json::object(),
+        "yes"
+    );
+
+    test_template(t, "is in (array, false)",
+        "{{ 'yes' if 5 is in([1, 2, 3]) else 'no' }}",
+        json::object(),
+        "no"
+    );
+
+    test_template(t, "is in (string)",
+        "{{ 'yes' if 'bc' is in('abcde') }}",
+        json::object(),
+        "yes"
+    );
+
+    test_template(t, "is in (object keys)",
+        "{{ 'yes' if 'a' is in(obj) }}",
+        {{"obj", {{"a", 1}, {"b", 2}}}},
+        "yes"
+    );
+
+    test_template(t, "reject with in test",
+        "{{ items | reject('in', skip) | join(', ') }}",
+        {{"items", json::array({"a", "b", "c", "d"})}, {"skip", json::array({"b", "d"})}},
+        "a, c"
+    );
+
+    test_template(t, "select with in test",
+        "{{ items | select('in', keep) | join(', ') }}",
+        {{"items", json::array({"a", "b", "c", "d"})}, {"keep", json::array({"b", "c"})}},
+        "b, c"
+    );
 }
 
 static void test_string_methods(testing & t) {

tools/completion/completion.cpp

@@ -674,15 +674,12 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
-                int n_eval = (int) embd.size() - i;
-                if (n_eval > params.n_batch) {
-                    n_eval = params.n_batch;
-                }
-
+            if (!embd.empty()) {
+                int n_eval = (int) embd.size();
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                GGML_ASSERT(n_eval <= params.n_batch);
+                if (llama_decode(ctx, llama_batch_get_one(embd.data(), n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -743,7 +740,7 @@ int main(int argc, char ** argv) {
                 common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
 
                 ++n_consumed;
-                if ((int) embd.size() >= params.n_batch) {
+                if ((int) embd.size() == params.n_batch) {
                     break;
                 }
             }

tools/mtmd/clip-impl.h

@@ -36,6 +36,8 @@
 // vision-specific
 #define KEY_VISION_PROJ_TYPE    "clip.vision.projector_type" // for models with mixed modalities
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_IMAGE_MIN_PIXELS    "clip.vision.image_min_pixels"
+#define KEY_IMAGE_MAX_PIXELS    "clip.vision.image_max_pixels"
 #define KEY_PREPROC_IMAGE_SIZE  "clip.vision.preproc_image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"

tools/server/server-models.cpp

@@ -767,7 +767,7 @@ static bool router_validate_model(const std::string & name, server_models & mode
     }
     auto meta = models.get_meta(name);
     if (!meta.has_value()) {
-        res_err(res, format_error_response("model not found", ERROR_TYPE_INVALID_REQUEST));
+        res_err(res, format_error_response(string_format("model '%s' not found", name.c_str()), ERROR_TYPE_INVALID_REQUEST));
         return false;
     }
     if (models_autoload) {

vendor/cpp-httplib/CMakeLists.txt

@@ -3,9 +3,14 @@ license_add_file("cpp-httplib" "LICENSE")
 
 find_package(Threads REQUIRED)
 
+llama_add_compile_flags()
+
 add_library(${TARGET} STATIC httplib.cpp httplib.h)
-if (NOT MSVC)
-    # disable warnings in 3rd party code
+
+# disable warnings in 3rd party code
+if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+    target_compile_options(${TARGET} PRIVATE /w)
+else()
     target_compile_options(${TARGET} PRIVATE -w)
 endif()
 
@@ -146,6 +151,23 @@ elseif (LLAMA_OPENSSL)
     endif()
 endif()
 
+# disable warnings in 3rd party code
+if(LLAMA_BUILD_BORINGSSL OR LLAMA_BUILD_LIBRESSL)
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        target_compile_options(ssl PRIVATE /w)
+        target_compile_options(crypto PRIVATE /w)
+        if(LLAMA_BUILD_BORINGSSL)
+            target_compile_options(fipsmodule PRIVATE /w)
+        endif()
+    else()
+        target_compile_options(ssl PRIVATE -w)
+        target_compile_options(crypto PRIVATE -w)
+        if(LLAMA_BUILD_BORINGSSL)
+            target_compile_options(fipsmodule PRIVATE -w)
+        endif()
+    endif()
+endif()
+
 if (CPPHTTPLIB_OPENSSL_SUPPORT)
     target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT) # used in server.cpp
     if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")

vendor/cpp-httplib/httplib.cpp

@@ -117,6 +117,8 @@ time_t parse_http_date(const std::string &date_str) {
 
 #ifdef _WIN32
   return _mkgmtime(&tm_buf);
+#elif defined _AIX
+  return mktime(&tm_buf);
 #else
   return timegm(&tm_buf);
 #endif
@@ -1376,7 +1378,7 @@ int getaddrinfo_with_timeout(const char *node, const char *service,
 
   // Allocate on the heap, so the resolver thread can keep using the data.
   auto state = std::make_shared<GetAddrInfoState>();
-  state->node = node;
+  if (node) { state->node = node; }
   state->service = service;
   state->hints = *hints;
 
@@ -2896,10 +2898,20 @@ bool parse_range_header(const std::string &s, Ranges &ranges) try {
         return;
       }
 
-      const auto first =
-          static_cast<ssize_t>(lhs.empty() ? -1 : std::stoll(lhs));
-      const auto last =
-          static_cast<ssize_t>(rhs.empty() ? -1 : std::stoll(rhs));
+      ssize_t first = -1;
+      if (!lhs.empty()) {
+        ssize_t v;
+        auto res = detail::from_chars(lhs.data(), lhs.data() + lhs.size(), v);
+        if (res.ec == std::errc{}) { first = v; }
+      }
+
+      ssize_t last = -1;
+      if (!rhs.empty()) {
+        ssize_t v;
+        auto res = detail::from_chars(rhs.data(), rhs.data() + rhs.size(), v);
+        if (res.ec == std::errc{}) { last = v; }
+      }
+
       if ((first == -1 && last == -1) ||
           (first != -1 && last != -1 && first > last)) {
         all_valid_ranges = false;
@@ -2974,25 +2986,17 @@ bool parse_accept_header(const std::string &s,
         return;
       }
 
-#ifdef CPPHTTPLIB_NO_EXCEPTIONS
       {
-        std::istringstream iss(quality_str);
-        iss >> accept_entry.quality;
-
-        // Check if conversion was successful and entire string was consumed
-        if (iss.fail() || !iss.eof()) {
+        double v = 0.0;
+        auto res = detail::from_chars(
+            quality_str.data(), quality_str.data() + quality_str.size(), v);
+        if (res.ec == std::errc{}) {
+          accept_entry.quality = v;
+        } else {
           has_invalid_entry = true;
           return;
         }
       }
-#else
-      try {
-        accept_entry.quality = std::stod(quality_str);
-      } catch (...) {
-        has_invalid_entry = true;
-        return;
-      }
-#endif
       // Check if quality is in valid range [0.0, 1.0]
       if (accept_entry.quality < 0.0 || accept_entry.quality > 1.0) {
         has_invalid_entry = true;
@@ -5570,13 +5574,26 @@ bool Server::read_content(Stream &strm, Request &req, Response &res) {
           strm, req, res,
           // Regular
           [&](const char *buf, size_t n) {
+            // Prevent arithmetic overflow when checking sizes.
+            // Avoid computing (req.body.size() + n) directly because
+            // adding two unsigned `size_t` values can wrap around and
+            // produce a small result instead of indicating overflow.
+            // Instead, check using subtraction: ensure `n` does not
+            // exceed the remaining capacity `max_size() - size()`.
+            if (req.body.size() >= req.body.max_size() ||
+                n > req.body.max_size() - req.body.size()) {
+              return false;
+            }
+
             // Limit decompressed body size to payload_max_length_ to protect
             // against "zip bomb" attacks where a small compressed payload
             // decompresses to a massive size.
-            if (req.body.size() + n > payload_max_length_ ||
-                req.body.size() + n > req.body.max_size()) {
+            if (payload_max_length_ > 0 &&
+                (req.body.size() >= payload_max_length_ ||
+                 n > payload_max_length_ - req.body.size())) {
               return false;
             }
+
             req.body.append(buf, n);
             return true;
           },
@@ -5666,22 +5683,29 @@ bool Server::read_content_core(
   // oversized request and fail early (causing connection close). For SSL
   // builds we cannot reliably peek the decrypted application bytes, so keep
   // the original behaviour.
-#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT) && !defined(_WIN32)
+#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT)
   if (!req.has_header("Content-Length") &&
       !detail::is_chunked_transfer_encoding(req.headers)) {
-    socket_t s = strm.socket();
-    if (s != INVALID_SOCKET) {
-      // Peek up to payload_max_length_ + 1 bytes. If more than
-      // payload_max_length_ bytes are pending, reject the request.
-      size_t to_peek =
-          (payload_max_length_ > 0)
-              ? (std::min)(payload_max_length_ + 1, static_cast<size_t>(4096))
-              : 1;
-      std::vector<char> peekbuf(to_peek);
-      ssize_t n = ::recv(s, peekbuf.data(), to_peek, MSG_PEEK);
-      if (n > 0 && static_cast<size_t>(n) > payload_max_length_) {
-        // Indicate failure so connection will be closed.
-        return false;
+    // Only peek if payload_max_length is set to a finite value
+    if (payload_max_length_ > 0 &&
+        payload_max_length_ < (std::numeric_limits<size_t>::max)()) {
+      socket_t s = strm.socket();
+      if (s != INVALID_SOCKET) {
+        // Peek to check if there is any pending data
+        char peekbuf[1];
+        ssize_t n = ::recv(s, peekbuf, 1, MSG_PEEK);
+        if (n > 0) {
+          // There is data, so read it with payload limit enforcement
+          auto result = detail::read_content_without_length(
+              strm, payload_max_length_, out);
+          if (result == detail::ReadContentResult::PayloadTooLarge) {
+            res.status = StatusCode::PayloadTooLarge_413;
+            return false;
+          } else if (result != detail::ReadContentResult::Success) {
+            return false;
+          }
+          return true;
+        }
       }
     }
     return true;
@@ -6656,7 +6680,8 @@ void ClientImpl::close_socket(Socket &socket) {
 }
 
 bool ClientImpl::read_response_line(Stream &strm, const Request &req,
-                                           Response &res) const {
+                                           Response &res,
+                                           bool skip_100_continue) const {
   std::array<char, 2048> buf{};
 
   detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
@@ -6677,8 +6702,8 @@ bool ClientImpl::read_response_line(Stream &strm, const Request &req,
   res.status = std::stoi(std::string(m[2]));
   res.reason = std::string(m[3]);
 
-  // Ignore '100 Continue'
-  while (res.status == StatusCode::Continue_100) {
+  // Ignore '100 Continue' (only when not using Expect: 100-continue explicitly)
+  while (skip_100_continue && res.status == StatusCode::Continue_100) {
     if (!line_reader.getline()) { return false; } // CRLF
     if (!line_reader.getline()) { return false; } // next response line
 
@@ -7463,7 +7488,8 @@ bool ClientImpl::write_content_with_provider(Stream &strm,
 }
 
 bool ClientImpl::write_request(Stream &strm, Request &req,
-                                      bool close_connection, Error &error) {
+                                      bool close_connection, Error &error,
+                                      bool skip_body) {
   // Prepare additional headers
   if (close_connection) {
     if (!req.has_header("Connection")) {
@@ -7582,7 +7608,59 @@ bool ClientImpl::write_request(Stream &strm, Request &req,
     }
   }
 
+  // After sending request line and headers, wait briefly for an early server
+  // response (e.g. 4xx) and avoid sending a potentially large request body
+  // unnecessarily. This workaround is only enabled on Windows because Unix
+  // platforms surface write errors (EPIPE) earlier; on Windows kernel send
+  // buffering can accept large writes even when the peer already responded.
+  // Check the stream first (which covers SSL via `is_readable()`), then
+  // fall back to select on the socket. Only perform the wait for very large
+  // request bodies to avoid interfering with normal small requests and
+  // reduce side-effects. Poll briefly (up to 50ms as default) for an early
+  // response. Skip this check when using Expect: 100-continue, as the protocol
+  // handles early responses properly.
+#if defined(_WIN32)
+  if (!skip_body &&
+      req.body.size() > CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD &&
+      req.path.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) {
+    auto start = std::chrono::high_resolution_clock::now();
+
+    for (;;) {
+      // Prefer socket-level readiness to avoid SSL_pending() false-positives
+      // from SSL internals. If the underlying socket is readable, assume an
+      // early response may be present.
+      auto sock = strm.socket();
+      if (sock != INVALID_SOCKET && detail::select_read(sock, 0, 0) > 0) {
+        return false;
+      }
+
+      // Fallback to stream-level check for non-socket streams or when the
+      // socket isn't reporting readable. Avoid using `is_readable()` for
+      // SSL, since `SSL_pending()` may report buffered records that do not
+      // indicate a complete application-level response yet.
+      if (!is_ssl() && strm.is_readable()) { return false; }
+
+      auto now = std::chrono::high_resolution_clock::now();
+      auto elapsed =
+          std::chrono::duration_cast<std::chrono::milliseconds>(now - start)
+              .count();
+      if (elapsed >= CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND) {
+        break;
+      }
+
+      std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+  }
+#endif
+
   // Body
+  if (skip_body) { return true; }
+
+  return write_request_body(strm, req, error);
+}
+
+bool ClientImpl::write_request_body(Stream &strm, Request &req,
+                                           Error &error) {
   if (req.body.empty()) {
     return write_content_with_provider(strm, req, error);
   }
@@ -7758,8 +7836,20 @@ void ClientImpl::output_error_log(const Error &err,
 bool ClientImpl::process_request(Stream &strm, Request &req,
                                         Response &res, bool close_connection,
                                         Error &error) {
-  // Send request
-  if (!write_request(strm, req, close_connection, error)) { return false; }
+  // Auto-add Expect: 100-continue for large bodies
+  if (CPPHTTPLIB_EXPECT_100_THRESHOLD > 0 && !req.has_header("Expect")) {
+    auto body_size = req.body.empty() ? req.content_length_ : req.body.size();
+    if (body_size >= CPPHTTPLIB_EXPECT_100_THRESHOLD) {
+      req.set_header("Expect", "100-continue");
+    }
+  }
+
+  // Check for Expect: 100-continue
+  auto expect_100_continue = req.get_header_value("Expect") == "100-continue";
+
+  // Send request (skip body if using Expect: 100-continue)
+  auto write_request_success =
+      write_request(strm, req, close_connection, error, expect_100_continue);
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
   if (is_ssl()) {
@@ -7774,14 +7864,48 @@ bool ClientImpl::process_request(Stream &strm, Request &req,
   }
 #endif
 
+  // Handle Expect: 100-continue with timeout
+  if (expect_100_continue && CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND > 0) {
+    time_t sec = CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND / 1000;
+    time_t usec = (CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND % 1000) * 1000;
+    auto ret = detail::select_read(strm.socket(), sec, usec);
+    if (ret <= 0) {
+      // Timeout or error: send body anyway (server didn't respond in time)
+      if (!write_request_body(strm, req, error)) { return false; }
+      expect_100_continue = false; // Switch to normal response handling
+    }
+  }
+
   // Receive response and headers
-  if (!read_response_line(strm, req, res) ||
+  // When using Expect: 100-continue, don't auto-skip `100 Continue` response
+  if (!read_response_line(strm, req, res, !expect_100_continue) ||
       !detail::read_headers(strm, res.headers)) {
-    error = Error::Read;
+    if (write_request_success) { error = Error::Read; }
     output_error_log(error, &req);
     return false;
   }
 
+  if (!write_request_success) { return false; }
+
+  // Handle Expect: 100-continue response
+  if (expect_100_continue) {
+    if (res.status == StatusCode::Continue_100) {
+      // Server accepted, send the body
+      if (!write_request_body(strm, req, error)) { return false; }
+
+      // Read the actual response
+      res.headers.clear();
+      res.body.clear();
+      if (!read_response_line(strm, req, res) ||
+          !detail::read_headers(strm, res.headers)) {
+        error = Error::Read;
+        output_error_log(error, &req);
+        return false;
+      }
+    }
+    // If not 100 Continue, server returned an error; proceed with that response
+  }
+
   // Body
   if ((res.status != StatusCode::NoContent_204) && req.method != "HEAD" &&
       req.method != "CONNECT") {
@@ -9543,7 +9667,7 @@ bool SSLClient::load_certs() {
         last_openssl_error_ = ERR_get_error();
         ret = false;
       }
-    } else {
+    } else if (!ca_cert_store_) {
       auto loaded = false;
 #ifdef _WIN32
       loaded =
@@ -9790,7 +9914,11 @@ bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
 
 bool SSLClient::check_host_name(const char *pattern,
                                        size_t pattern_len) const {
-  if (host_.size() == pattern_len && host_ == pattern) { return true; }
+  // Exact match (case-insensitive)
+  if (host_.size() == pattern_len &&
+      detail::case_ignore::equal(host_, std::string(pattern, pattern_len))) {
+    return true;
+  }
 
   // Wildcard match
   // https://bugs.launchpad.net/ubuntu/+source/firefox-3.0/+bug/376484
@@ -9805,9 +9933,23 @@ bool SSLClient::check_host_name(const char *pattern,
   auto itr = pattern_components.begin();
   for (const auto &h : host_components_) {
     auto &p = *itr;
-    if (p != h && p != "*") {
-      auto partial_match = (p.size() > 0 && p[p.size() - 1] == '*' &&
-                            !p.compare(0, p.size() - 1, h));
+    if (!httplib::detail::case_ignore::equal(p, h) && p != "*") {
+      bool partial_match = false;
+      if (!p.empty() && p[p.size() - 1] == '*') {
+        const auto prefix_length = p.size() - 1;
+        if (prefix_length == 0) {
+          partial_match = true;
+        } else if (h.size() >= prefix_length) {
+          partial_match =
+              std::equal(p.begin(),
+                         p.begin() + static_cast<std::string::difference_type>(
+                                         prefix_length),
+                         h.begin(), [](const char ca, const char cb) {
+                           return httplib::detail::case_ignore::to_lower(ca) ==
+                                  httplib::detail::case_ignore::to_lower(cb);
+                         });
+        }
+      }
       if (!partial_match) { return false; }
     }
     ++itr;

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.30.1"
-#define CPPHTTPLIB_VERSION_NUM "0x001E01"
+#define CPPHTTPLIB_VERSION "0.30.2"
+#define CPPHTTPLIB_VERSION_NUM "0x001E02"
 
 /*
  * Platform compatibility check
@@ -98,6 +98,22 @@
 #define CPPHTTPLIB_CLIENT_MAX_TIMEOUT_MSECOND 0
 #endif
 
+#ifndef CPPHTTPLIB_EXPECT_100_THRESHOLD
+#define CPPHTTPLIB_EXPECT_100_THRESHOLD 1024
+#endif
+
+#ifndef CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND
+#define CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND 1000
+#endif
+
+#ifndef CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD
+#define CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD (1024 * 1024)
+#endif
+
+#ifndef CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND
+#define CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND 50
+#endif
+
 #ifndef CPPHTTPLIB_IDLE_INTERVAL_SECOND
 #define CPPHTTPLIB_IDLE_INTERVAL_SECOND 0
 #endif
@@ -286,8 +302,10 @@ using socket_t = int;
 #include <atomic>
 #include <cassert>
 #include <cctype>
+#include <chrono>
 #include <climits>
 #include <condition_variable>
+#include <cstdlib>
 #include <cstring>
 #include <errno.h>
 #include <exception>
@@ -305,6 +323,7 @@ using socket_t = int;
 #include <sstream>
 #include <string>
 #include <sys/stat.h>
+#include <system_error>
 #include <thread>
 #include <unordered_map>
 #include <unordered_set>
@@ -494,6 +513,69 @@ private:
   bool execute_on_destruction;
 };
 
+// Simple from_chars implementation for integer and double types (C++17
+// substitute)
+template <typename T> struct from_chars_result {
+  const char *ptr;
+  std::errc ec;
+};
+
+template <typename T>
+inline from_chars_result<T> from_chars(const char *first, const char *last,
+                                       T &value, int base = 10) {
+  value = 0;
+  const char *p = first;
+  bool negative = false;
+
+  if (p != last && *p == '-') {
+    negative = true;
+    ++p;
+  }
+  if (p == last) { return {first, std::errc::invalid_argument}; }
+
+  T result = 0;
+  for (; p != last; ++p) {
+    char c = *p;
+    int digit = -1;
+    if ('0' <= c && c <= '9') {
+      digit = c - '0';
+    } else if ('a' <= c && c <= 'z') {
+      digit = c - 'a' + 10;
+    } else if ('A' <= c && c <= 'Z') {
+      digit = c - 'A' + 10;
+    } else {
+      break;
+    }
+
+    if (digit < 0 || digit >= base) { break; }
+    if (result > ((std::numeric_limits<T>::max)() - digit) / base) {
+      return {p, std::errc::result_out_of_range};
+    }
+    result = result * base + digit;
+  }
+
+  if (p == first || (negative && p == first + 1)) {
+    return {first, std::errc::invalid_argument};
+  }
+
+  value = negative ? -result : result;
+  return {p, std::errc{}};
+}
+
+// from_chars for double (simple wrapper for strtod)
+inline from_chars_result<double> from_chars(const char *first, const char *last,
+                                            double &value) {
+  std::string s(first, last);
+  char *endptr = nullptr;
+  errno = 0;
+  value = std::strtod(s.c_str(), &endptr);
+  if (endptr == s.c_str()) { return {first, std::errc::invalid_argument}; }
+  if (errno == ERANGE) {
+    return {first + (endptr - s.c_str()), std::errc::result_out_of_range};
+  }
+  return {first + (endptr - s.c_str()), std::errc{}};
+}
+
 } // namespace detail
 
 enum SSLVerifierResponse {
@@ -1848,10 +1930,11 @@ private:
   Result send_(Request &&req);
 
   socket_t create_client_socket(Error &error) const;
-  bool read_response_line(Stream &strm, const Request &req,
-                          Response &res) const;
+  bool read_response_line(Stream &strm, const Request &req, Response &res,
+                          bool skip_100_continue = true) const;
   bool write_request(Stream &strm, Request &req, bool close_connection,
-                     Error &error);
+                     Error &error, bool skip_body = false);
+  bool write_request_body(Stream &strm, Request &req, Error &error);
   void prepare_default_headers(Request &r, bool for_stream,
                                const std::string &ct);
   bool redirect(Request &req, Response &res, Error &error);
@@ -3243,10 +3326,11 @@ private:
       msg.id = value;
     } else if (field == "retry") {
       // Parse retry interval in milliseconds
-      try {
-        retry_ms = std::stoi(value);
-      } catch (...) {
-        // Invalid retry value, ignore
+      {
+        int v = 0;
+        auto res =
+            detail::from_chars(value.data(), value.data() + value.size(), v);
+        if (res.ec == std::errc{}) { retry_ms = v; }
       }
     }
     // Unknown fields are ignored per SSE spec