llama-cpp-bin 9326.0.0__tar.gz → 9341.0.0__tar.gz

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9341.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9326.0.0
+Version: 9341.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9341.0.0}/llama.cpp/ggml/src/ggml-metal/ggml-metal-device.h

@@ -215,6 +215,30 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
 // device
 //
 
+enum ggml_metal_device_id {
+    GGML_METAL_DEVICE_GENERIC = 0,
+
+    GGML_METAL_DEVICE_M1,
+    GGML_METAL_DEVICE_M1_PRO,
+    GGML_METAL_DEVICE_M1_MAX,
+    GGML_METAL_DEVICE_M1_ULTRA,
+    GGML_METAL_DEVICE_M2,
+    GGML_METAL_DEVICE_M2_PRO,
+    GGML_METAL_DEVICE_M2_MAX,
+    GGML_METAL_DEVICE_M2_ULTRA,
+    GGML_METAL_DEVICE_M3,
+    GGML_METAL_DEVICE_M3_PRO,
+    GGML_METAL_DEVICE_M3_MAX,
+    GGML_METAL_DEVICE_M3_ULTRA,
+    GGML_METAL_DEVICE_M4,
+    GGML_METAL_DEVICE_M4_PRO,
+    GGML_METAL_DEVICE_M4_MAX,
+    GGML_METAL_DEVICE_M5,
+    GGML_METAL_DEVICE_M5_PRO,
+    GGML_METAL_DEVICE_M5_MAX,
+    GGML_METAL_DEVICE_M5_ULTRA,
+};
+
 struct ggml_metal_device_props {
     int device;
     char name[128];
@@ -234,6 +258,8 @@ struct ggml_metal_device_props {
 
     bool supports_gpu_family_apple7;
 
+    enum ggml_metal_device_id device_id;
+
     int op_offload_min_batch_size;
 };
 

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9341.0.0}/llama.cpp/ggml/src/ggml-sycl/common.hpp

@@ -224,6 +224,7 @@ struct sycl_device_info {
     int max_wg_per_cu; // max work groups per compute unit - refer to
                        // cudaOccupancyMaxActiveBlocksPerMultiprocessor
     bool    vmm;                // virtual memory support
+    size_t  vmm_granularity;    // granularity of virtual memory
     size_t  total_vram;
     sycl_hw_info hw_info;
     optimize_feature opt_feature;
@@ -244,6 +245,8 @@ struct ggml_sycl_device_info {
 
 const ggml_sycl_device_info & ggml_sycl_info();
 
+static constexpr size_t SYCL_BUFFER_ALIGNMENT = 128;
+
 struct ggml_sycl_pool {
     virtual ~ggml_sycl_pool() = default;
 

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9341.0.0}/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -19,6 +19,7 @@
 #include <cstdlib>
 #include <float.h>
 #include <limits>
+#include <optional>
 #include <stdint.h>
 #include <stdio.h>
 #include <vector>
@@ -37,6 +38,11 @@
 #if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
 #    include <sycl/ext/oneapi/experimental/async_alloc/async_alloc.hpp>
 #endif
+#if SYCL_EXT_ONEAPI_VIRTUAL_MEM
+#    include <sycl/ext/oneapi/virtual_mem/physical_mem.hpp>
+#    include <sycl/ext/oneapi/virtual_mem/virtual_mem.hpp>
+#    define GGML_SYCL_USE_VMM
+#endif
 #include <sycl/half_type.hpp>
 
 #include "ggml.h"
@@ -70,6 +76,7 @@ int g_ggml_sycl_debug = 0;
 int g_ggml_sycl_disable_optimize = 0;
 int g_ggml_sycl_disable_graph = 0;
 int g_ggml_sycl_disable_dnn = 0;
+int g_ggml_sycl_enable_vmm = 1;
 int g_ggml_sycl_prioritize_dmmv = 0;
 int g_ggml_sycl_use_async_mem_op = 0;
 int g_ggml_sycl_use_async_mem_op_requested = 1;
@@ -96,13 +103,30 @@ static ggml_sycl_device_info ggml_sycl_init() {
 //     GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__);
 // #endif
     for (int i = 0; i < info.device_count; ++i) {
-        info.devices[i].vmm = 0;
         dpct::device_info prop;
         auto & device = dpct::dev_mgr::instance().get_device(i);
 
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
             prop, device)));
 
+#if !defined(GGML_SYCL_USE_VMM)
+        info.devices[i].vmm = 0;
+#else
+        info.devices[i].vmm = device.has(sycl::aspect::ext_oneapi_virtual_mem);
+        if (info.devices[i].vmm) {
+            // NB: SYCL's get_mem_granularity always returns the _minimum_ granularity,
+            // but the L0 API requires a larger page size for allocs above 2 MiB and
+            // rejects non-multiples with UR_RESULT_ERROR_INVALID_VALUE [sic].
+            // Here we clamp it to 2 MiB for simplicity, but other devices may require
+            // calling zeVirtualMemQueryPageSize or yet unexposed public API.
+            const size_t physical_page = 2ull << 20; // 2 MiB
+            info.devices[i].vmm_granularity = std::max<size_t>(
+                sycl::ext::oneapi::experimental::get_mem_granularity(
+                    device, sycl::context(device)),
+                physical_page);
+        }
+#endif
+
         info.default_tensor_split[i] = total_vram;
         total_vram += prop.get_global_mem_size();
 
@@ -234,6 +258,7 @@ static void ggml_check_sycl() try {
         g_ggml_sycl_disable_optimize = get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
         g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
+        g_ggml_sycl_enable_vmm = get_sycl_env("GGML_SYCL_ENABLE_VMM", 1);
         g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
 #ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO
         g_ggml_sycl_enable_level_zero = get_sycl_env("GGML_SYCL_ENABLE_LEVEL_ZERO", ggml_sycl_info().ext_oneapi_level_zero);
@@ -275,6 +300,11 @@ static void ggml_check_sycl() try {
 #else
         GGML_LOG_INFO("  GGML_SYCL_SUPPORT_LEVEL_ZERO: no\n");
 #endif
+#if defined(GGML_SYCL_USE_VMM)
+        GGML_LOG_INFO("  GGML_SYCL_USE_VMM: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_USE_VMM: no\n");
+#endif
 
         GGML_LOG_INFO("Running with Environment Variables:\n");
         GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
@@ -293,6 +323,11 @@ static void ggml_check_sycl() try {
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: %d\n", g_ggml_sycl_disable_dnn);
 #else
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n");
+#endif
+#if defined(GGML_SYCL_USE_VMM)
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: %d\n", g_ggml_sycl_enable_vmm);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: virtual memory extension is not available\n");
 #endif
         GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
         g_ggml_sycl_use_async_mem_op_requested = get_sycl_env("GGML_SYCL_USE_ASYNC_MEM_OP", 1);
@@ -754,7 +789,7 @@ catch (sycl::exception const &exc) {
 }
 
 static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return 128;
+    return SYCL_BUFFER_ALIGNMENT;
     GGML_UNUSED(buft);
 }
 
@@ -1177,7 +1212,7 @@ static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(gg
 }
 
 static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return 128;
+    return SYCL_BUFFER_ALIGNMENT;
     GGML_UNUSED(buft);
 }
 
@@ -1462,6 +1497,121 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
     }
 };
 
+// pool with virtual memory management
+#if defined(GGML_SYCL_USE_VMM)
+struct ggml_sycl_pool_vmm : public ggml_sycl_pool {
+    static const size_t SYCL_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
+
+    int           device;
+    sycl::context ctx;
+    sycl::device  dev;
+
+    uintptr_t pool_addr = 0;
+    size_t    pool_used = 0;
+    size_t    pool_size = 0;
+    size_t    granularity;
+
+    // physical_mem owns the commits (unlike cuMemMap)
+    struct mapping {
+        sycl::ext::oneapi::experimental::physical_mem phys;
+        void * map_ptr;
+    };
+    std::vector<mapping> mappings;
+
+    explicit ggml_sycl_pool_vmm(queue_ptr qptr_, int device_) :
+        device(device_),
+        ctx(qptr_->get_context()),
+        dev(qptr_->get_device()),
+        granularity(ggml_sycl_info().devices[device_].vmm_granularity) {
+    }
+
+    ~ggml_sycl_pool_vmm() {
+        if (pool_addr == 0) {
+            return;
+        }
+
+        // Per spec, unmap must (a) match the exact (ptr, size) of an earlier
+        // physical_mem::map() call and (b) precede destruction of the
+        // physical_mem objects (their dtors won't unmap).
+        for (auto & m : mappings) {
+            SYCL_CHECK(CHECK_TRY_ERROR(sycl::ext::oneapi::experimental::unmap(
+                m.map_ptr, m.phys.size(), ctx)));
+        }
+        SYCL_CHECK(CHECK_TRY_ERROR(sycl::ext::oneapi::experimental::free_virtual_mem(
+            pool_addr, SYCL_POOL_VMM_MAX_SIZE, ctx)));
+    }
+
+    void * alloc(size_t size, size_t * actual_size) override {
+        // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
+        size = GGML_PAD(size, SYCL_BUFFER_ALIGNMENT);
+
+        size_t avail = pool_size - pool_used;
+
+        if (size > avail) {
+            // round up to the next multiple of the granularity
+            size_t reserve_size = GGML_PAD(size - avail, granularity);
+
+            GGML_ASSERT(pool_size + reserve_size <= SYCL_POOL_VMM_MAX_SIZE);
+
+            // allocate more physical memory
+            std::optional<sycl::ext::oneapi::experimental::physical_mem> phys;
+            SYCL_CHECK(CHECK_TRY_ERROR(phys.emplace(dev, ctx, reserve_size)));
+
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                SYCL_CHECK(CHECK_TRY_ERROR(
+                    pool_addr = sycl::ext::oneapi::experimental::reserve_virtual_mem(
+                        SYCL_POOL_VMM_MAX_SIZE, ctx)));
+            }
+
+            // map at the end of the pool
+            void * map_ptr = nullptr;
+            SYCL_CHECK(CHECK_TRY_ERROR(
+                map_ptr = phys->map(pool_addr + pool_size, reserve_size,
+                                    sycl::ext::oneapi::experimental::address_access_mode::read_write)));
+
+            // stash these so we could unmap this exact range in dtor
+            mappings.push_back({
+                std::move(*phys),
+                map_ptr,
+            });
+
+            // add to the pool
+            pool_size += reserve_size;
+
+#ifdef DEBUG_SYCL_MALLOC
+            GGML_LOG_INFO("sycl pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+                          device, (unsigned long long) (pool_size/1024/1024),
+                          (unsigned long long) (reserve_size/1024/1024));
+#endif
+        }
+
+        GGML_ASSERT(pool_addr != 0);
+
+        void * ptr = reinterpret_cast<void *>(pool_addr + pool_used);
+        *actual_size = size;
+        pool_used += size;
+
+#ifdef DEBUG_SYCL_MALLOC
+        GGML_LOG_INFO("sycl pool[%d]: allocated %llu bytes at %p\n", device, (unsigned long long) size, ptr);
+#endif
+
+        return ptr;
+    }
+
+    void free(void * ptr, size_t size) override {
+#ifdef DEBUG_SYCL_MALLOC
+        GGML_LOG_INFO("sycl pool[%d]: freed %llu bytes at %p\n", device, (unsigned long long) size, ptr);
+#endif
+
+        pool_used -= size;
+
+        // all deallocations must be in reverse order of the allocations
+        GGML_ASSERT(ptr == reinterpret_cast<void *>(pool_addr + pool_used));
+    }
+};
+#endif // defined(GGML_SYCL_USE_VMM)
+
 struct ggml_sycl_pool_host : public ggml_sycl_pool {
     queue_ptr qptr;
     int       device;
@@ -1542,20 +1692,19 @@ std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_host(que
 }
 
 std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_device(queue_ptr qptr, int device) {
-    // TBD: NO VMM support
-    // if (ggml_sycl_info().devices[device].vmm) {
-    //     return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_vmm(device));
-    // }
-   return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_leg(qptr, device));
+#if defined(GGML_SYCL_USE_VMM)
+    if (g_ggml_sycl_enable_vmm && ggml_sycl_info().devices[device].vmm) {
+        return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_vmm(qptr, device));
+    }
+#endif // defined(GGML_SYCL_USE_VMM)
+    return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_leg(qptr, device));
 }
 
+
 std::unique_ptr<ggml_sycl_fattn_kv_buffers> ggml_backend_sycl_context::new_fattn_kv_buffers(queue_ptr qptr, int device) {
     return std::unique_ptr<ggml_sycl_fattn_kv_buffers>(new ggml_sycl_fattn_kv_buffers(qptr, device));
 }
 
-// TBD pool with virtual memory management
-// struct ggml_sycl_pool_vmm : public ggml_sycl_pool
-
 /// kernels
 typedef void (*ggml_sycl_op_mul_mat_t)(
     ggml_backend_sycl_context & ctx,

{llama_cpp_bin-9326.0.0 → llama_cpp_bin-9341.0.0}/llama.cpp/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -52,7 +52,7 @@
 #define WEBGPU_MUL_MAT_VEC_LEGACY_Q_OUTPUTS_PER_WG 4
 #define WEBGPU_MUL_MAT_VEC_K_Q_OUTPUTS_PER_WG      4
 
-// default size for legacy matrix multiplication
+// default size for reg-tile matrix multiplication
 #define WEBGPU_MUL_MAT_WG_SIZE 256
 
 // Same hash combine function as in boost
@@ -93,6 +93,8 @@ struct ggml_webgpu_shader_lib_context {
     uint32_t sg_mat_k                 = 0;
     uint32_t min_subgroup_size        = 0;
     uint32_t max_subgroup_size        = 0;
+    bool     supports_dot_product     = false;
+    std::string vendor;
 };
 
 struct webgpu_pipeline {
@@ -850,31 +852,15 @@ inline ggml_webgpu_flash_attn_decisions ggml_webgpu_flash_attn_get_decisions(
 
 /** Matrix Multiplication **/
 
-struct ggml_webgpu_legacy_mul_mat_pipeline_key {
-    ggml_type src0_type;
-    ggml_type src1_type;
-
-    bool operator==(const ggml_webgpu_legacy_mul_mat_pipeline_key & other) const {
-        return src0_type == other.src0_type && src1_type == other.src1_type;
-    }
-};
-
-struct ggml_webgpu_legacy_mul_mat_pipeline_key_hash {
-    size_t operator()(const ggml_webgpu_legacy_mul_mat_pipeline_key & key) const {
-        size_t seed = 0;
-        ggml_webgpu_hash_combine(seed, key.src0_type);
-        ggml_webgpu_hash_combine(seed, key.src1_type);
-        return seed;
-    }
-};
-
 struct ggml_webgpu_mul_mat_vec_pipeline_key {
     ggml_type src0_type;
     ggml_type src1_type;
     int       vectorized;
+    bool      use_mmvq;
 
     bool operator==(const ggml_webgpu_mul_mat_vec_pipeline_key & other) const {
-        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized;
+        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
+               use_mmvq == other.use_mmvq;
     }
 };
 
@@ -884,6 +870,7 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key_hash {
         ggml_webgpu_hash_combine(seed, key.src0_type);
         ggml_webgpu_hash_combine(seed, key.src1_type);
         ggml_webgpu_hash_combine(seed, key.vectorized);
+        ggml_webgpu_hash_combine(seed, key.use_mmvq);
         return seed;
     }
 };
@@ -894,6 +881,20 @@ struct ggml_webgpu_mul_mat_vec_shader_decisions {
     uint32_t vec_size;
 };
 
+struct ggml_webgpu_quantize_q8_pipeline_key {
+    ggml_type src0_type;
+
+    bool operator==(const ggml_webgpu_quantize_q8_pipeline_key & other) const { return src0_type == other.src0_type; }
+};
+
+struct ggml_webgpu_quantize_q8_pipeline_key_hash {
+    size_t operator()(const ggml_webgpu_quantize_q8_pipeline_key & key) const {
+        size_t seed = 0;
+        ggml_webgpu_hash_combine(seed, key.src0_type);
+        return seed;
+    }
+};
+
 struct ggml_webgpu_mul_mat_pipeline_key {
     ggml_type src0_type;
     ggml_type src1_type;
@@ -1051,6 +1052,36 @@ struct ggml_webgpu_soft_max_pipeline_key_hash {
     }
 };
 
+/** MMVQ **/
+
+inline bool ggml_webgpu_can_use_mmvq(const ggml_tensor * src0,
+                                     const ggml_tensor * src1,
+                                     bool                supports_dot_product,
+                                     const std::string & vendor) {
+    if (src1->ne[1] == 1) {
+        bool supports_dp4a = vendor == "amd" || vendor == "intel" || vendor == "nvidia";
+        if (supports_dp4a && supports_dot_product) {
+            switch (src1->type) {
+                case GGML_TYPE_F32:
+                    switch (src0->type) {
+                        case GGML_TYPE_Q4_0:
+                        case GGML_TYPE_Q4_1:
+                        case GGML_TYPE_Q8_0:
+                        case GGML_TYPE_Q2_K:
+                        case GGML_TYPE_Q4_K:
+                            return src0->ne[0] % 4 == 0;
+                        default:
+                            break;
+                    }
+                    break;
+                default:
+                    break;
+            }
+        }
+    }
+    return false;
+}
+
 class ggml_webgpu_shader_lib {
     wgpu::Device           device;
     pre_wgsl::Preprocessor preprocessor;
@@ -1099,14 +1130,12 @@ class ggml_webgpu_shader_lib {
                        webgpu_pipeline,
                        ggml_webgpu_flash_attn_blk_pipeline_key_hash>
         flash_attn_blk_pipelines;
-    std::unordered_map<ggml_webgpu_legacy_mul_mat_pipeline_key,
-                       webgpu_pipeline,
-                       ggml_webgpu_legacy_mul_mat_pipeline_key_hash>
-        mul_mat_legacy_pipelines;  // legacy mul_mat (non-subgroup/non-regtile/non-vec)
     std::unordered_map<ggml_webgpu_mul_mat_vec_pipeline_key, webgpu_pipeline, ggml_webgpu_mul_mat_vec_pipeline_key_hash>
         mul_mat_vec_pipelines;     // fast mat-vec (n==1)
     std::unordered_map<ggml_webgpu_mul_mat_pipeline_key, webgpu_pipeline, ggml_webgpu_mul_mat_pipeline_key_hash>
                                              mul_mat_fast_pipelines;       // fast mat-mat (reg-tile or subgroup)
+    std::unordered_map<ggml_webgpu_quantize_q8_pipeline_key, webgpu_pipeline, ggml_webgpu_quantize_q8_pipeline_key_hash>
+                                             quantize_q8_pipelines;
     std::unordered_map<int, webgpu_pipeline> mul_mat_id_gather_pipelines;  // key is fixed
     std::unordered_map<ggml_webgpu_mul_mat_id_pipeline_key, webgpu_pipeline, ggml_webgpu_mul_mat_id_pipeline_key_hash>
         mul_mat_id_pipelines;                                              // src0_type/src1_type
@@ -1631,7 +1660,7 @@ class ggml_webgpu_shader_lib {
         key.type                              = context.dst->type;
         key.d_state                           = (int) context.src0->ne[0];
         key.xbc_overlap                       = ggml_webgpu_tensor_overlap(context.src1, context.src4) &&
-                          ggml_webgpu_tensor_overlap(context.src1, context.src5);
+                                                ggml_webgpu_tensor_overlap(context.src1, context.src5);
 
         auto it = ssm_scan_pipelines.find(key);
         if (it != ssm_scan_pipelines.end()) {
@@ -1744,6 +1773,44 @@ class ggml_webgpu_shader_lib {
         return pad_pipelines[key];
     }
 
+    webgpu_pipeline get_quantize_q8_pipeline(const ggml_webgpu_shader_lib_context & context) {
+        ggml_webgpu_quantize_q8_pipeline_key key = {};
+        key.src0_type                            = context.src0->type;
+
+        auto it = quantize_q8_pipelines.find(key);
+        if (it != quantize_q8_pipelines.end()) {
+            return it->second;
+        }
+        const char *             shader_src = wgsl_quantize_q8;
+        std::vector<std::string> defines;
+        std::string              variant = "quantize_q8";
+
+        uint32_t wg_size = WEBGPU_MUL_MAT_VEC_WG_SIZE;
+
+        defines.push_back("SRC1_INNER_TYPE=f32");
+        defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
+
+        const struct ggml_type_traits * src0_traits = ggml_get_type_traits(context.src0->type);
+        std::string                     src0_name   = src0_traits->type_name;
+        std::string                     type_upper  = src0_name;
+        variant += "_" + src0_name;
+        std::transform(type_upper.begin(), type_upper.end(), type_upper.begin(), ::toupper);
+
+        defines.push_back("MUL_ACC_" + type_upper);
+        defines.push_back("Q8_1_T");
+
+        defines.push_back(context.supports_subgroups ? "USE_SUBGROUP_REDUCTION" : "USE_WORKGROUP_REDUCTION");
+        variant += context.supports_subgroups ? "_sg_reduce" : "_wg_reduce";
+
+        auto processed             = preprocessor.preprocess(shader_src, defines);
+        auto decisions             = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        decisions->wg_size         = wg_size;
+        webgpu_pipeline pipeline   = ggml_webgpu_create_pipeline(device, processed, variant);
+        pipeline.context           = decisions;
+        quantize_q8_pipelines[key] = pipeline;
+        return quantize_q8_pipelines[key];
+    }
+
     webgpu_pipeline get_mul_mat_vec_pipeline(const ggml_webgpu_shader_lib_context & context) {
         ggml_webgpu_mul_mat_vec_pipeline_key key = {};
         key.src0_type                            = context.src0->type;
@@ -1752,6 +1819,8 @@ class ggml_webgpu_shader_lib {
                           (context.src0->type == GGML_TYPE_F32 || context.src0->type == GGML_TYPE_F16)) ?
                                                        1 :
                                                        0;
+        key.use_mmvq                             =
+            ggml_webgpu_can_use_mmvq(context.src0, context.src1, context.supports_dot_product, context.vendor);
 
         auto it = mul_mat_vec_pipelines.find(key);
         if (it != mul_mat_vec_pipelines.end()) {
@@ -1788,6 +1857,19 @@ class ggml_webgpu_shader_lib {
                     defines.push_back("U32_DEQUANT_HELPERS");
                     defines.push_back("SRC0_INNER_TYPE=u32");
                     switch (context.src0->type) {
+                        case GGML_TYPE_Q8_0:
+                        case GGML_TYPE_Q4_0:
+                        case GGML_TYPE_Q4_1:
+                            if (key.use_mmvq) {
+                                defines.push_back("LEGACY_QUANTS");
+                            }
+                            break;
+                        case GGML_TYPE_Q2_K:
+                        case GGML_TYPE_Q4_K:
+                            if (key.use_mmvq) {
+                                defines.push_back("K_QUANTS");
+                            }
+                            break;
                         case GGML_TYPE_IQ1_S:
                         case GGML_TYPE_IQ1_M:
                         case GGML_TYPE_IQ2_S:
@@ -1840,6 +1922,11 @@ class ggml_webgpu_shader_lib {
             outputs_per_wg = WEBGPU_MUL_MAT_VEC_LEGACY_Q_OUTPUTS_PER_WG;
         }
 
+        if (key.use_mmvq) {
+            defines.push_back("MMVQ");
+            defines.push_back("Q8_1_T");
+        }
+
         defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
         defines.push_back(std::string("OUTPUTS_PER_WG=") + std::to_string(outputs_per_wg));
         defines.push_back(context.supports_subgroups ? "USE_SUBGROUP_REDUCTION" : "USE_WORKGROUP_REDUCTION");
@@ -2018,100 +2105,6 @@ class ggml_webgpu_shader_lib {
         return mul_mat_fast_pipelines[key];
     }
 
-    webgpu_pipeline get_mul_mat_legacy_pipeline(const ggml_webgpu_shader_lib_context & context) {
-        ggml_webgpu_legacy_mul_mat_pipeline_key key = {};
-        key.src0_type                               = context.src0->type;
-        key.src1_type                               = context.src1->type;
-
-        auto it = mul_mat_legacy_pipelines.find(key);
-        if (it != mul_mat_legacy_pipelines.end()) {
-            return it->second;
-        }
-
-        std::vector<std::string> defines;
-        std::string              variant = "mul_mat";
-
-        switch (context.src1->type) {
-            case GGML_TYPE_F32:
-                defines.push_back("SRC1_TYPE=f32");
-                variant += "_f32";
-                break;
-            case GGML_TYPE_F16:
-                defines.push_back("SRC1_TYPE=f16");
-                variant += "_f16";
-                break;
-            default:
-                GGML_ABORT("Unsupported src1 type for mul_mat legacy shader");
-        }
-
-        const struct ggml_type_traits * src0_traits = ggml_get_type_traits(context.src0->type);
-        const char *                    src0_name   = src0_traits->type_name;
-
-        switch (context.src0->type) {
-            case GGML_TYPE_F32:
-                defines.push_back("SRC0_TYPE=f32");
-                defines.push_back("FLOAT");
-                variant += "_f32";
-                break;
-            case GGML_TYPE_F16:
-                defines.push_back("SRC0_TYPE=f16");
-                defines.push_back("FLOAT");
-                variant += "_f16";
-                break;
-            default:
-                {
-                    std::string type_upper = src0_name;
-                    std::transform(type_upper.begin(), type_upper.end(), type_upper.begin(), ::toupper);
-
-                    switch (context.src0->type) {
-                        case GGML_TYPE_Q4_0:
-                        case GGML_TYPE_Q5_0:
-                        case GGML_TYPE_Q8_0:
-                        case GGML_TYPE_Q3_K:
-                        case GGML_TYPE_Q6_K:
-                        case GGML_TYPE_IQ2_XXS:
-                        case GGML_TYPE_IQ2_XS:
-                        case GGML_TYPE_IQ2_S:
-                        case GGML_TYPE_IQ3_XXS:
-                        case GGML_TYPE_IQ3_S:
-                        case GGML_TYPE_IQ1_S:
-                        case GGML_TYPE_IQ4_NL:
-                        case GGML_TYPE_MXFP4:
-                            {
-                                // Quantized types using u32 buffers for portability.
-                                defines.push_back("SRC0_TYPE=u32");
-                                defines.push_back("U32_DEQUANT_HELPERS");
-                                break;
-                            }
-                        default:
-                            {
-                                defines.push_back(std::string("SRC0_TYPE=") + src0_name);
-                            }
-                    }
-
-                    defines.push_back("BYTE_HELPERS");
-                    defines.push_back(type_upper + "_T");
-                    defines.push_back(type_upper);
-                    defines.push_back(type_upper + "_SCALE_MIN");
-                    defines.push_back(type_upper + "_TABLES");
-                    defines.push_back(type_upper + "_GRID");
-
-                    variant += std::string("_") + src0_name;
-                    break;
-                }
-        }
-
-        auto processed = preprocessor.preprocess(wgsl_mul_mat, defines);
-
-        auto decisions     = std::make_shared<ggml_webgpu_generic_shader_decisions>();
-        decisions->wg_size = WEBGPU_MUL_MAT_WG_SIZE;
-
-        webgpu_pipeline pipeline      = ggml_webgpu_create_pipeline(device, processed, variant);
-        pipeline.context              = decisions;
-        mul_mat_legacy_pipelines[key] = pipeline;
-        return mul_mat_legacy_pipelines[key];
-    }
-
     webgpu_pipeline get_mul_mat_id_gather_pipeline(const ggml_webgpu_shader_lib_context & context) {
         auto it = mul_mat_id_gather_pipelines.find(1);
         if (it != mul_mat_id_gather_pipelines.end()) {