npm - @fugood/llama.node - Versions diffs - 0.3.7 → 0.3.8 - Mend

@fugood/llama.node 0.3.7 → 0.3.8

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 (186) hide show

package/README.md +17 -2
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-cuda/arm64/llama-node.node +0 -0
package/bin/linux-cuda/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/lib/binding.ts +2 -0
package/lib/index.js +16 -1
package/lib/index.ts +16 -0
package/package.json +1 -1
package/src/EmbeddingWorker.cpp +4 -3
package/src/LlamaCompletionWorker.cpp +4 -2
package/src/LlamaContext.cpp +61 -6
package/src/LlamaContext.h +1 -0
package/src/common.hpp +6 -11
package/src/llama.cpp/.github/workflows/build.yml +19 -17
package/src/llama.cpp/.github/workflows/docker.yml +77 -30
package/src/llama.cpp/.github/workflows/editorconfig.yml +3 -1
package/src/llama.cpp/.github/workflows/server.yml +22 -3
package/src/llama.cpp/CMakeLists.txt +49 -24
package/src/llama.cpp/common/arg.cpp +82 -26
package/src/llama.cpp/common/arg.h +3 -0
package/src/llama.cpp/common/common.cpp +192 -72
package/src/llama.cpp/common/common.h +51 -18
package/src/llama.cpp/common/ngram-cache.cpp +12 -12
package/src/llama.cpp/common/ngram-cache.h +2 -2
package/src/llama.cpp/common/sampling.cpp +11 -6
package/src/llama.cpp/common/speculative.cpp +18 -15
package/src/llama.cpp/docs/build.md +2 -0
package/src/llama.cpp/examples/batched/batched.cpp +9 -7
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +3 -3
package/src/llama.cpp/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +10 -8
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +11 -8
package/src/llama.cpp/examples/cvector-generator/mean.hpp +1 -1
package/src/llama.cpp/examples/cvector-generator/pca.hpp +1 -1
package/src/llama.cpp/examples/embedding/embedding.cpp +8 -7
package/src/llama.cpp/examples/eval-callback/eval-callback.cpp +7 -6
package/src/llama.cpp/examples/export-lora/export-lora.cpp +8 -7
package/src/llama.cpp/examples/gguf/gguf.cpp +10 -6
package/src/llama.cpp/examples/gguf-hash/gguf-hash.cpp +1 -0
package/src/llama.cpp/examples/gguf-split/gguf-split.cpp +8 -7
package/src/llama.cpp/examples/gritlm/gritlm.cpp +13 -10
package/src/llama.cpp/examples/imatrix/imatrix.cpp +13 -12
package/src/llama.cpp/examples/infill/infill.cpp +23 -24
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +44 -13
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +11 -6
package/src/llama.cpp/examples/llava/clip.cpp +4 -2
package/src/llama.cpp/examples/llava/llava-cli.cpp +9 -6
package/src/llama.cpp/examples/llava/llava.cpp +2 -2
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +8 -4
package/src/llama.cpp/examples/llava/qwen2vl-cli.cpp +11 -8
package/src/llama.cpp/examples/lookahead/lookahead.cpp +6 -7
package/src/llama.cpp/examples/lookup/lookup-create.cpp +4 -9
package/src/llama.cpp/examples/lookup/lookup-stats.cpp +3 -7
package/src/llama.cpp/examples/lookup/lookup.cpp +5 -6
package/src/llama.cpp/examples/main/main.cpp +51 -29
package/src/llama.cpp/examples/parallel/parallel.cpp +5 -6
package/src/llama.cpp/examples/passkey/passkey.cpp +7 -5
package/src/llama.cpp/examples/perplexity/perplexity.cpp +37 -23
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +12 -14
package/src/llama.cpp/examples/retrieval/retrieval.cpp +8 -8
package/src/llama.cpp/examples/rpc/rpc-server.cpp +12 -0
package/src/llama.cpp/examples/run/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.cpp +1351 -0
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.h +114 -0
package/src/llama.cpp/examples/run/run.cpp +175 -61
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +4 -25
package/src/llama.cpp/examples/server/CMakeLists.txt +1 -0
package/src/llama.cpp/examples/server/httplib.h +1295 -409
package/src/llama.cpp/examples/server/server.cpp +387 -181
package/src/llama.cpp/examples/server/tests/requirements.txt +1 -0
package/src/llama.cpp/examples/server/utils.hpp +170 -58
package/src/llama.cpp/examples/simple/simple.cpp +9 -8
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +16 -12
package/src/llama.cpp/examples/speculative/speculative.cpp +22 -23
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +8 -12
package/src/llama.cpp/examples/tokenize/tokenize.cpp +17 -5
package/src/llama.cpp/examples/tts/tts.cpp +64 -23
package/src/llama.cpp/ggml/CMakeLists.txt +5 -21
package/src/llama.cpp/ggml/include/ggml-backend.h +2 -0
package/src/llama.cpp/ggml/include/ggml-cpp.h +1 -0
package/src/llama.cpp/ggml/include/ggml.h +36 -145
package/src/llama.cpp/ggml/include/gguf.h +202 -0
package/src/llama.cpp/ggml/src/CMakeLists.txt +6 -3
package/src/llama.cpp/ggml/src/ggml-alloc.c +5 -0
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +0 -1
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +79 -49
package/src/llama.cpp/ggml/src/ggml-backend.cpp +5 -2
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +33 -23
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +57 -72
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +87 -2
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +335 -66
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +10 -2
package/src/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +1090 -378
package/src/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.h +2 -2
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/cuda.h +1 -0
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +3 -0
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +3 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +3 -1
package/src/llama.cpp/ggml/src/ggml-impl.h +11 -16
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +16 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +6 -6
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +154 -35
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +9 -3
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +18 -0
package/src/llama.cpp/ggml/src/ggml-sycl/concat.cpp +3 -2
package/src/llama.cpp/ggml/src/ggml-sycl/concat.hpp +1 -2
package/src/llama.cpp/ggml/src/ggml-sycl/conv.cpp +3 -2
package/src/llama.cpp/ggml/src/ggml-sycl/conv.hpp +1 -2
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +40 -95
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +48 -48
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.hpp +24 -24
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +238 -164
package/src/llama.cpp/ggml/src/ggml-sycl/gla.cpp +105 -0
package/src/llama.cpp/ggml/src/ggml-sycl/gla.hpp +8 -0
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +3 -3
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.hpp +1 -2
package/src/llama.cpp/ggml/src/ggml-sycl/tsembd.cpp +3 -2
package/src/llama.cpp/ggml/src/ggml-sycl/tsembd.hpp +1 -2
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +7 -5
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +1 -2
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +74 -4
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +314 -116
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -2
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +9 -3
package/src/llama.cpp/ggml/src/ggml.c +117 -1327
package/src/llama.cpp/ggml/src/gguf.cpp +1329 -0
package/src/llama.cpp/include/llama-cpp.h +6 -1
package/src/llama.cpp/include/llama.h +138 -75
package/src/llama.cpp/src/CMakeLists.txt +13 -1
package/src/llama.cpp/src/llama-adapter.cpp +347 -0
package/src/llama.cpp/src/llama-adapter.h +74 -0
package/src/llama.cpp/src/llama-arch.cpp +1487 -0
package/src/llama.cpp/src/llama-arch.h +400 -0
package/src/llama.cpp/src/llama-batch.cpp +368 -0
package/src/llama.cpp/src/llama-batch.h +88 -0
package/src/llama.cpp/src/llama-chat.cpp +578 -0
package/src/llama.cpp/src/llama-chat.h +52 -0
package/src/llama.cpp/src/llama-context.cpp +1775 -0
package/src/llama.cpp/src/llama-context.h +128 -0
package/src/llama.cpp/src/llama-cparams.cpp +1 -0
package/src/llama.cpp/src/llama-cparams.h +37 -0
package/src/llama.cpp/src/llama-grammar.cpp +5 -4
package/src/llama.cpp/src/llama-grammar.h +3 -1
package/src/llama.cpp/src/llama-hparams.cpp +71 -0
package/src/llama.cpp/src/llama-hparams.h +139 -0
package/src/llama.cpp/src/llama-impl.cpp +167 -0
package/src/llama.cpp/src/llama-impl.h +16 -136
package/src/llama.cpp/src/llama-kv-cache.cpp +718 -0
package/src/llama.cpp/src/llama-kv-cache.h +218 -0
package/src/llama.cpp/src/llama-mmap.cpp +589 -0
package/src/llama.cpp/src/llama-mmap.h +67 -0
package/src/llama.cpp/src/llama-model-loader.cpp +1124 -0
package/src/llama.cpp/src/llama-model-loader.h +167 -0
package/src/llama.cpp/src/llama-model.cpp +3953 -0
package/src/llama.cpp/src/llama-model.h +370 -0
package/src/llama.cpp/src/llama-quant.cpp +934 -0
package/src/llama.cpp/src/llama-quant.h +1 -0
package/src/llama.cpp/src/llama-sampling.cpp +147 -32
package/src/llama.cpp/src/llama-sampling.h +3 -19
package/src/llama.cpp/src/llama-vocab.cpp +1832 -575
package/src/llama.cpp/src/llama-vocab.h +97 -142
package/src/llama.cpp/src/llama.cpp +7160 -20314
package/src/llama.cpp/src/unicode.cpp +8 -3
package/src/llama.cpp/tests/CMakeLists.txt +2 -0
package/src/llama.cpp/tests/test-autorelease.cpp +3 -3
package/src/llama.cpp/tests/test-backend-ops.cpp +370 -59
package/src/llama.cpp/tests/test-chat-template.cpp +162 -125
package/src/llama.cpp/tests/test-gguf.cpp +222 -187
package/src/llama.cpp/tests/test-model-load-cancel.cpp +1 -1
package/src/llama.cpp/tests/test-sampling.cpp +0 -1
package/src/llama.cpp/tests/test-tokenizer-0.cpp +4 -4
package/src/llama.cpp/tests/test-tokenizer-1-bpe.cpp +9 -7
package/src/llama.cpp/tests/test-tokenizer-1-spm.cpp +8 -6

package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp CHANGED Viewed

@@ -145,6 +145,8 @@ class vk_perf_logger;
 #endif
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
+static constexpr uint32_t mul_mat_vec_max_cols = 8;
 struct vk_device_struct {
     std::mutex mutex;
@@ -202,8 +204,8 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
@@ -226,6 +228,8 @@ struct vk_device_struct {
     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_cpy_f32_quant[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
@@ -382,10 +386,13 @@ struct vk_flash_attn_push_constants {
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nb01;
     uint32_t nb02;
     uint32_t nb03;
+    uint32_t nb11;
     uint32_t nb12;
     uint32_t nb13;
+    uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
     uint32_t nb31;
@@ -411,7 +418,7 @@ struct vk_op_unary_push_constants {
     uint32_t ne;
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2;
     uint32_t ne0_012mp; uint32_t ne0_012L;
     uint32_t ne0_01mp;  uint32_t ne0_01L;
@@ -459,7 +466,7 @@ struct vk_op_binary_push_constants {
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
     uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2; int32_t param3;
 };
@@ -546,7 +553,7 @@ struct vk_staging_memcpy {
 };
 struct vk_op_upscale_push_constants {
-    uint32_t ne; uint32_t d_offset;
+    uint32_t ne; uint32_t a_offset; uint32_t d_offset;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -1404,10 +1411,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // spec constants and tile sizes for non-quant matmul/matmul_id
         l_warptile = { 256, 128, 256, 64 };
         m_warptile = { 256, 128, 128, 64 };
-        s_warptile = { 128,  32,  16, 64 };
+        s_warptile = { 128,  64,  64, 64 };
         l_wg_denoms = {128, 256, 1 };
         m_wg_denoms = {128, 128, 1 };
-        s_wg_denoms = { 32,  16, 1 };
+        s_wg_denoms = { 64,  64, 1 };
         // spec constants and tile sizes for quant matmul (non-Qi_K)
         l_warptile_mmq = { 256, 128, 256, 64 };
@@ -1643,6 +1650,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #undef CREATE_MM2
     } else
 #endif  // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
     if (device->coopmat_support) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -1737,7 +1745,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
         }
 #undef CREATE_MM2
 #undef CREATE_MM
-    } else if (device->fp16) {
+    } else
+#endif  // defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+    if (device->fp16) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l) \
@@ -1855,53 +1865,60 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // mul mat vec
-    // AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
-    uint32_t rm = 1;
-    if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
-        rm = 2;
-    // computing additional 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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    // the number of rows computed per shader depends on GPU model and quant
+    uint32_t rm_stdq = 1;
+    uint32_t rm_kq = 2;
+    if (device->vendor_id == VK_VENDOR_ID_AMD) {
+        if (device->subgroup_min_size == 64 && device->subgroup_max_size == 64) { // GCN
+            rm_stdq = 2;
+            rm_kq = 4;
+        }
+    } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
+        rm_stdq = 2;
+    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  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, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  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, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  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, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  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, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), 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*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+    }
     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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    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*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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}, {subgroup_size_16}, 1, true);
-    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*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    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*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    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*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
+    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), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    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), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    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), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    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), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    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), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    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*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
     // 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);
@@ -1953,6 +1970,20 @@ static void ggml_vk_load_shaders(vk_device& device) {
     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_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_0], "cpy_q5_0_f32", cpy_q5_0_f32_len, cpy_q5_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q5_1], "cpy_q5_1_f32", cpy_q5_1_f32_len, cpy_q5_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 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);
@@ -2012,11 +2043,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     }
     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);
@@ -2031,6 +2062,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     std::cerr << "Done!" << std::endl;
 }
+static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props);
 static vk_device ggml_vk_get_device(size_t idx) {
     VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")");
@@ -2166,9 +2199,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
-        if (device->vendor_id == VK_VENDOR_ID_INTEL || (device->vendor_id == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
-            // Intel drivers don't support coopmat properly yet
-            // Only RADV supports coopmat properly on AMD
+        if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props)) {
             device->coopmat_support = false;
         }
@@ -2233,6 +2264,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
             last_struct = (VkBaseOutStructure *)&subgroup_size_control_features;
         }
+#if defined(VK_KHR_cooperative_matrix)
         VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
         coopmat_features.pNext = nullptr;
         coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
@@ -2242,6 +2274,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
             last_struct->pNext = (VkBaseOutStructure *)&coopmat_features;
             last_struct = (VkBaseOutStructure *)&coopmat_features;
         }
+#endif
 #if defined(VK_NV_cooperative_matrix2)
         VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {};
@@ -2263,6 +2296,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
             device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
+            device_extensions.push_back("VK_EXT_subgroup_size_control");
         }
         device->subgroup_size_control = device->subgroup_size_control &&
@@ -2271,10 +2305,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (device->subgroup_size_control) {
             device->subgroup_require_full_support = subgroup_size_control_features.computeFullSubgroups;
-            device_extensions.push_back("VK_EXT_subgroup_size_control");
         }
+#if defined(VK_KHR_cooperative_matrix)
         device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix;
+#endif
         if (coopmat2_support) {
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
@@ -2367,6 +2402,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_KHR_shader_float16_int8");
         }
+#if defined(VK_KHR_cooperative_matrix)
         if (device->coopmat_support) {
             // Query supported shapes
             std::vector<VkCooperativeMatrixPropertiesKHR> cm_props;
@@ -2433,7 +2469,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (device->coopmat_support) {
             device_extensions.push_back("VK_KHR_cooperative_matrix");
         }
+#endif
         device->name = GGML_VK_NAME + std::to_string(idx);
         device_create_info = {
@@ -2506,7 +2542,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
     return vk_instance.devices[idx];
 }
 static void ggml_vk_print_gpu_info(size_t idx) {
     GGML_ASSERT(idx < vk_instance.device_indices.size());
     size_t dev_num = vk_instance.device_indices[idx];
@@ -2545,9 +2580,11 @@ static void ggml_vk_print_gpu_info(size_t idx) {
             fp16_storage = true;
         } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
             fp16_compute = true;
-        } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+       } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
                    !getenv("GGML_VK_DISABLE_COOPMAT")) {
             coopmat_support = true;
+#endif
 #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
         } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
                    !getenv("GGML_VK_DISABLE_COOPMAT2")) {
@@ -2556,9 +2593,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         }
     }
-    if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
-        // Intel drivers don't support coopmat properly yet
-        // Only RADV supports coopmat properly on AMD
+    if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props)) {
         coopmat_support = false;
     }
@@ -2587,6 +2622,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     // Pointer to the last chain element
     VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features;
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
     VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
     coopmat_features.pNext = nullptr;
     coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
@@ -2602,6 +2638,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     fp16 = fp16 && vk12_features.shaderFloat16;
     coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix;
+#endif
     std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none";
@@ -2887,9 +2924,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : 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) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
     GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+    GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
     switch (a_type) {
         case GGML_TYPE_F32:
@@ -2910,7 +2948,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1];
 }
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
@@ -3205,8 +3243,8 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
         GGML_ABORT("fatal error");
     }
     // Check if src is pinned memory
-    vk_buffer buf;
-    size_t buf_offset;
+    vk_buffer buf = nullptr;
+    size_t buf_offset = 0;
     ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset);
     const uint64_t ne0 = tensor->ne[0];
@@ -3269,7 +3307,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
     VkBufferCopy buf_copy{ 0, offset, copy_size };
     ggml_vk_sync_buffers(subctx);
-    vkCmdCopyBuffer(subctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
+    vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
     for (uint64_t i3 = 0; i3 < ne3; i3++) {
         for (uint64_t i2 = 0; i2 < ne2; i2++) {
@@ -3302,7 +3340,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
     }
     // Check if src is pinned memory
     vk_buffer buf = nullptr;
-    size_t buf_offset;
+    size_t buf_offset = 0;
     ggml_vk_host_get(dst->device, src, buf, buf_offset);
     if (buf != nullptr) {
@@ -3344,7 +3382,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
         copy_size};
     ggml_vk_sync_buffers(subctx);
-    vkCmdCopyBuffer(subctx->s->buffer, staging_buffer->buffer, dst->buffer, 1, &buf_copy);
+    vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
     if (width == spitch) {
         deferred_memcpy((uint8_t *)staging_buffer->ptr, src, width * height, &subctx->in_memcpys);
@@ -3400,7 +3438,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
     // Check if dst is pinned memory
     vk_buffer buf = nullptr;
-    size_t buf_offset;
+    size_t buf_offset = 0;
     ggml_vk_host_get(src->device, dst, buf, buf_offset);
     std::vector<vk::BufferCopy> slices(1);
@@ -3480,7 +3518,7 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds
     VkBufferCopy bc{ src_offset, dst_offset, size };
-    vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
+    vkCmdCopyBuffer(ctx->s->buffer, (VkBuffer)src->buffer, (VkBuffer)dst->buffer, 1, &bc);
 }
 static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
@@ -3670,6 +3708,33 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cpy_f16_f16;
         }
     }
+    if (src->type == GGML_TYPE_F32) {
+        switch (to) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
+            return ctx->device->pipeline_cpy_f32_quant[to];
+        default:
+            break;
+        }
+    }
+    if (to == GGML_TYPE_F32) {
+        switch (src->type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
+            return ctx->device->pipeline_cpy_quant_f32[src->type];
+        default:
+            break;
+        }
+    }
     std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
     GGML_ABORT("fatal error");
@@ -3732,9 +3797,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
     bool src0_uma = false;
@@ -3920,8 +3985,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t ne12 = src1->ne[2];
     const uint64_t ne13 = src1->ne[3];
-    GGML_ASSERT(ne11 == 1);
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
     const uint64_t ne22 = dst->ne[2];
@@ -3930,13 +3993,18 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t r2 = ne12 / ne02;
     const uint64_t r3 = ne13 / ne03;
+    // batch_n indicates that we need to compute a few vector results, and this assumes
+    // ne12 and ne13 are 1. It overloads the batch_strides to hold the row strides.
+    GGML_ASSERT(ne11 == 1 || ne12 * ne13 == 1);
+    bool batch_n = ne11 > 1;
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
     bool src0_uma = false;
@@ -3980,7 +4048,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
@@ -4052,8 +4120,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
     }
-    uint32_t stride_batch_x = ne00*ne01;
-    uint32_t stride_batch_y = ne10*ne11;
+    // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
+    uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
+    uint32_t stride_batch_y = batch_n ? ne10 : (ne10*ne11);
+    uint32_t stride_batch_d = batch_n ? ne20 : (ne20*ne21);
     if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
         stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
@@ -4076,7 +4146,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
+        stride_batch_x, stride_batch_y, stride_batch_d,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_sync_buffers(subctx);
@@ -4112,7 +4182,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
     bool src1_uma = false;
@@ -4256,7 +4326,10 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     } 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))) {
+    // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
+    // when ne12 and ne13 are one.
+    } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 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);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -4300,11 +4373,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
     ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
-    vk_buffer d_ids;
+    vk_buffer d_ids = nullptr;
     size_t ids_buf_offset = 0;
     bool src0_uma = false;
@@ -4505,11 +4578,11 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
     ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
-    vk_buffer d_ids;
+    vk_buffer d_ids = nullptr;
     size_t ids_buf_offset = 0;
     bool src0_uma = false;
@@ -4739,7 +4812,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
     assert(pipelines);
-    bool aligned = (KV % pipelines[1]->align) == 0;
+    const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
+    const uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
+    const uint32_t v_stride = (uint32_t)(nbv1 / ggml_type_size(v->type));
+    bool aligned = (KV % pipelines[1]->align) == 0 &&
+                   // the "aligned" shader variant will forcibly align strides, for performance
+                   (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;
     vk_pipeline pipeline = pipelines[aligned];
     assert(pipeline);
@@ -4768,22 +4848,22 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     ggml_vk_sync_buffers(subctx);
-    vk_buffer d_Q, d_K, d_V, d_D, d_M;
-    uint64_t q_buf_offset, k_buf_offset, v_buf_offset, d_buf_offset, m_buf_offset;
+    vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr;
+    size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0;
     bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false;
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset);
-        ggml_vk_host_get(ctx->device, k->data, d_K, q_buf_offset);
-        ggml_vk_host_get(ctx->device, v->data, d_V, q_buf_offset);
-        ggml_vk_host_get(ctx->device, dst->data, d_D, q_buf_offset);
+        ggml_vk_host_get(ctx->device, k->data, d_K, k_buf_offset);
+        ggml_vk_host_get(ctx->device, v->data, d_V, v_buf_offset);
+        ggml_vk_host_get(ctx->device, dst->data, d_D, d_buf_offset);
         Q_uma = d_Q != nullptr;
         K_uma = d_K != nullptr;
         V_uma = d_V != nullptr;
         D_uma = d_D != nullptr;
         if (mask) {
-            ggml_vk_host_get(ctx->device, mask->data, d_M, q_buf_offset);
+            ggml_vk_host_get(ctx->device, mask->data, d_M, m_buf_offset);
             M_uma = d_M != nullptr;
         }
     }
@@ -4821,7 +4901,18 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
-    const vk_flash_attn_push_constants pc = { N, KV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, (uint32_t)neq2, (uint32_t)neq3, (uint32_t)nek2, (uint32_t)nek3, (uint32_t)nev2, (uint32_t)nev3, nem1, (uint32_t)nbq2, (uint32_t)nbq3, (uint32_t)nbk2, (uint32_t)nbk3, (uint32_t)nbv2, (uint32_t)nbv3, nbm1, scale, max_bias, logit_softcap, mask != nullptr, n_head_log2, m0, m1 };
+    const vk_flash_attn_push_constants pc = { N, KV,
+                                              (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
+                                              (uint32_t)neq2, (uint32_t)neq3,
+                                              (uint32_t)nek2, (uint32_t)nek3,
+                                              (uint32_t)nev2, (uint32_t)nev3,
+                                              nem1,
+                                              q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
+                                              k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
+                                              v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
+                                              nbm1,
+                                              scale, max_bias, logit_softcap,
+                                              mask != nullptr, n_head_log2, m0, m1 };
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                 {
                                     vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE},
@@ -5071,6 +5162,57 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     }
 }
+static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_tensor * t)
+{
+    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
+}
+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    GGML_UNUSED(p);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(dst);
+    static_assert(!std::is_const<T>::value, "unexpected type");
+    GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
+    GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
+    GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
+    GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
+}
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+    p.misalign_offsets = (a_offset << 16) | d_offset;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+    GGML_ASSERT(dst->op != GGML_OP_GET_ROWS || (a_offset == 0 && b_offset == 0 && d_offset == 0));
+    p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
+    GGML_UNUSED(src2);
+}
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+    p.a_offset = a_offset;
+    p.d_offset = d_offset;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
@@ -5082,7 +5224,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
-    GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
+    GGML_ASSERT(op == GGML_OP_GET_ROWS || op == GGML_OP_CPY || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
     GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
     GGML_ASSERT(dst->buffer != nullptr);
     const uint64_t ne00 = src0->ne[0];
@@ -5174,8 +5316,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
     GGML_ASSERT(d_D != nullptr);
-    uint64_t d_buf_offset = ((vk_tensor_offset(dst) + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    GGML_ASSERT(d_buf_offset == vk_tensor_offset(dst) || op == GGML_OP_CPY);  // NOLINT
+    uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
     if(!src0_uma) {
         d_X = src0_buf_ctx->dev_buffer;
         x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
@@ -5191,6 +5332,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
         GGML_ASSERT(d_Z != nullptr);
     }
+    // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
+    x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
@@ -5378,7 +5525,6 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
     int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
     int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
@@ -5390,7 +5536,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f, offset,
     }, dryrun);
 }
@@ -5474,8 +5620,8 @@ static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subc
     ggml_vk_sync_buffers(subctx);
-    vk_buffer d_D, d_K, d_V, d_R, d_TF, d_TD, d_State;
-    uint64_t k_offset, v_offset, r_offset, tf_offset, td_offset, state_offset, dst_offset;
+    vk_buffer d_D = nullptr, d_K = nullptr, d_V = nullptr, d_R = nullptr, d_TF = nullptr, d_TD = nullptr, d_State = nullptr;
+    size_t k_offset = 0, v_offset = 0, r_offset = 0, tf_offset = 0, td_offset = 0, state_offset = 0, dst_offset = 0;
     bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false;
     if (ctx->device->uma) {
@@ -5551,9 +5697,9 @@ static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subc
 }
 static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
-    const size_t seq_length = dst->src[0]->ne[3];
+    const size_t seq_length = dst->src[0]->ne[2];
     const size_t n_embed = dst->ne[0];
-    const size_t n_heads = dst->src[0]->ne[2];
+    const size_t n_heads = dst->src[0]->ne[1];
     const size_t n_seqs = dst->src[5]->ne[1];
     ggml_vk_op_f32_rwkv6(
@@ -5594,7 +5740,7 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
     const float sf3 = (float)dst->ne[3] / src0->ne[3];
     ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
-        (uint32_t)ggml_nelements(dst), 0,
+        (uint32_t)ggml_nelements(dst), 0, 0,
         (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
         sf0, sf1, sf2, sf3,
@@ -5704,13 +5850,12 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
     ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
@@ -7824,12 +7969,36 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1] != nullptr ? op->src[1]->type : src0_type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
+                if (src0_type == GGML_TYPE_F32) {
+                    switch (src1_type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        break;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
-                    return true;
+                if (src1_type == GGML_TYPE_F32) {
+                    switch (src0_type) {
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        break;
+                    }
                 }
                 if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
                     return true;
                 }
@@ -8016,6 +8185,25 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
     UNUSED(instance_extensions);
 }
+static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props) {
+    switch (props.vendorID) {
+    case VK_VENDOR_ID_INTEL:
+        // Intel drivers don't support coopmat properly yet
+        return false;
+    case VK_VENDOR_ID_AMD:
+        if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
+            // Workaround for AMD proprietary driver reporting support on all GPUs
+            const std::string name = props.deviceName;
+            return name.rfind("AMD Radeon RX 7", 0) == 0   || name.rfind("AMD Radeon(TM) RX 7", 0) == 0   || // RDNA 3 consumer GPUs
+                   name.rfind("AMD Radeon PRO W7", 0) == 0 || name.rfind("AMD Radeon(TM) PRO W7", 0) == 0 || // RDNA 3 workstation GPUs
+                   name.rfind("AMD Radeon 7", 0) == 0      || name.rfind("AMD Radeon(TM) 7", 0) == 0;        // RDNA 3 APUs
+        }
+        return true;
+    default:
+        return true;
+    }
+}
 // checks
 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -8501,6 +8689,7 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
     ggml_tensor * src2 = tensor->src[2];
+    ggml_tensor * src3 = tensor->src[3];
     void * tensor_data = tensor->data;
@@ -8563,6 +8752,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
                         if (src2 != nullptr) {
                             std::cerr << "src2=" << src2 << " src2->name=" << src2->name << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
                         }
+                        if (src3 != nullptr) {
+                            std::cerr << "src3=" << src3 << " src3->name=" << src3->name << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+                        }
                         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
                         std::cerr << std::endl << "Result:" << std::endl;
                         ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3);
@@ -8607,6 +8799,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
         if (src2 != nullptr) {
             std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
         }
+        if (src3 != nullptr) {
+            std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
@@ -8629,6 +8824,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
         if (src2 != nullptr) {
             std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
         }
+        if (src3 != nullptr) {
+            std::cerr << "src3=" << src3 << " op=" << ggml_op_name(src3->op) << " type=" << ggml_type_name(src3->type) << " ne0=" << src3->ne[0] << " nb0=" << src3->nb[0] << " ne1=" << src3->ne[1] << " nb1=" << src3->nb[1] << " ne2=" << src3->ne[2] << " nb2=" << src3->nb[2] << " ne3=" << src3->ne[3] << " nb3=" << src3->nb[3] << " offset=" << src3->view_offs << std::endl;
+        }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
         ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]);