npm - @fugood/llama.node - Versions diffs - 0.3.2 → 0.3.3 - Mend

@fugood/llama.node 0.3.2 → 0.3.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (190) hide show

package/CMakeLists.txt +2 -0
package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/package.json +1 -1
package/src/DetokenizeWorker.cpp +1 -1
package/src/EmbeddingWorker.cpp +2 -2
package/src/LlamaCompletionWorker.cpp +8 -8
package/src/LlamaCompletionWorker.h +2 -2
package/src/LlamaContext.cpp +8 -9
package/src/TokenizeWorker.cpp +1 -1
package/src/common.hpp +4 -4
package/src/llama.cpp/.github/workflows/build.yml +43 -9
package/src/llama.cpp/.github/workflows/docker.yml +3 -0
package/src/llama.cpp/CMakeLists.txt +7 -4
package/src/llama.cpp/cmake/arm64-apple-clang.cmake +16 -0
package/src/llama.cpp/common/CMakeLists.txt +0 -2
package/src/llama.cpp/common/arg.cpp +642 -607
package/src/llama.cpp/common/arg.h +22 -22
package/src/llama.cpp/common/common.cpp +79 -281
package/src/llama.cpp/common/common.h +130 -100
package/src/llama.cpp/common/json-schema-to-grammar.cpp +1 -1
package/src/llama.cpp/common/log.cpp +50 -50
package/src/llama.cpp/common/log.h +18 -18
package/src/llama.cpp/common/ngram-cache.cpp +36 -36
package/src/llama.cpp/common/ngram-cache.h +19 -19
package/src/llama.cpp/common/sampling.cpp +116 -108
package/src/llama.cpp/common/sampling.h +20 -20
package/src/llama.cpp/docs/build.md +37 -17
package/src/llama.cpp/examples/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/batched/batched.cpp +14 -14
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +10 -11
package/src/llama.cpp/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +1 -1
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +9 -9
package/src/llama.cpp/examples/embedding/embedding.cpp +12 -12
package/src/llama.cpp/examples/eval-callback/eval-callback.cpp +8 -8
package/src/llama.cpp/examples/export-lora/export-lora.cpp +5 -5
package/src/llama.cpp/examples/gen-docs/gen-docs.cpp +7 -7
package/src/llama.cpp/examples/gritlm/gritlm.cpp +18 -18
package/src/llama.cpp/examples/imatrix/imatrix.cpp +20 -11
package/src/llama.cpp/examples/infill/infill.cpp +40 -86
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +42 -151
package/src/llama.cpp/examples/llama.android/llama/build.gradle.kts +1 -0
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +11 -14
package/src/llama.cpp/examples/llava/clip.cpp +1 -0
package/src/llama.cpp/examples/llava/llava-cli.cpp +23 -23
package/src/llama.cpp/examples/llava/llava.cpp +37 -3
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +21 -21
package/src/llama.cpp/examples/lookahead/lookahead.cpp +26 -26
package/src/llama.cpp/examples/lookup/lookup-create.cpp +7 -7
package/src/llama.cpp/examples/lookup/lookup-merge.cpp +4 -4
package/src/llama.cpp/examples/lookup/lookup-stats.cpp +14 -14
package/src/llama.cpp/examples/lookup/lookup.cpp +29 -29
package/src/llama.cpp/examples/main/main.cpp +64 -109
package/src/llama.cpp/examples/parallel/parallel.cpp +18 -19
package/src/llama.cpp/examples/passkey/passkey.cpp +14 -14
package/src/llama.cpp/examples/perplexity/perplexity.cpp +99 -120
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +10 -9
package/src/llama.cpp/examples/retrieval/retrieval.cpp +13 -13
package/src/llama.cpp/examples/rpc/rpc-server.cpp +3 -1
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +34 -17
package/src/llama.cpp/examples/server/CMakeLists.txt +4 -13
package/src/llama.cpp/examples/server/server.cpp +553 -691
package/src/llama.cpp/examples/server/utils.hpp +312 -25
package/src/llama.cpp/examples/simple/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple/simple.cpp +128 -96
package/src/llama.cpp/examples/simple-chat/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +197 -0
package/src/llama.cpp/examples/speculative/speculative.cpp +54 -51
package/src/llama.cpp/examples/tokenize/tokenize.cpp +2 -2
package/src/llama.cpp/ggml/CMakeLists.txt +15 -9
package/src/llama.cpp/ggml/include/ggml-amx.h +25 -0
package/src/llama.cpp/ggml/include/ggml-backend.h +46 -33
package/src/llama.cpp/ggml/include/ggml-blas.h +5 -3
package/src/llama.cpp/ggml/include/ggml-cann.h +9 -7
package/src/llama.cpp/ggml/include/ggml-cpp.h +38 -0
package/src/llama.cpp/ggml/include/ggml-cpu.h +177 -0
package/src/llama.cpp/ggml/include/ggml-cuda.h +12 -12
package/src/llama.cpp/ggml/include/ggml-kompute.h +7 -3
package/src/llama.cpp/ggml/include/ggml-metal.h +11 -7
package/src/llama.cpp/ggml/include/ggml-opt.h +216 -0
package/src/llama.cpp/ggml/include/ggml-rpc.h +9 -5
package/src/llama.cpp/ggml/include/ggml-sycl.h +18 -11
package/src/llama.cpp/ggml/include/ggml-vulkan.h +10 -8
package/src/llama.cpp/ggml/include/ggml.h +53 -393
package/src/llama.cpp/ggml/src/CMakeLists.txt +66 -1149
package/src/llama.cpp/ggml/src/ggml-aarch64.c +46 -3126
package/src/llama.cpp/ggml/src/ggml-aarch64.h +0 -20
package/src/llama.cpp/ggml/src/ggml-alloc.c +23 -27
package/src/llama.cpp/ggml/src/ggml-amx/CMakeLists.txt +107 -0
package/src/llama.cpp/ggml/src/ggml-amx/common.h +94 -0
package/src/llama.cpp/ggml/src/ggml-amx/ggml-amx.cpp +446 -0
package/src/llama.cpp/ggml/src/ggml-amx/mmq.cpp +2510 -0
package/src/llama.cpp/ggml/src/ggml-amx/mmq.h +17 -0
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +6 -25
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +195 -0
package/src/llama.cpp/ggml/src/ggml-backend.cpp +303 -864
package/src/llama.cpp/ggml/src/ggml-blas/CMakeLists.txt +91 -0
package/src/llama.cpp/ggml/src/{ggml-blas.cpp → ggml-blas/ggml-blas.cpp} +213 -65
package/src/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +46 -0
package/src/llama.cpp/ggml/src/{ggml-cann.cpp → ggml-cann/ggml-cann.cpp} +255 -149
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +261 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.c +3560 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +30 -0
package/src/llama.cpp/ggml/src/{ggml-cpu-impl.h → ggml-cpu/ggml-cpu-impl.h} +0 -243
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +10822 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.h +63 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +13970 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +663 -0
package/src/llama.cpp/ggml/src/{llamafile → ggml-cpu/llamafile}/sgemm.cpp +667 -1
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +155 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +106 -0
package/src/llama.cpp/ggml/src/ggml-impl.h +366 -16
package/src/llama.cpp/ggml/src/ggml-kompute/CMakeLists.txt +162 -0
package/src/llama.cpp/ggml/src/{ggml-kompute.cpp → ggml-kompute/ggml-kompute.cpp} +238 -72
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +108 -0
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +249 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +100 -0
package/src/llama.cpp/ggml/src/ggml-opt.cpp +867 -0
package/src/llama.cpp/ggml/src/ggml-quants.c +187 -10692
package/src/llama.cpp/ggml/src/ggml-quants.h +78 -125
package/src/llama.cpp/ggml/src/ggml-rpc/CMakeLists.txt +11 -0
package/src/llama.cpp/ggml/src/{ggml-rpc.cpp → ggml-rpc/ggml-rpc.cpp} +475 -300
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +81 -0
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +3 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +40 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +258 -0
package/src/llama.cpp/ggml/src/ggml-sycl/concat.cpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +2 -22
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +1011 -0
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.hpp +76 -0
package/src/llama.cpp/ggml/src/{ggml-sycl.cpp → ggml-sycl/ggml-sycl.cpp} +3584 -4142
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +69 -67
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +3 -3
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +56 -0
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.hpp +11 -0
package/src/llama.cpp/ggml/src/ggml-sycl/presets.hpp +6 -0
package/src/llama.cpp/ggml/src/ggml-sycl/vecdotq.hpp +4 -4
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +138 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +10 -0
package/src/llama.cpp/ggml/src/ggml-threading.cpp +12 -0
package/src/llama.cpp/ggml/src/ggml-threading.h +12 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +78 -0
package/src/llama.cpp/ggml/src/{ggml-vulkan.cpp → ggml-vulkan/ggml-vulkan.cpp} +555 -623
package/src/llama.cpp/ggml/src/{vulkan-shaders → ggml-vulkan/vulkan-shaders}/vulkan-shaders-gen.cpp +125 -206
package/src/llama.cpp/ggml/src/ggml.c +4032 -19890
package/src/llama.cpp/include/llama.h +67 -33
package/src/llama.cpp/pocs/vdot/q8dot.cpp +4 -3
package/src/llama.cpp/pocs/vdot/vdot.cpp +8 -7
package/src/llama.cpp/src/CMakeLists.txt +2 -1
package/src/llama.cpp/src/llama-sampling.cpp +745 -105
package/src/llama.cpp/src/llama-sampling.h +21 -2
package/src/llama.cpp/src/llama-vocab.cpp +49 -9
package/src/llama.cpp/src/llama-vocab.h +35 -11
package/src/llama.cpp/src/llama.cpp +2636 -2406
package/src/llama.cpp/src/unicode-data.cpp +2 -2
package/src/llama.cpp/tests/CMakeLists.txt +1 -2
package/src/llama.cpp/tests/test-arg-parser.cpp +14 -14
package/src/llama.cpp/tests/test-backend-ops.cpp +185 -60
package/src/llama.cpp/tests/test-barrier.cpp +1 -0
package/src/llama.cpp/tests/test-chat-template.cpp +9 -5
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +17 -4
package/src/llama.cpp/tests/test-log.cpp +2 -2
package/src/llama.cpp/tests/test-opt.cpp +853 -142
package/src/llama.cpp/tests/test-quantize-fns.cpp +22 -19
package/src/llama.cpp/tests/test-quantize-perf.cpp +16 -14
package/src/llama.cpp/tests/test-rope.cpp +1 -0
package/src/llama.cpp/tests/test-sampling.cpp +162 -137
package/src/llama.cpp/tests/test-tokenizer-0.cpp +7 -7
package/src/llama.cpp/tests/test-tokenizer-1-bpe.cpp +5 -5
package/src/llama.cpp/tests/test-tokenizer-1-spm.cpp +5 -5
package/src/llama.cpp/common/train.cpp +0 -1515
package/src/llama.cpp/common/train.h +0 -233
package/src/llama.cpp/examples/baby-llama/CMakeLists.txt +0 -5
package/src/llama.cpp/examples/baby-llama/baby-llama.cpp +0 -1639
package/src/llama.cpp/tests/test-grad0.cpp +0 -1683
/package/src/llama.cpp/ggml/{cmake → src/ggml-cpu/cmake}/FindSIMD.cmake +0 -0
/package/src/llama.cpp/ggml/src/{llamafile → ggml-cpu/llamafile}/sgemm.h +0 -0
/package/src/llama.cpp/ggml/src/{vulkan-shaders → ggml-vulkan/vulkan-shaders}/CMakeLists.txt +0 -0

package/src/llama.cpp/ggml/src/{ggml-vulkan.cpp → ggml-vulkan/ggml-vulkan.cpp} RENAMED Viewed

@@ -106,6 +106,15 @@ struct vk_matmul_pipeline_struct {
 typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
+struct vk_matmul_pipeline2 {
+    vk_matmul_pipeline2() {
+        f16acc = std::make_shared<vk_matmul_pipeline_struct>();
+        f32acc = std::make_shared<vk_matmul_pipeline_struct>();
+    }
+    vk_matmul_pipeline f32acc;
+    vk_matmul_pipeline f16acc;
+};
 struct vk_device_struct;
 typedef std::shared_ptr<vk_device_struct> vk_device;
 typedef std::weak_ptr<vk_device_struct> vk_device_ref;
@@ -161,11 +170,11 @@ struct vk_device_struct {
     vk_matmul_pipeline pipeline_matmul_f32;
     vk_matmul_pipeline pipeline_matmul_f32_f16;
-    vk_matmul_pipeline pipeline_matmul_f16;
-    vk_matmul_pipeline pipeline_matmul_f16_f32;
+    vk_matmul_pipeline2 pipeline_matmul_f16;
+    vk_matmul_pipeline2 pipeline_matmul_f16_f32;
     vk_pipeline pipeline_matmul_split_k_reduce;
-    vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
+    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
     vk_matmul_pipeline pipeline_matmul_id_f32;
     vk_matmul_pipeline pipeline_matmul_id_f16;
@@ -183,9 +192,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_acc_f32;
-    vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
-    vk_pipeline pipeline_mul_f32;
-    vk_pipeline pipeline_div_f32;
+    vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat;
+    vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat;
+    vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat;
+    vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat;
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
     vk_pipeline pipeline_upscale_f32;
     vk_pipeline pipeline_scale_f32;
@@ -196,6 +206,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_repeat_f32;
     vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
+    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
@@ -213,6 +224,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_sum_rows_f32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
+    vk_pipeline pipeline_pool2d_f32;
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
     std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
@@ -403,6 +415,17 @@ struct vk_op_timestep_embedding_push_constants {
     uint32_t max_period;
 };
+struct vk_op_pool2d_push_constants {
+    uint32_t IW; uint32_t IH;
+    uint32_t OW; uint32_t OH;
+    uint32_t OC;
+    uint32_t pelements;
+    uint32_t op;
+    int32_t k0; int32_t k1;
+    int32_t s0; int32_t s1;
+    int32_t p0; int32_t p1;
+};
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -710,6 +733,12 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         std::lock_guard<std::mutex> guard(compile_count_mutex);
         assert(compile_count > 0);
         compile_count--;
+        // "Progress bar" for shader compiles
+        static uint32_t total_compile_count = 0;
+        if ((total_compile_count++ % 10) == 0) {
+            std::cerr << ".";
+        }
     }
     compile_count_cond.notify_all();
 }
@@ -1035,7 +1064,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
         return buf;
     }
-    buf->size = size;
     vk::BufferCreateInfo buffer_create_info{
         vk::BufferCreateFlags(),
         size,
@@ -1063,7 +1091,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
     if (memory_type_index == UINT32_MAX) {
         device->device.destroyBuffer(buf->buffer);
-        buf->size = 0;
         throw vk::OutOfDeviceMemoryError("No suitable memory type found");
     }
@@ -1080,13 +1107,11 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
             }
             catch (const vk::SystemError& e) {
                 device->device.destroyBuffer(buf->buffer);
-                buf->size = 0;
                 throw e;
             }
         } else {
             // Out of Host/Device memory, clean up buffer
             device->device.destroyBuffer(buf->buffer);
-            buf->size = 0;
             throw e;
         }
     }
@@ -1099,6 +1124,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
     device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0);
     buf->device = device;
+    buf->size = size;
 #ifdef GGML_VULKAN_MEMORY_DEBUG
     device->memory_logger->log_allocation(buf, size);
@@ -1191,38 +1217,31 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 static void ggml_vk_load_shaders(vk_device& device) {
     VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
+    std::cerr << "ggml_vulkan: Compiling shaders";
     // mulmat
-    std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_m = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_s = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
+    std::vector<uint32_t> l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq;
+    std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms;
+    uint32_t l_align, m_align, s_align;
-    std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_mmq_m = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_mmq_s = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
+    l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    m_warptile = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    s_warptile = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
-    std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
-    std::array<uint32_t, 3> m_wg_denoms = { 64,  64, 1 };
-    std::array<uint32_t, 3> s_wg_denoms = { 32,  32, 1 };
+    l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    m_warptile_mmq = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    s_warptile_mmq = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
-    uint32_t l_align = 128;
-    uint32_t m_align =  64;
-    uint32_t s_align =  32;
+    l_wg_denoms = {128, 128, 1 };
+    m_wg_denoms = { 64,  64, 1 };
+    s_wg_denoms = { 32,  32, 1 };
+    l_align = 128;
+    m_align =  64;
+    s_align =  32;
     device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
@@ -1240,7 +1259,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1254,455 +1273,140 @@ static void ggml_vk_load_shaders(vk_device& device) {
     };
     if (device->fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+#undef CREATE_MM
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+#undef CREATE_MM
     }
     // mul mat vec
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1750,13 +1454,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16_norepeat, "add_f16_f32_f16_norepeat", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32_norepeat, "div_f32_norepeat", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -1803,9 +1515,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
     for (auto &c : compiles) {
         c.wait();
     }
+    std::cerr << "Done!" << std::endl;
 }
 static vk_device ggml_vk_get_device(size_t idx) {
@@ -1941,7 +1656,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (device->fp16) {
             device_extensions.push_back("VK_KHR_shader_float16_int8");
         }
-        device->name = device->properties.deviceName.data();
+        device->name = GGML_VK_NAME + std::to_string(idx);
         device_create_info = {
             vk::DeviceCreateFlags(),
@@ -1968,7 +1683,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->buffer_type = {
             /* .iface    = */ ggml_backend_vk_buffer_type_interface,
-            /* .device   = */ nullptr,
+            /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), idx),
             /* .context  = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
         };
@@ -2049,7 +1764,8 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     fp16 = fp16 && vk12_features.shaderFloat16;
     std::string device_name = props2.properties.deviceName.data();
-    std::cerr << GGML_VK_NAME << idx << ": " << device_name << " (" << driver_props.driverName << ") | uma: " << uma << " | fp16: " << fp16 << " | warp size: " << subgroup_size << std::endl;
+    GGML_LOG_DEBUG("ggml_vulkan: %d = %s (%s) | uma: %d | fp16: %d | warp size: %d\n",
+              idx, device_name.c_str(), driver_props.driverName, uma, fp16, subgroup_size);
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
         std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
@@ -2107,8 +1823,7 @@ void ggml_vk_instance_init() {
         };
         validation_features.setPNext(nullptr);
         instance_create_info.setPNext(&validation_features);
-        std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
+        GGML_LOG_DEBUG("ggml_vulkan: Validation layers enabled\n");
     }
     vk_instance.instance = vk::createInstance(instance_create_info);
@@ -2222,8 +1937,8 @@ void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(0);
         }
     }
+    GGML_LOG_DEBUG("ggml_vulkan: Found %d Vulkan devices:\n", vk_instance.device_indices.size());
-    std::cerr << "ggml_vulkan: Found " << vk_instance.device_indices.size() << " Vulkan devices:" << std::endl;
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
         ggml_vk_print_gpu_info(i);
@@ -2288,10 +2003,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
         return ctx->device->pipeline_matmul_f32_f16;
     }
     if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-        return ctx->device->pipeline_matmul_f16_f32;
+        return ctx->device->pipeline_matmul_f16_f32.f32acc;
     }
     if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-        return ctx->device->pipeline_matmul_f16;
+        return ctx->device->pipeline_matmul_f16.f32acc;
     }
     if (src1_type != GGML_TYPE_F32) {
@@ -2315,7 +2030,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
             return nullptr;
     }
-    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
+    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -3050,18 +2765,34 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
-static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) {
-    if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
-        return ctx->device->pipeline_cpy_f32_f32;
+static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to) {
+    // Choose "contiguous copy" shader if src/dst are contiguous
+    bool contig = ggml_is_contiguous(src) && (!dst || ggml_is_contiguous(dst));
+    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f32_f32;
+        } else {
+            return ctx->device->pipeline_cpy_f32_f32;
+        }
     }
-    if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
-        return ctx->device->pipeline_cpy_f32_f16;
+    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f32_f16;
+        } else {
+            return ctx->device->pipeline_cpy_f32_f16;
+        }
     }
-    if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
-        return ctx->device->pipeline_cpy_f16_f16;
+    if (src->type == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f16_f16;
+        } else {
+            return ctx->device->pipeline_cpy_f16_f16;
+        }
     }
-    std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
+    std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
     GGML_ABORT("fatal error");
 }
@@ -3071,6 +2802,15 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     const int tensor_type_size = ggml_type_size(tensor->type);
     const uint32_t ne = ggml_nelements(tensor);
+    std::array<uint32_t, 3> elements;
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
     const vk_op_unary_push_constants pc = {
         (uint32_t)ne,
@@ -3080,7 +2820,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
         0.0f, 0.0f,
     };
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
 }
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -3136,7 +2876,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
-    if (mmp == nullptr) {
+    if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
         mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16);
     }
@@ -3165,12 +2905,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     vk_pipeline to_fp16_vk_1 = nullptr;
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -3350,10 +3090,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -3619,9 +3359,19 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")");
-    if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) {
+    if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1 &&
+        // detect 0213 permutation, and batch size of 1
+        src0->nb[0] <= src0->nb[2] &&
+        src0->nb[2] <= src0->nb[1] &&
+        src0->nb[1] <= src0->nb[3] &&
+        src1->nb[0] <= src1->nb[2] &&
+        src1->nb[2] <= src1->nb[1] &&
+        src1->nb[1] <= src1->nb[3] &&
+        src0->ne[3] == 1 &&
+        src1->ne[3] == 1) {
         ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
-    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1) {
+    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
+               !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
         ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
     } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -3697,7 +3447,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
-    if (mmp == nullptr) {
+    if (qx_needs_dequant) {
         GGML_ABORT("fatal error");
     }
@@ -3724,12 +3474,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_1 = nullptr;
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -3917,10 +3667,10 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, src1->type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -4057,20 +3807,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_ADD:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_add_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f32_norepeat : ctx->device->pipeline_add_f32;
         }
         if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
-            return ctx->device->pipeline_add_f16_f32_f16;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16;
         }
         return nullptr;
     case GGML_OP_MUL:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_mul_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32;
         }
         return nullptr;
     case GGML_OP_DIV:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_div_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_f32_norepeat : ctx->device->pipeline_div_f32;
         }
         return nullptr;
     case GGML_OP_CONCAT:
@@ -4127,7 +3877,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
-        return ggml_vk_get_cpy_pipeline(ctx, src0->type, dst->type);
+        return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
     case GGML_OP_NORM:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_norm_f32;
@@ -4234,6 +3984,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_timestep_embedding_f32;
         }
         return nullptr;
+    case GGML_OP_POOL_2D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pool2d_f32;
+        }
+        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -4255,7 +4010,6 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_DIV:
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
-    case GGML_OP_SCALE:
     case GGML_OP_SQR:
     case GGML_OP_SIN:
     case GGML_OP_COS:
@@ -4454,7 +4208,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             const uint32_t OH = is_2D ? dst->ne[2] : 1;
             const uint32_t OW =         dst->ne[1];
-            const uint32_t batch = src1->ne[3];
+            const uint32_t batch = src1->ne[is_2D ? 3 : 2];
             elements = { OW * KW * KH, OH, batch * IC };
         } break;
@@ -4464,6 +4218,14 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             uint32_t half_ceil = (dim + 1) / 2;
             elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
         } break;
+    case GGML_OP_POOL_2D:
+        {
+            const uint32_t N = dst->ne[3];
+            const uint32_t OC = dst->ne[2];
+            const uint32_t OH = dst->ne[1];
+            const uint32_t OW = dst->ne[0];
+            elements = { N * OC * OH * OW, 1, 1};
+        } break;
     case GGML_OP_ADD:
     case GGML_OP_DIV:
     case GGML_OP_MUL:
@@ -4891,7 +4653,7 @@ static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, co
     const uint32_t OW =         dst->ne[1];
     const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const uint32_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+    const uint32_t batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
     const uint32_t pelements = OW * KW * KH;
@@ -4914,6 +4676,34 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
     }, dryrun);
 }
+static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
+    const int32_t k1 = dst->op_params[1];
+    const int32_t k0 = dst->op_params[2];
+    const int32_t s1 = dst->op_params[3];
+    const int32_t s0 = dst->op_params[4];
+    const int32_t p1 = dst->op_params[5];
+    const int32_t p0 = dst->op_params[6];
+    const uint32_t IH = src0->ne[1];
+    const uint32_t IW = src0->ne[0];
+    const uint32_t N = dst->ne[3];
+    const uint32_t OC = dst->ne[2];
+    const uint32_t OH = dst->ne[1];
+    const uint32_t OW = dst->ne[0];
+    const uint32_t parallel_elements = N * OC * OH * OW;
+    ggml_vk_op_f32<vk_op_pool2d_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_POOL_2D, {
+        IW, IH, OW, OH, OC,
+        parallel_elements,
+        op,
+        k0, k1, s0, s1, p0, p1,
+    }, dryrun);
+}
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const float * op_params = (const float *)dst->op_params;
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
@@ -4970,10 +4760,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_s;
             shname = "F32_F16_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_s;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_s;
             shname = "F16_F32_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_s;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_s;
             shname = "F16_ALIGNED_S";
         } else {
             GGML_ABORT("fatal error");
@@ -4986,10 +4776,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_m;
             shname = "F32_F16_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_m;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_m;
             shname = "F16_F32_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_m;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_m;
             shname = "F16_ALIGNED_M";
         } else {
             GGML_ABORT("fatal error");
@@ -5002,10 +4792,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_l;
             shname = "F32_F16_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_l;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_l;
             shname = "F16_F32_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_l;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_l;
             shname = "F16_ALIGNED_L";
         } else {
             GGML_ABORT("fatal error");
@@ -5025,10 +4815,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->s;
                 shname = "F32_F16_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->s;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->s;
                 shname = "F16_F32_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->s;
+                p = ctx->device->pipeline_matmul_f16.f32acc->s;
                 shname = "F16_S";
             }
         } else if (shader_size == 1) {
@@ -5039,10 +4829,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->m;
                 shname = "F32_F16_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->m;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->m;
                 shname = "F16_F32_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->m;
+                p = ctx->device->pipeline_matmul_f16.f32acc->m;
                 shname = "F16_M";
             }
         } else if (shader_size == 2) {
@@ -5053,10 +4843,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->l;
                 shname = "F32_F16_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->l;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->l;
                 shname = "F16_F32_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->l;
+                p = ctx->device->pipeline_matmul_f16.f32acc->l;
                 shname = "F16_L";
             }
         }
@@ -5287,9 +5077,9 @@ static void ggml_vk_dequantize_data(const void * from, float * to, size_t ne, gg
         return;
     }
-    ggml_type_traits_t tt = ggml_internal_get_type_traits(quant);
+    const auto * tt = ggml_get_type_traits(quant);
-    ggml_to_float_t dequant_fn = tt.to_float;
+    ggml_to_float_t dequant_fn = tt->to_float;
     dequant_fn(from, to, ne);
 }
@@ -5385,13 +5175,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     vk_pipeline p;
     std::string shname;
     if (shader_size == 0) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
     } else if (shader_size == 1) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
     } else if (shader_size == 2) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l;
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
     } else {
         GGML_ASSERT(0);
@@ -5401,13 +5191,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     if (k != kpad) {
         if (shader_size == 0) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
             shname = std::string(ggml_type_name(quant)) + "_S";
         } else if (shader_size == 1) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
             shname = std::string(ggml_type_name(quant)) + "_M";
         } else if (shader_size == 2) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l;
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
             shname = std::string(ggml_type_name(quant)) + "_L";
         } else {
             GGML_ASSERT(0);
@@ -5792,6 +5582,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_SUM_ROWS:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_POOL_2D:
     case GGML_OP_LEAKY_RELU:
         break;
     default:
@@ -5927,6 +5718,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_TIMESTEP_EMBEDDING:
         ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
+        break;
+    case GGML_OP_POOL_2D:
+        ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
         break;
     case GGML_OP_LEAKY_RELU:
         ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
@@ -6018,6 +5813,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_SUM_ROWS:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_POOL_2D:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_REPEAT:
         buf = tensor->buffer;
@@ -6186,13 +5982,8 @@ static void ggml_vk_get_device_description(int device, char * description, size_
 // device backend
-static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
-    ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
-    return ctx->name.c_str();
-}
 static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
-    return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
+    return buffer->buft->iface.get_name == ggml_backend_vk_buffer_type_name;
 }
 static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@@ -6256,7 +6047,6 @@ static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t v
 }
 static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
-    /* .get_name        = */ ggml_backend_vk_buffer_get_name,
     /* .free_buffer     = */ ggml_backend_vk_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_vk_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_vk_buffer_init_tensor,
@@ -6352,7 +6142,6 @@ static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
     buffer->buft = buft;
-    buffer->iface.get_name = ggml_backend_vk_host_buffer_name;
     buffer->iface.free_buffer = ggml_backend_vk_host_buffer_free_buffer;
     return buffer;
@@ -6378,7 +6167,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
-        /* .device   = */ nullptr,
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), 0),
         /* .context  = */ nullptr,
     };
@@ -6581,9 +6370,132 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     UNUSED(backend);
 }
-static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
-    // ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
+// TODO: enable async and synchronize
+static ggml_backend_i ggml_backend_vk_interface = {
+    /* .get_name                = */ ggml_backend_vk_name,
+    /* .free                    = */ ggml_backend_vk_free,
+    /* .set_tensor_async        = */ NULL,  // ggml_backend_vk_set_tensor_async,
+    /* .get_tensor_async        = */ NULL,  // ggml_backend_vk_get_tensor_async,
+    /* .cpy_tensor_async        = */ NULL,  // ggml_backend_vk_cpy_tensor_async,
+    /* .synchronize             = */ NULL,  // ggml_backend_vk_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_vk_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+static ggml_guid_t ggml_backend_vk_guid() {
+    static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
+    return &guid;
+}
+ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
+    VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
+    ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
+    ggml_vk_init(ctx, dev_num);
+    ggml_backend_t vk_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_vk_guid(),
+        /* .interface = */ ggml_backend_vk_interface,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
+        /* .context   = */ ctx,
+    };
+    return vk_backend;
+}
+bool ggml_backend_is_vk(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
+}
+int ggml_backend_vk_get_device_count() {
+    return ggml_vk_get_device_count();
+}
+void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    int dev_idx = vk_instance.device_indices[device];
+    ggml_vk_get_device_description(dev_idx, description, description_size);
+}
+void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
+        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
+            *total = heap.size;
+            *free = heap.size;
+            break;
+        }
+    }
+}
+//////////////////////////
+struct ggml_backend_vk_device_context {
+    size_t device;
+    std::string name;
+    std::string description;
+};
+static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+static void ggml_backend_vk_device_get_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)device->context;
+    ggml_backend_vk_get_device_memory(ctx->device, free, total);
+}
+static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+    return ggml_backend_vk_buffer_type(ctx->device);
+}
+static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    UNUSED(dev);
+    return ggml_backend_vk_host_buffer_type();
+}
+static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
+    UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_vk_device_get_name(dev);
+    props->description = ggml_backend_vk_device_get_description(dev);
+    props->type        = ggml_backend_vk_device_get_type(dev);
+    ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ true,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
+    UNUSED(params);
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+    return ggml_backend_vk_init(ctx->device);
+}
+static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -6630,6 +6542,11 @@ static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tenso
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
+                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) ||
+                    !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) {
+                    return false;
+                }
                 return true;
             } break;
         case GGML_OP_GET_ROWS:
@@ -6695,103 +6612,108 @@ static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tenso
         case GGML_OP_SUM_ROWS:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_POOL_2D:
         case GGML_OP_LEAKY_RELU:
             return true;
         default:
             return false;
     }
-    UNUSED(backend);
+    UNUSED(dev);
 }
-static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-    UNUSED(backend);
-}
-static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+static bool ggml_backend_vk_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
     if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
         return false;
     }
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
     ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
-    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
-    return buft_ctx->device == ctx->device;
+    return buft_ctx->device->idx == ctx->device;
 }
-// TODO: enable async and synchronize
-static ggml_backend_i ggml_backend_vk_interface = {
-    /* .get_name                = */ ggml_backend_vk_name,
-    /* .free                    = */ ggml_backend_vk_free,
-    /* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
-    /* .set_tensor_async        = */ NULL,  // ggml_backend_vk_set_tensor_async,
-    /* .get_tensor_async        = */ NULL,  // ggml_backend_vk_get_tensor_async,
-    /* .cpy_tensor_async        = */ NULL,  // ggml_backend_vk_cpy_tensor_async,
-    /* .synchronize             = */ NULL,  // ggml_backend_vk_synchronize,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_vk_graph_compute,
-    /* .supports_op             = */ ggml_backend_vk_supports_op,
-    /* .supports_buft           = */ ggml_backend_vk_supports_buft,
-    /* .offload_op              = */ ggml_backend_vk_offload_op,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-static ggml_guid_t ggml_backend_vk_guid() {
-    static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
-    return &guid;
-}
-ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
-    VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
+static bool ggml_backend_vk_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    const int min_batch_size = 32;
-    ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
-    ggml_vk_init(ctx, dev_num);
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-    ggml_backend_t vk_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_vk_guid(),
-        /* .interface = */ ggml_backend_vk_interface,
-        /* .device    = */ nullptr,
-        /* .context   = */ ctx,
-    };
+    UNUSED(dev);
+}
+static const struct ggml_backend_device_i ggml_backend_vk_device_i = {
+    /* .get_name             = */ ggml_backend_vk_device_get_name,
+    /* .get_description      = */ ggml_backend_vk_device_get_description,
+    /* .get_memory           = */ ggml_backend_vk_device_get_memory,
+    /* .get_type             = */ ggml_backend_vk_device_get_type,
+    /* .get_props            = */ ggml_backend_vk_device_get_props,
+    /* .init_backend         = */ ggml_backend_vk_device_init,
+    /* .get_buffer_type      = */ ggml_backend_vk_device_get_buffer_type,
+    /* .get_host_buffer_type = */ ggml_backend_vk_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_vk_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_vk_device_supports_buft,
+    /* .offload_op           = */ ggml_backend_vk_device_offload_op,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
-    return vk_backend;
+static const char * ggml_backend_vk_reg_get_name(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+    return GGML_VK_NAME;
 }
-bool ggml_backend_is_vk(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
+static size_t ggml_backend_vk_reg_get_device_count(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+    return ggml_backend_vk_get_device_count();
 }
-int ggml_backend_vk_get_device_count() {
-    return ggml_vk_get_device_count();
-}
+static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg, size_t device) {
+    static std::vector<ggml_backend_dev_t> devices;
-void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
-    ggml_vk_get_device_description(device, description, description_size);
-}
+    static bool initialized = false;
-void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
-    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
+                ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
+                char desc[256];
+                ggml_backend_vk_get_device_description(i, desc, sizeof(desc));
+                ctx->device = i;
+                ctx->name = GGML_VK_NAME + std::to_string(i);
+                ctx->description = desc;
+                devices.push_back(new ggml_backend_device {
+                    /* .iface   = */ ggml_backend_vk_device_i,
+                    /* .reg     = */ reg,
+                    /* .context = */ ctx,
+                });
+            }
+            initialized = true;
+        }
+    }
-    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+    GGML_ASSERT(device < devices.size());
+    return devices[device];
+}
-    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
+    /* .get_name         = */ ggml_backend_vk_reg_get_name,
+    /* .get_device_count = */ ggml_backend_vk_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_vk_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
-    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
-        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-            *total = heap.size;
-            *free = heap.size;
-            break;
-        }
-    }
+ggml_backend_reg_t ggml_backend_vk_reg() {
+    static ggml_backend_reg reg = {
+        /* .iface   = */ ggml_backend_vk_reg_i,
+        /* .context = */ nullptr,
+    };
+    return &reg;
 }
 // Extension availability
@@ -7204,6 +7126,16 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t dim = tensor->op_params[0];
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
+    } else if (tensor->op == GGML_OP_POOL_2D) {
+        enum ggml_op_pool op = static_cast<ggml_op_pool>(dst->op_params[0]);
+        const int32_t k0 = tensor->op_params[1];
+        const int32_t k1 = tensor->op_params[2];
+        const int32_t s0 = tensor->op_params[3];
+        const int32_t s1 = tensor->op_params[4];
+        const int32_t p0 = tensor->op_params[5];
+        const int32_t p1 = tensor->op_params[6];
+        tensor_clone = ggml_pool_2d(ggml_ctx, src0_clone, op, k0, k1, s0, s1, p0, p1);
     } else if (tensor->op == GGML_OP_LEAKY_RELU) {
         const float * op_params = (const float *)tensor->op_params;
         tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);