npm - @fugood/llama.node - Versions diffs - 0.3.13 → 0.3.15 - Mend

@fugood/llama.node 0.3.13 → 0.3.15

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 (184) hide show

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 +1 -1
package/package.json +1 -1
package/src/LlamaContext.cpp +98 -76
package/src/LlamaContext.h +1 -1
package/src/common.hpp +1 -2
package/src/llama.cpp/.github/workflows/build.yml +89 -10
package/src/llama.cpp/.github/workflows/server.yml +2 -0
package/src/llama.cpp/CMakeLists.txt +9 -1
package/src/llama.cpp/cmake/common.cmake +2 -0
package/src/llama.cpp/common/CMakeLists.txt +3 -3
package/src/llama.cpp/common/arg.cpp +132 -13
package/src/llama.cpp/common/chat.cpp +960 -266
package/src/llama.cpp/common/chat.h +135 -0
package/src/llama.cpp/common/common.cpp +33 -174
package/src/llama.cpp/common/common.h +27 -67
package/src/llama.cpp/common/json-schema-to-grammar.cpp +4 -5
package/src/llama.cpp/common/json-schema-to-grammar.h +0 -1
package/src/llama.cpp/common/{minja.hpp → minja/minja.hpp} +37 -5
package/src/llama.cpp/common/ngram-cache.cpp +1 -0
package/src/llama.cpp/common/sampling.cpp +45 -7
package/src/llama.cpp/common/speculative.cpp +10 -9
package/src/llama.cpp/common/speculative.h +1 -1
package/src/llama.cpp/docs/build.md +45 -7
package/src/llama.cpp/examples/batched-bench/batched-bench.cpp +2 -2
package/src/llama.cpp/examples/cvector-generator/cvector-generator.cpp +4 -2
package/src/llama.cpp/examples/embedding/embedding.cpp +2 -1
package/src/llama.cpp/examples/export-lora/export-lora.cpp +4 -2
package/src/llama.cpp/examples/gritlm/gritlm.cpp +2 -2
package/src/llama.cpp/examples/imatrix/imatrix.cpp +3 -4
package/src/llama.cpp/examples/infill/infill.cpp +2 -2
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +2 -2
package/src/llama.cpp/examples/llama.android/llama/src/main/cpp/llama-android.cpp +5 -5
package/src/llama.cpp/examples/llava/CMakeLists.txt +7 -0
package/src/llama.cpp/examples/llava/clip.cpp +373 -107
package/src/llama.cpp/examples/llava/clip.h +19 -3
package/src/llama.cpp/examples/llava/gemma3-cli.cpp +341 -0
package/src/llama.cpp/examples/llava/llava.cpp +4 -2
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +30 -11
package/src/llama.cpp/examples/lookahead/lookahead.cpp +7 -6
package/src/llama.cpp/examples/lookup/lookup.cpp +1 -1
package/src/llama.cpp/examples/main/main.cpp +79 -34
package/src/llama.cpp/examples/parallel/parallel.cpp +6 -5
package/src/llama.cpp/examples/passkey/passkey.cpp +15 -14
package/src/llama.cpp/examples/perplexity/perplexity.cpp +6 -6
package/src/llama.cpp/examples/quantize/quantize.cpp +1 -0
package/src/llama.cpp/examples/quantize-stats/quantize-stats.cpp +2 -2
package/src/llama.cpp/examples/retrieval/retrieval.cpp +1 -1
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.cpp +882 -237
package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.h +35 -26
package/src/llama.cpp/examples/run/run.cpp +196 -108
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +2 -2
package/src/llama.cpp/examples/server/server.cpp +113 -101
package/src/llama.cpp/examples/server/utils.hpp +94 -105
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +2 -2
package/src/llama.cpp/examples/speculative/speculative.cpp +14 -14
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +1 -1
package/src/llama.cpp/examples/sycl/run-llama2.sh +2 -2
package/src/llama.cpp/examples/tts/tts.cpp +263 -151
package/src/llama.cpp/ggml/CMakeLists.txt +14 -1
package/src/llama.cpp/ggml/cmake/common.cmake +26 -0
package/src/llama.cpp/ggml/include/ggml-alloc.h +1 -1
package/src/llama.cpp/ggml/include/ggml-backend.h +3 -3
package/src/llama.cpp/ggml/include/ggml-cpu.h +3 -0
package/src/llama.cpp/ggml/include/ggml.h +29 -1
package/src/llama.cpp/ggml/src/CMakeLists.txt +15 -34
package/src/llama.cpp/ggml/src/ggml-alloc.c +24 -15
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +58 -54
package/src/llama.cpp/ggml/src/ggml-backend.cpp +10 -8
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +6 -2
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +3 -7
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +3 -5
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +139 -16
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +4 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +151 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +1546 -387
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +1645 -113
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +22 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +259 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +61 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +288 -0
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +15 -2
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +2 -1
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +3 -1
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +14 -0
package/src/llama.cpp/ggml/src/ggml-impl.h +1 -1
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +4 -5
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +242 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +6 -6
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +1 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +315 -138
package/src/llama.cpp/ggml/src/ggml-quants.c +114 -114
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +5 -0
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +2 -1
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +17 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +117 -36
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +33 -4
package/src/llama.cpp/ggml/src/ggml-sycl/convert.hpp +2 -2
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.cpp +701 -0
package/src/llama.cpp/ggml/src/ggml-sycl/cpy.hpp +11 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dequantize.hpp +55 -0
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +147 -16
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +40 -40
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.cpp +307 -0
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +262 -746
package/src/llama.cpp/ggml/src/ggml-sycl/mmq.cpp +0 -1
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +75 -78
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +114 -6
package/src/llama.cpp/ggml/src/ggml-sycl/norm.hpp +6 -0
package/src/llama.cpp/ggml/src/ggml-sycl/softmax.cpp +4 -1
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
package/src/llama.cpp/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.cpp +305 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv.hpp +10 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +498 -188
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +0 -4
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +16 -3
package/src/llama.cpp/ggml/src/ggml.c +93 -5
package/src/llama.cpp/include/llama.h +105 -27
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-gpt-4o.gguf.out +46 -0
package/src/llama.cpp/requirements/requirements-all.txt +1 -0
package/src/llama.cpp/requirements/requirements-tool_bench.txt +12 -0
package/src/llama.cpp/requirements.txt +1 -0
package/src/llama.cpp/src/CMakeLists.txt +5 -2
package/src/llama.cpp/src/llama-adapter.cpp +19 -20
package/src/llama.cpp/src/llama-adapter.h +11 -9
package/src/llama.cpp/src/llama-arch.cpp +123 -16
package/src/llama.cpp/src/llama-arch.h +19 -0
package/src/llama.cpp/src/llama-batch.h +2 -2
package/src/llama.cpp/src/llama-chat.cpp +1 -0
package/src/llama.cpp/src/llama-context.cpp +2253 -1222
package/src/llama.cpp/src/llama-context.h +214 -77
package/src/llama.cpp/src/llama-cparams.h +1 -0
package/src/llama.cpp/src/llama-grammar.cpp +182 -182
package/src/llama.cpp/src/llama-grammar.h +12 -3
package/src/llama.cpp/src/llama-graph.cpp +1662 -0
package/src/llama.cpp/src/llama-graph.h +574 -0
package/src/llama.cpp/src/llama-hparams.cpp +8 -0
package/src/llama.cpp/src/llama-hparams.h +9 -0
package/src/llama.cpp/src/llama-io.cpp +15 -0
package/src/llama.cpp/src/llama-io.h +35 -0
package/src/llama.cpp/src/llama-kv-cache.cpp +1006 -291
package/src/llama.cpp/src/llama-kv-cache.h +178 -109
package/src/llama.cpp/src/llama-memory.cpp +1 -0
package/src/llama.cpp/src/llama-memory.h +21 -0
package/src/llama.cpp/src/llama-mmap.cpp +11 -1
package/src/llama.cpp/src/llama-model.cpp +8230 -122
package/src/llama.cpp/src/llama-model.h +34 -1
package/src/llama.cpp/src/llama-quant.cpp +10 -1
package/src/llama.cpp/src/llama-sampling.cpp +43 -10
package/src/llama.cpp/src/llama-vocab.cpp +12 -0
package/src/llama.cpp/src/llama.cpp +51 -9837
package/src/llama.cpp/tests/test-backend-ops.cpp +247 -112
package/src/llama.cpp/tests/test-chat-template.cpp +32 -22
package/src/llama.cpp/tests/test-chat.cpp +593 -395
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +63 -63
package/src/llama.cpp/tests/test-quantize-fns.cpp +1 -9
package/src/llama.cpp/Sources/llama/llama.h +0 -4
package/src/llama.cpp/common/chat.hpp +0 -55
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +0 -143
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp +0 -9
/package/src/llama.cpp/common/{chat-template.hpp → minja/chat-template.hpp} +0 -0

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

@@ -29,6 +29,7 @@
 #include "ggml-vulkan-shaders.hpp"
+#define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 #define VK_VENDOR_ID_AMD 0x1002
@@ -149,6 +150,66 @@ static void ggml_vk_destroy_buffer(vk_buffer& buf);
 static constexpr uint32_t mul_mat_vec_max_cols = 8;
+enum vk_device_architecture {
+    OTHER,
+    AMD_GCN,
+    AMD_RDNA1,
+    AMD_RDNA2,
+    AMD_RDNA3,
+};
+static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
+    vk::PhysicalDeviceProperties props = device.getProperties();
+    if (props.vendorID == VK_VENDOR_ID_AMD) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+        bool amd_shader_core_properties = false;
+        bool integer_dot_product = false;
+        bool subgroup_size_control = false;
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_AMD_shader_core_properties", properties.extensionName) == 0) {
+                amd_shader_core_properties = true;
+            } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) {
+                integer_dot_product = true;
+            } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                subgroup_size_control = true;
+            }
+        }
+        if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) {
+            return vk_device_architecture::OTHER;
+        }
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd;
+        vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+        props2.pNext = &shader_core_props_amd;
+        shader_core_props_amd.pNext = &integer_dot_props;
+        integer_dot_props.pNext = &subgroup_size_control_props;
+        device.getProperties2(&props2);
+        if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) {
+            return vk_device_architecture::AMD_GCN;
+        }
+        if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) {
+            // RDNA
+            if (shader_core_props_amd.wavefrontsPerSimd == 20) {
+                return vk_device_architecture::AMD_RDNA1;
+            }
+            if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) {
+                return vk_device_architecture::AMD_RDNA3;
+            }
+            return vk_device_architecture::AMD_RDNA2;
+        }
+    }
+    return vk_device_architecture::OTHER;
+}
 struct vk_device_struct {
     std::mutex mutex;
@@ -161,6 +222,7 @@ struct vk_device_struct {
     bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
+    vk_device_architecture architecture;
     vk_queue compute_queue;
     vk_queue transfer_queue;
     bool single_queue;
@@ -241,15 +303,20 @@ struct vk_device_struct {
     vk_pipeline pipeline_norm_f32;
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
+    vk_pipeline pipeline_rms_norm_back_f32;
+    vk_pipeline pipeline_l2_norm_f32;
     vk_pipeline pipeline_gelu_f32;
     vk_pipeline pipeline_gelu_quick_f32;
     vk_pipeline pipeline_silu_f32;
+    vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_relu_f32;
     vk_pipeline pipeline_leaky_relu_f32;
     vk_pipeline pipeline_tanh_f32;
+    vk_pipeline pipeline_sigmoid_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
+    vk_pipeline pipeline_soft_max_back_f32;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
@@ -262,6 +329,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_timestep_embedding_f32;
     vk_pipeline pipeline_pool2d_f32;
     vk_pipeline pipeline_rwkv_wkv6_f32;
+    vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
@@ -364,6 +432,7 @@ struct vk_mat_mat_push_constants {
     uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
     uint32_t k_split;
     uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+    uint32_t padded_N;
 };
 struct vk_mat_vec_push_constants {
     uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
@@ -376,6 +445,7 @@ struct vk_mat_mat_id_push_constants {
     uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
     uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
     uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11;
+    uint32_t padded_N;
 };
 struct vk_mat_vec_id_push_constants {
     uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
@@ -504,6 +574,7 @@ struct vk_op_rope_push_constants {
     uint32_t s1;
     uint32_t s2;
     int32_t sections[4];
+    uint32_t is_back;
 };
 struct vk_op_soft_max_push_constants {
@@ -560,6 +631,13 @@ struct vk_op_rwkv_wkv6_push_constants {
     uint32_t H;
 };
+struct vk_op_rwkv_wkv7_push_constants {
+    uint32_t B;
+    uint32_t T;
+    uint32_t C;
+    uint32_t H;
+};
 // 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) {}
@@ -1440,6 +1518,73 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
     return supported;
 }
+struct GpuPipelineConfig {
+    // GPU architecture identifier.
+    // Example: vk_device_architecture::AMD_GCN
+    vk_device_architecture arch;
+    // Mapping of pipeline names to their specific subgroup sizes.
+    // Example: {"soft_max_f32", 64}
+    std::unordered_map<std::string, uint32_t> pipelines;
+    // Default subgroup size for this GPU.
+    // Defaults to 0 if not explicitly provided.
+    uint32_t default_subgroup_size = 0;
+};
+// Pipeline configuration for RDNA1 GPUs.
+static const std::unordered_map<std::string, uint32_t> rdna1_pipelines = {
+    {"soft_max", 64}, {"im2col", 64},
+    {"argmax", 64}, {"mul_mat_vec", 64},
+    {"mul_mat_vec_f16", 32}, {"mul_mat_vec_f32_f16", 32}
+};
+// Pipeline configuration for RDNA2 GPUs.
+static const std::unordered_map<std::string, uint32_t> rdna2_pipelines = {
+    {"soft_max", 64}, {"im2col", 64},
+};
+static constexpr uint32_t RDNA_DEFAULT_SUBGROUP_SIZE = 32;
+// Define configurations for different GPUs.
+static std::vector<GpuPipelineConfig> gpu_pipeline_configs = {
+    {
+        vk_device_architecture::AMD_RDNA1,
+        {
+            rdna1_pipelines,
+        },
+        RDNA_DEFAULT_SUBGROUP_SIZE
+    },
+    {
+        vk_device_architecture::AMD_RDNA2,
+        {
+            rdna2_pipelines,
+        },
+        RDNA_DEFAULT_SUBGROUP_SIZE
+    },
+};
+static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_device_architecture &arch) {
+    for (const auto &config : gpu_pipeline_configs) {
+        if (config.arch == arch) {
+            auto pipIt = config.pipelines.find(pipeline_name);
+            if (pipIt != config.pipelines.end()) {
+                return pipIt->second;
+            }
+            std::vector<std::pair<std::string, uint32_t>> sorted_pipelines(config.pipelines.begin(), config.pipelines.end());
+            std::sort(sorted_pipelines.begin(), sorted_pipelines.end(),
+                      [](const auto &a, const auto &b) { return a.first.size() > b.first.size(); });
+            for (const auto &entry : sorted_pipelines) {
+                if (pipeline_name.find(entry.first) != std::string::npos) {
+                    return entry.second;
+                }
+            }
+            return config.default_subgroup_size;
+        }
+    }
+    return 0; // If no matching configuration is found
+}
 static void ggml_vk_load_shaders(vk_device& device) {
     VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
@@ -1461,36 +1606,36 @@ static void ggml_vk_load_shaders(vk_device& device) {
     uint32_t l_align, m_align, s_align;
     if (device->coopmat2) {
         // 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,  64,  64, 64 };
+        l_warptile = { 256, 128, 256, 64, 1 };
+        m_warptile = { 256, 128, 128, 64, 0 };
+        s_warptile = { 128,  64,  64, 64, 0 };
         l_wg_denoms = {128, 256, 1 };
         m_wg_denoms = {128, 128, 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 };
-        m_warptile_mmq = { 256, 128, 128, 64 };
-        s_warptile_mmq = { 256, 128, 128, 64 };
+        l_warptile_mmq = { 256, 128, 256, 64, 1 };
+        m_warptile_mmq = { 256, 128, 128, 64, 1 };
+        s_warptile_mmq = { 256, 32,  64, 128, 0 };
         l_mmq_wg_denoms = { 128, 256, 1 };
         m_mmq_wg_denoms = { 128, 128, 1 };
-        s_mmq_wg_denoms = { 128, 128, 1 };
+        s_mmq_wg_denoms = { 32,  64,  1 };
         // spec constants and tile sizes for quant matmul (Qi_K)
-        l_warptile_mmq_k = { 256, 128, 512, 16 };
-        m_warptile_mmq_k = { 256, 128, 256, 16 };
-        s_warptile_mmq_k = { 256, 32, 128, 64 };
-        l_mmq_wg_denoms_k = { 128, 512, 1 };
-        m_mmq_wg_denoms_k = { 128, 256, 1 };
-        s_mmq_wg_denoms_k = { 32, 128, 1 };
+        l_warptile_mmq_k = { 256, 64, 128, 64,  1 };
+        m_warptile_mmq_k = { 256, 32,  64, 64,  0 };
+        s_warptile_mmq_k = { 256, 32,  32, 128, 0 };
+        l_mmq_wg_denoms_k = { 64, 128, 1 };
+        m_mmq_wg_denoms_k = { 32,  64, 1 };
+        s_mmq_wg_denoms_k = { 32,  32, 1 };
         // spec constants and tile sizes for quant matmul_id
-        l_warptile_mmqid = { 256, 128, 128, 16 };
-        m_warptile_mmqid = { 256, 128, 64, 16 };
-        s_warptile_mmqid = { 256, 64, 64, 16 };
-        l_mmqid_wg_denoms = { 128, 128, 1 };
+        l_warptile_mmqid = { 256, 128, 64, 16, 0 };
+        m_warptile_mmqid = { 256, 128, 64, 16, 0 };
+        s_warptile_mmqid = { 256, 128, 64, 16, 0 };
+        l_mmqid_wg_denoms = { 128, 64, 1 };
         m_mmqid_wg_denoms = { 128, 64, 1 };
-        s_mmqid_wg_denoms = { 64, 64, 1 };
+        s_mmqid_wg_denoms = { 128, 64, 1 };
         l_align = 128;
         m_align =  64;
@@ -1566,6 +1711,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                               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, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
+        if (!require_full_subgroups && required_subgroup_size == 0) {
+            required_subgroup_size = get_subgroup_size(name, device->architecture);
+        }
         if (!pipeline) {
             pipeline = std::make_shared<vk_pipeline_struct>();
             pipeline->name = name;
@@ -1987,6 +2136,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         }
     } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
         rm_stdq = 2;
+    uint32_t rm_iq = 2 * rm_kq;
     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);
@@ -2001,15 +2151,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
         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_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_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_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_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_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_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_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_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_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_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_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_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_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_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_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_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_f32_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, 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), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, 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);
@@ -2023,15 +2173,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
         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_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_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_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_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_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_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_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_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_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_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_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_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_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_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_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_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_f16_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, 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), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, 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);
@@ -2046,15 +2196,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
     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_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   mul_mat_vec_id_iq1_s_f32_len,   mul_mat_vec_id_iq1_s_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_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   mul_mat_vec_id_iq1_m_f32_len,   mul_mat_vec_id_iq1_m_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_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_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_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_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_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_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_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_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_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_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_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_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);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S],   "mul_mat_vec_id_iq1_s_f32",   mul_mat_vec_id_iq1_s_f32_len,   mul_mat_vec_id_iq1_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M],   "mul_mat_vec_id_iq1_m_f32",   mul_mat_vec_id_iq1_m_f32_len,   mul_mat_vec_id_iq1_m_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS],  "mul_mat_vec_id_iq4_xs_f32",  mul_mat_vec_id_iq4_xs_f32_len,  mul_mat_vec_id_iq4_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 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), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 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);
@@ -2121,6 +2271,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     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);
@@ -2180,9 +2332,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sigmoid_f32, "sigmoid_f32", sigmoid_f32_len, sigmoid_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
@@ -2190,6 +2344,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -2229,6 +2384,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
     for (auto &c : compiles) {
@@ -2237,7 +2394,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->need_compiles = false;
 }
-static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props);
+static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch);
 static vk_device ggml_vk_get_device(size_t idx) {
     VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")");
@@ -2266,6 +2423,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->physical_device = physical_devices[dev_num];
         const std::vector<vk::ExtensionProperties> ext_props = device->physical_device.enumerateDeviceExtensionProperties();
+        device->architecture = get_device_architecture(device->physical_device);
         const char* GGML_VK_PREFER_HOST_MEMORY = getenv("GGML_VK_PREFER_HOST_MEMORY");
         device->prefer_host_memory = GGML_VK_PREFER_HOST_MEMORY != nullptr;
@@ -2278,7 +2437,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool coopmat2_support = false;
         device->coopmat_support = false;
-        // Check if maintenance4 is supported
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
                 maintenance4_support = true;
@@ -2366,13 +2524,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (GGML_VK_SUBALLOCATION_BLOCK_SIZE != nullptr) {
             device->suballocation_block_size = std::stoul(GGML_VK_SUBALLOCATION_BLOCK_SIZE);
-#if defined(_WIN32)
-        } else if (device->vendor_id == VK_VENDOR_ID_NVIDIA) {
+        } else {
             // Limit batching of allocations to 1GB by default to avoid fragmentation issues
             device->suballocation_block_size = 1024*1024*1024;
-#endif
-        } else {
-            device->suballocation_block_size = device->max_memory_allocation_size;
         }
         device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);
@@ -2391,7 +2545,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
-        if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props)) {
+        if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props, device->architecture)) {
             device->coopmat_support = false;
         }
@@ -2769,7 +2923,10 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     subgroup_props.pNext = &driver_props;
     physical_device.getProperties2(&props2);
-    const size_t subgroup_size = subgroup_props.subgroupSize;
+    vk_device_architecture arch = get_device_architecture(physical_device);
+    uint32_t default_subgroup_size = get_subgroup_size("", arch);
+    const size_t subgroup_size = (default_subgroup_size != 0) ? default_subgroup_size : subgroup_props.subgroupSize;
     const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
     bool fp16_storage = false;
@@ -2795,7 +2952,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         }
     }
-    if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props)) {
+    const vk_device_architecture device_architecture = get_device_architecture(physical_device);
+    if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture)) {
         coopmat_support = false;
     }
@@ -3840,10 +3999,14 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")");
     if (ctx->device->coopmat2) {
-        if ((ctx->device->mul_mat_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
+        // Use large shader when the N dimension is greater than the medium shader's tile size
+        uint32_t crossover_large = mmp->m->wg_denoms[1];
+        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
             return aligned ? mmp->a_l : mmp->l;
         }
-        if ((ctx->device->mul_mat_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s[src0_type]) {
+        // Use medium shader when the N dimension is greater than the small shader's tile size
+        uint32_t crossover_medium = mmp->s->wg_denoms[1];
+        if ((ctx->device->mul_mat_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_s[src0_type]) {
             return aligned ? mmp->a_m : mmp->m;
         }
         return aligned ? mmp->a_s : mmp->s;
@@ -3868,18 +4031,19 @@ static void ggml_vk_matmul(
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
-        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) {
+        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3,
+        uint32_t padded_n) {
         VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")");
     ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
-        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 };
+        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
         return;
     }
     GGML_ASSERT(batch_stride_d == m * n);
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 };
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);
@@ -3888,13 +4052,17 @@ static void ggml_vk_matmul(
 }
 static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type src0_type) {
-    VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")");
+    VK_LOG_DEBUG("ggml_vk_guess_matmul_id_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")");
     if (ctx->device->coopmat2) {
-        if ((ctx->device->mul_mat_id_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) {
+        // Use large shader when the N dimension is greater than the medium shader's tile size
+        uint32_t crossover_large = mmp->m->wg_denoms[1];
+        if ((ctx->device->mul_mat_id_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) {
             return aligned ? mmp->a_l : mmp->l;
         }
-        if ((ctx->device->mul_mat_id_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s[src0_type]) {
+        // Use medium shader when the N dimension is greater than the small shader's tile size
+        uint32_t crossover_medium = mmp->s->wg_denoms[1];
+        if ((ctx->device->mul_mat_id_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_id_s[src0_type]) {
             return aligned ? mmp->a_m : mmp->m;
         }
         return aligned ? mmp->a_s : mmp->s;
@@ -3919,14 +4087,15 @@ static void ggml_vk_matmul_id(
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
         uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
         uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
-        uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) {
+        uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11,
+        uint32_t padded_n) {
     VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " <<
         "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
         "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
         "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")");
     ggml_vk_sync_buffers(subctx);
     const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
-                                              nei0, nei1, nbi1, ne11 };
+                                              nei0, nei1, nbi1, ne11, padded_n };
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as });
 }
@@ -4088,15 +4257,17 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
-    const int x_ne = ne01 * ne00;
-    const int y_ne = ne11 * ne10;
-    const int d_ne = ne11 * ne01;
     const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
     vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type);
+    // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
+    uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
+    const int x_ne = ne01 * ne00;
+    const int y_ne = padded_n * ne10;
+    const int d_ne = ne11 * ne01;
     const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline);
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
@@ -4183,7 +4354,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
     if (qy_needs_dequant) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
+        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -4219,7 +4390,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
         ne01, ne11, ne10,
         ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
-        split_k, ne12*ne13, ne02, ne12, r2, r3
+        split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n
     );  // NOLINT
 }
@@ -4670,15 +4841,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne21 * ne20;
     const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
     vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type);
+    // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
+    uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = padded_n * ne10;
+    const uint64_t d_ne = ne21 * ne20;
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
@@ -4760,7 +4933,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
     if (qy_needs_dequant) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
+        GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -4797,7 +4970,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
         ne01, ne21, ne10, ne10, ne10, ne01,
         stride_batch_x, stride_batch_y, ne20*ne21,
-        n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11
+        n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
     );  // NOLINT
 }
@@ -5283,6 +5456,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONT:
     case GGML_OP_DUP:
         return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
+    case GGML_OP_SILU_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_silu_back_f32;
+        }
+        return nullptr;
     case GGML_OP_NORM:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_norm_f32;
@@ -5298,6 +5476,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_rms_norm_f32;
         }
         return nullptr;
+    case GGML_OP_RMS_NORM_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_rms_norm_back_f32;
+        }
+        return nullptr;
+    case GGML_OP_L2_NORM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_l2_norm_f32;
+        }
+        return nullptr;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(dst)) {
             case GGML_UNARY_OP_SILU:
@@ -5325,6 +5513,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                     return ctx->device->pipeline_tanh_f32;
                 }
                 break;
+            case GGML_UNARY_OP_SIGMOID:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_sigmoid_f32;
+                }
+                break;
             default:
                 break;
         }
@@ -5344,7 +5537,13 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
+    case GGML_OP_SOFT_MAX_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_soft_max_back_f32;
+        }
+        return nullptr;
     case GGML_OP_ROPE:
+    case GGML_OP_ROPE_BACK:
         {
             const int mode = ((const int32_t *) dst->op_params)[2];
             const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
@@ -5426,6 +5625,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_rwkv_wkv6_f32;
         }
         return nullptr;
+    case GGML_OP_RWKV_WKV7:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_rwkv_wkv7_f32;
+        }
+        return nullptr;
     case GGML_OP_OPT_STEP_ADAMW:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_opt_step_adamw_f32;
@@ -5672,7 +5876,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     switch (op) {
     case GGML_OP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
+    case GGML_OP_L2_NORM:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
     case GGML_OP_ARGMAX:
         {
@@ -5696,6 +5903,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         } break;
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_ROPE:
+    case GGML_OP_ROPE_BACK:
         elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
         break;
     case GGML_OP_GET_ROWS:
@@ -5791,7 +5999,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
-    } else if (op == GGML_OP_ROPE) {
+    } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
         vk_subbuffer subbuf_z;
         if (use_src2) {
@@ -5919,23 +6127,17 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
-static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, bool dryrun = false) {
-    const ggml_tensor * k = dst->src[0];
-    const ggml_tensor * v = dst->src[1];
-    const ggml_tensor * r = dst->src[2];
-    const ggml_tensor * tf = dst->src[3];
-    const ggml_tensor * td = dst->src[4];
-    const ggml_tensor * state = dst->src[5];
-    GGML_ASSERT(!ggml_is_quantized(k->type));
-    GGML_ASSERT(!ggml_is_quantized(v->type));
-    GGML_ASSERT(!ggml_is_quantized(r->type));
-    GGML_ASSERT(!ggml_is_quantized(tf->type));
-    GGML_ASSERT(!ggml_is_quantized(td->type));
-    GGML_ASSERT(!ggml_is_quantized(state->type));
+static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, int version, bool dryrun = false) {
+    GGML_ASSERT(version == 6 || version == 7);
+    int num_srcs = version == 6 ? 6 : 7;
+    for (int i = 0; i < num_srcs; i++) {
+        GGML_ASSERT(!ggml_is_quantized(dst->src[i]->type));
+    }
     GGML_ASSERT(dst->buffer != nullptr);
-    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, k, v, r, dst, GGML_OP_RWKV_WKV6);
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, dst->src[0], dst->src[1], dst->src[2], dst, dst->op);
     GGML_ASSERT(pipeline != nullptr);
     if (dryrun) {
@@ -5944,89 +6146,73 @@ static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subc
     }
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
-    ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
-    ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
-    ggml_backend_vk_buffer_context * r_buf_ctx = (ggml_backend_vk_buffer_context *)r->buffer->context;
-    ggml_backend_vk_buffer_context * tf_buf_ctx = (ggml_backend_vk_buffer_context *)tf->buffer->context;
-    ggml_backend_vk_buffer_context * td_buf_ctx = (ggml_backend_vk_buffer_context *)td->buffer->context;
-    ggml_backend_vk_buffer_context * state_buf_ctx = (ggml_backend_vk_buffer_context *)state->buffer->context;
+    ggml_backend_vk_buffer_context * src_buf_ctxs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
+    for (int i = 0; i < num_srcs; i++) {
+        src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context;
+    }
     ggml_vk_sync_buffers(subctx);
-    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;
+    vk_buffer d_D = nullptr, d_srcs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
+    size_t dst_offset = 0, src_offsets[7] = { 0, 0, 0, 0, 0, 0, 0 };
+    bool dst_uma = false, srcs_uma[7] = { false, false, false, false, false, false, false };
     if (ctx->device->uma) {
-        ggml_vk_host_get(ctx->device, k->data, d_K, k_offset);
-        ggml_vk_host_get(ctx->device, v->data, d_V, v_offset);
-        ggml_vk_host_get(ctx->device, r->data, d_R, r_offset);
-        ggml_vk_host_get(ctx->device, tf->data, d_TF, tf_offset);
-        ggml_vk_host_get(ctx->device, td->data, d_TD, td_offset);
-        ggml_vk_host_get(ctx->device, state->data, d_State, state_offset);
-        ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
+        for (int i = 0; i < num_srcs; i++) {
+            ggml_vk_host_get(ctx->device, dst->src[i]->data, d_srcs[i], src_offsets[i]);
+            srcs_uma[i] = d_srcs[i] != nullptr;
+        }
-        K_uma = d_K != nullptr;
-        V_uma = d_V != nullptr;
-        R_uma = d_R != nullptr;
-        TF_uma = d_TF != nullptr;
-        TD_uma = d_TD != nullptr;
-        STATE_uma = d_State != nullptr;
-        DST_uma = d_D != nullptr;
+        ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
+        dst_uma = d_D != nullptr;
     }
-    if (!K_uma) {
-        d_K = k_buf_ctx->dev_buffer;
-        k_offset = vk_tensor_offset(k) + k->view_offs;
-    }
-    if (!V_uma) {
-        d_V = v_buf_ctx->dev_buffer;
-        v_offset = vk_tensor_offset(v) + v->view_offs;
-    }
-    if (!R_uma) {
-        d_R = r_buf_ctx->dev_buffer;
-        r_offset = vk_tensor_offset(r) + r->view_offs;
-    }
-    if (!TF_uma) {
-        d_TF = tf_buf_ctx->dev_buffer;
-        tf_offset = vk_tensor_offset(tf) + tf->view_offs;
-    }
-    if (!TD_uma) {
-        d_TD = td_buf_ctx->dev_buffer;
-        td_offset = vk_tensor_offset(td) + td->view_offs;
-    }
-    if (!STATE_uma) {
-        d_State = state_buf_ctx->dev_buffer;
-        state_offset = vk_tensor_offset(state) + state->view_offs;
+    uint64_t src_sizes[7] = { 0, 0, 0, 0, 0, 0, 0 };
+    for (int i = 0; i < num_srcs; i++) {
+        src_sizes[i] = ggml_nbytes(dst->src[i]);
+        if (!srcs_uma[i]) {
+            d_srcs[i] = src_buf_ctxs[i]->dev_buffer;
+            src_offsets[i] = vk_tensor_offset(dst->src[i]) + dst->src[i]->view_offs;
+        }
     }
-    if (!DST_uma) {
+    const uint64_t dst_size = ggml_nbytes(dst);
+    if (!dst_uma) {
         d_D = dst_buf_ctx->dev_buffer;
         dst_offset = vk_tensor_offset(dst) + dst->view_offs;
     }
-    const uint64_t k_size = ggml_nbytes(k);
-    const uint64_t v_size = ggml_nbytes(v);
-    const uint64_t r_size = ggml_nbytes(r);
-    const uint64_t tf_size = ggml_nbytes(tf);
-    const uint64_t td_size = ggml_nbytes(td);
-    const uint64_t state_size = ggml_nbytes(state);
-    const uint64_t dst_size = ggml_nbytes(dst);
     std::array<uint32_t, 3> elements = {
         (uint32_t)(pc.B * pc.H),
         1,
         1
     };
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
-        vk_subbuffer{ d_K, k_offset, k_size },
-        vk_subbuffer{ d_V, v_offset, v_size },
-        vk_subbuffer{ d_R, r_offset, r_size },
-        vk_subbuffer{ d_TF, tf_offset, tf_size },
-        vk_subbuffer{ d_TD, td_offset, td_size },
-        vk_subbuffer{ d_State, state_offset, state_size },
-        vk_subbuffer{ d_D, dst_offset, dst_size }
-    }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
+    if (version == 6) {
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
+            vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
+            vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] },
+            vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] },
+            vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] },
+            vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
+            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
+            vk_subbuffer{ d_D, dst_offset, dst_size }
+        }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
+    } else if (version == 7) {
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
+            vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
+            vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] },
+            vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] },
+            vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] },
+            vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
+            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
+            vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] },
+            vk_subbuffer{ d_D, dst_offset, dst_size }
+        }, sizeof(vk_op_rwkv_wkv7_push_constants), &pc, elements);
+    } else {
+        // shouldn't happen
+        GGML_ASSERT(false);
+    }
 }
 static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
@@ -6035,7 +6221,7 @@ static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx,
     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(
+    ggml_vk_op_f32_wkv(
         ctx, subctx, dst,
         {
             (uint32_t)n_seqs,
@@ -6043,6 +6229,26 @@ static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx,
             (uint32_t)n_embed,
             (uint32_t)n_heads,
         },
+        6,
+        dryrun
+    );
+}
+static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+    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[1];
+    const size_t n_seqs = dst->src[6]->ne[1];
+    ggml_vk_op_f32_wkv(
+        ctx, subctx, dst,
+        {
+            (uint32_t)n_seqs,
+            (uint32_t)seq_length,
+            (uint32_t)n_embed,
+            (uint32_t)n_heads,
+        },
+        7,
         dryrun
     );
 }
@@ -6313,6 +6519,10 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
+static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+}
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     float * op_params = (float *)dst->op_params;
@@ -6335,6 +6545,16 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx,
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
 }
+static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+}
+static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+}
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
@@ -6370,7 +6590,12 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     }, dryrun);
 }
-static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], op_params[1] }, dryrun);
+}
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) {
     const int n_dims        = ((int32_t *) dst->op_params)[1];
     const int mode          = ((int32_t *) dst->op_params)[2];
     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
@@ -6398,7 +6623,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        sections[0], sections[1], sections[2], sections[3],
+        sections[0], sections[1], sections[2], sections[3], backprop
     }, dryrun);
 }
@@ -6719,7 +6944,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k),
             m, n, k,
             k, k, m, k*m, k*n, m*n,
-            split_k, batch, batch, batch, 1, 1
+            split_k, batch, batch, batch, 1, 1, n
         );
     }
     ggml_vk_ctx_end(subctx);
@@ -7064,7 +7289,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
             ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k),
             m, n, k,
             k, k, m, k*m, k*n, m*n,
-            split_k, batch, batch, batch, 1, 1
+            split_k, batch, batch, batch, 1, 1, n
         );
     }
     ggml_vk_ctx_end(subctx);
@@ -7295,6 +7520,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
+        case GGML_UNARY_OP_SIGMOID:
             break;
         default:
             return false;
@@ -7319,12 +7545,17 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
+    case GGML_OP_SILU_BACK:
     case GGML_OP_NORM:
     case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
+    case GGML_OP_L2_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_ROPE:
+    case GGML_OP_ROPE_BACK:
     case GGML_OP_MUL_MAT:
     case GGML_OP_MUL_MAT_ID:
     case GGML_OP_ARGSORT:
@@ -7336,6 +7567,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
     case GGML_OP_RWKV_WKV6:
+    case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_FLASH_ATTN_EXT:
     case GGML_OP_OPT_STEP_ADAMW:
@@ -7377,13 +7609,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_CPY:
         case GGML_OP_CONT:
         case GGML_OP_DUP:
+        case GGML_OP_SILU_BACK:
         case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_L2_NORM:
         case GGML_OP_UNARY:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ROPE:
+        case GGML_OP_ROPE_BACK:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
@@ -7475,6 +7712,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_DUP:
         ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun);
+        break;
+    case GGML_OP_SILU_BACK:
+        ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun);
         break;
     case GGML_OP_NORM:
         ggml_vk_norm(ctx, compute_ctx, src0, node, dryrun);
@@ -7487,6 +7728,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_RMS_NORM:
         ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun);
+        break;
+    case GGML_OP_RMS_NORM_BACK:
+        ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun);
+        break;
+    case GGML_OP_L2_NORM:
+        ggml_vk_l2_norm(ctx, compute_ctx, src0, node, dryrun);
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
@@ -7495,6 +7744,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
+        case GGML_UNARY_OP_SIGMOID:
             ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun);
             break;
         default:
@@ -7508,9 +7758,17 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_SOFT_MAX:
         ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun);
+        break;
+    case GGML_OP_SOFT_MAX_BACK:
+        ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node, dryrun);
         break;
     case GGML_OP_ROPE:
-        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun);
+        break;
+    case GGML_OP_ROPE_BACK:
+        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, true, dryrun);
         break;
     case GGML_OP_ARGSORT:
@@ -7568,6 +7826,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         break;
+    case GGML_OP_RWKV_WKV7:
+        ggml_vk_rwkv_wkv7(ctx, compute_ctx, node, dryrun);
+        break;
     case GGML_OP_OPT_STEP_ADAMW:
         ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun);
@@ -7636,12 +7899,17 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_CPY:
     case GGML_OP_CONT:
     case GGML_OP_DUP:
+    case GGML_OP_SILU_BACK:
     case GGML_OP_NORM:
     case GGML_OP_GROUP_NORM:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_RMS_NORM_BACK:
+    case GGML_OP_L2_NORM:
     case GGML_OP_DIAG_MASK_INF:
     case GGML_OP_SOFT_MAX:
+    case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_ROPE:
+    case GGML_OP_ROPE_BACK:
     case GGML_OP_RESHAPE:
     case GGML_OP_VIEW:
     case GGML_OP_PERMUTE:
@@ -7656,6 +7924,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
     case GGML_OP_RWKV_WKV6:
+    case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -7670,6 +7939,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
+        case GGML_UNARY_OP_SIGMOID:
             buf = tensor->buffer;
             break;
         default:
@@ -7844,11 +8114,12 @@ static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
     UNUSED(buffer);
 }
-static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+static enum ggml_status ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
     VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")");
     if (tensor->view_src != nullptr) {
         GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
     }
+    return GGML_STATUS_SUCCESS;
 }
 static void ggml_backend_vk_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
@@ -8165,8 +8436,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
     }
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
@@ -8187,17 +8462,27 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     bool first_node_in_batch = true; // true if next node will be first node in a batch
     int submit_node_idx = 0; // index to first node in a batch
-    // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution.
-    // Start with a smaller count to get work submitted right away, and increase it after each submit.
-    int nodes_per_submit = 20;
+    // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
+    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
+    // (and scaled down based on model size, so smaller models submit earlier).
+    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
+    int nodes_per_submit = 100;
     int submitted_nodes = 0;
     int submit_count = 0;
+    uint64_t mul_mat_bytes = 0;
+    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
-        bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
+        bool submit = (submitted_nodes >= nodes_per_submit) ||
+                      (mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+                      (i == last_node);
         bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
@@ -8214,13 +8499,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         if (submit) {
             first_node_in_batch = true;
             submitted_nodes = 0;
-            switch (submit_count) {
-            case 0:
-                nodes_per_submit = 50;
-                break;
-            default:
-                nodes_per_submit = 100;
-                break;
+            mul_mat_bytes = 0;
+            if (submit_count < 3) {
+                mul_mat_bytes_per_submit *= 2;
             }
             submit_count++;
         }
@@ -8371,7 +8652,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_TANH:
-                    return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_SIGMOID:
+                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
             }
@@ -8560,6 +8842,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_REPEAT_BACK:
             return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ROPE:
+        case GGML_OP_ROPE_BACK:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -8569,22 +8852,27 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_NORM:
         case GGML_OP_GROUP_NORM:
         case GGML_OP_RMS_NORM:
+        case GGML_OP_L2_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
-        case GGML_OP_CONCAT:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_SCALE:
+        case GGML_OP_SILU_BACK:
+        case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SQR:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ACC:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_SCALE:
         case GGML_OP_PAD:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
+        case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
@@ -8594,6 +8882,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
+        case GGML_OP_RWKV_WKV7:
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_OPT_STEP_ADAMW:
             return true;
@@ -8740,7 +9029,7 @@ 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) {
+static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
     switch (props.vendorID) {
     case VK_VENDOR_ID_INTEL:
         // Intel drivers don't support coopmat properly yet
@@ -8748,10 +9037,7 @@ static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDevicePrope
     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 arch == vk_device_architecture::AMD_RDNA3;
         }
         return true;
     default:
@@ -8976,15 +9262,25 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         tensor_clone = ggml_group_norm(ggml_ctx, src_clone[0], *(int *)tensor->op_params, ((float *)tensor->op_params)[1]);
     } else if (tensor->op == GGML_OP_RMS_NORM) {
         tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_RMS_NORM_BACK) {
+        const float eps = ((float *) tensor->op_params)[0];
+        tensor_clone = ggml_rms_norm_back(ggml_ctx, src_clone[0], src_clone[1], eps);
+    } else if (tensor->op == GGML_OP_SILU_BACK) {
+        tensor_clone = ggml_silu_back(ggml_ctx, src_clone[0], src_clone[1]);
+    } else if (tensor->op == GGML_OP_L2_NORM) {
+        const float eps = ((float *) tensor->op_params)[0];
+        tensor_clone = ggml_l2_norm(ggml_ctx, src_clone[0], eps);
     } else if (tensor->op == GGML_OP_SOFT_MAX) {
         if (src1 != nullptr) {
             tensor_clone = ggml_soft_max_ext(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
         } else {
             tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]);
         }
+    } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) {
+        tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
     } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
         tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], *(int *)tensor->op_params);
-    } else if (tensor->op == GGML_OP_ROPE) {
+    } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) {
         const int n_dims      = ((int32_t *) tensor->op_params)[1];
         const int mode        = ((int32_t *) tensor->op_params)[2];
         //const int n_ctx_ggml       = ((int32_t *) tensor->op_params)[3];
@@ -8997,9 +9293,17 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const float beta_slow       = ((float *) tensor->op_params)[10];
         if (mode & GGML_ROPE_TYPE_MROPE) {
             int32_t *sections = ((int32_t *) tensor->op_params) + 11;
-            tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            if (tensor->op == GGML_OP_ROPE) {
+                tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                tensor_clone = ggml_rope_multi_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
         } else {
-            tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            if (tensor->op == GGML_OP_ROPE) {
+                tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+            }
         }
     } else if (tensor->op == GGML_OP_UNARY) {
         switch (ggml_get_unary_op(tensor)) {
@@ -9018,6 +9322,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         case GGML_UNARY_OP_TANH:
             tensor_clone = ggml_tanh(ggml_ctx, src_clone[0]);
             break;
+        case GGML_UNARY_OP_SIGMOID:
+            tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]);
+            break;
         default:
             std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
             GGML_ABORT("fatal error");
@@ -9082,6 +9389,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     } else if (tensor->op == GGML_OP_RWKV_WKV6) {
         tensor_clone = ggml_rwkv_wkv6(ggml_ctx, src_clone[0], src_clone[1],
         src_clone[2], src_clone[3], src_clone[4], src_clone[5]);
+    } else if (tensor->op == GGML_OP_RWKV_WKV7) {
+        tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3],
+        src_clone[4], src_clone[5], src_clone[6]);
     } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) {
         src_clone[0]->flags = src0->flags;
         tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],