whispercpp 1.3.4 → 1.3.5

data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -7,34 +7,110 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-impl.h"
+#include "ggml-webgpu-shader-lib.hpp"
 #include "ggml-wgsl-shaders.hpp"
+#include "pre_wgsl.hpp"
+
+#ifdef __EMSCRIPTEN__
+#    include <emscripten/emscripten.h>
+#endif
 
 #include <webgpu/webgpu_cpp.h>
 
+#include <atomic>
 #include <condition_variable>
+#include <cstdint>
 #include <cstring>
 #include <iostream>
+#include <map>
 #include <mutex>
+#include <optional>
 #include <string>
 #include <vector>
 
+#define ROUNDUP_POW2(x, pow2) (((x) + ((pow2) - 1)) & ~((pow2) - 1))
+#define CEIL_DIV(M, N)        (((M) + (N) - 1) / (N))
+
 #ifdef GGML_WEBGPU_DEBUG
 #    define WEBGPU_LOG_DEBUG(msg)  std::cout << msg << std::endl
-#    define WEBGPU_DEBUG_BUF_ELEMS 32
+#    define WEBGPU_DEBUG_BUF_ELEMS 512
 #else
 #    define WEBGPU_LOG_DEBUG(msg) ((void) 0)
 #endif  // GGML_WEBGPU_DEBUG
 
+#ifdef GGML_WEBGPU_CPU_PROFILE
+// total timing (aggregated)
+#    define WEBGPU_CPU_PROFILE_TOTAL_START(id) auto cpu_total_start_##id = std::chrono::high_resolution_clock::now();
+
+#    define WEBGPU_CPU_PROFILE_TOTAL_END(id, ctx)                                                         \
+        auto   cpu_total_end_##id = std::chrono::high_resolution_clock::now();                            \
+        double cpu_total_time_##id =                                                                      \
+            std::chrono::duration<double, std::milli>(cpu_total_end_##id - cpu_total_start_##id).count(); \
+        (ctx)->cpu_time_ms[#id] += cpu_total_time_##id;
+
+// fine-grained timing (not included in totals)
+#    define WEBGPU_CPU_PROFILE_DETAIL_START(id) auto cpu_detail_start_##id = std::chrono::high_resolution_clock::now();
+
+#    define WEBGPU_CPU_PROFILE_DETAIL_END(id, ctx)                                                          \
+        auto   cpu_detail_end_##id = std::chrono::high_resolution_clock::now();                             \
+        double cpu_detail_time_##id =                                                                       \
+            std::chrono::duration<double, std::milli>(cpu_detail_end_##id - cpu_detail_start_##id).count(); \
+        (ctx)->cpu_detail_ms[#id] += cpu_detail_time_##id;
+#else
+#    define WEBGPU_CPU_PROFILE_TOTAL_START(id)
+#    define WEBGPU_CPU_PROFILE_TOTAL_END(id, ctx)
+#    define WEBGPU_CPU_PROFILE_DETAIL_START(id)
+#    define WEBGPU_CPU_PROFILE_DETAIL_END(id, ctx)
+#endif  // GGML_WEBGPU_CPU_PROFILE
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+#    define WEBGPU_NUM_TIMESTAMP_QUERY_BUFS       24
+#    define WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES 16  // e.g. enough for two timestamps
+#endif
+
 /* Constants */
 
-#define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     16
-#define WEBGPU_MUL_MAT_WG_SIZE               64
-#define WEBGPU_NUM_PARAM_BUFS                100
+// Track https://github.com/gpuweb/gpuweb/issues/5315 for fixes to implementations so this can be removed.
+#define WEBGPU_MAX_WG_SIZE 288
+
+#define WEBGPU_MUL_MAT_WG_SIZE               256
+#define WEBGPU_NUM_PARAM_BUFS                32u
+#define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     8u
+#define WEBGPU_WAIT_ANY_TIMEOUT_MS           0
+// Maximum number of in-flight submissions per-thread, to avoid exhausting the parameter buffer pool
+#define WEBGPU_MAX_INFLIGHT_SUBS_PER_THREAD  WEBGPU_NUM_PARAM_BUFS / WEBGPU_COMMAND_SUBMIT_BATCH_SIZE
 #define WEBGPU_PARAMS_BUF_SIZE_BYTES         128  // enough for 32 parameters
 #define WEBGPU_NUM_SET_ROWS_ERROR_BUFS       32
 #define WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES 4
 #define WEBGPU_STORAGE_BUF_BINDING_MULT      4  // a storage buffer binding size must be a multiple of 4
 
+// For operations which process a row in parallel, this seems like a reasonable default
+#define WEBGPU_ROW_SPLIT_WG_SIZE 64
+
+// Matrix multiplication parameters
+
+// Register tiling parameters
+#define WEBGPU_MUL_MAT_TILE_M    8
+#define WEBGPU_MUL_MAT_TILE_N    8
+#define WEBGPU_MUL_MAT_WG_SIZE_M 8
+#define WEBGPU_MUL_MAT_WG_SIZE_N 8
+#define WEBGPU_MUL_MAT_TILE_K    32
+
+// Subgroup matrix parameters
+// The number of subgroups in the M dimension
+#define WEBGPU_MUL_MAT_SUBGROUP_M        2
+// The number of subgroups in the N dimension
+#define WEBGPU_MUL_MAT_SUBGROUP_N        2
+// The number of subgroup matrices each subgroup accumulates over
+#define WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M 4
+#define WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N 2
+
+// Matrix-vector multiplication parameters
+#define WEBGPU_MUL_MAT_VEC_WG_SIZE        256
+// Must be multiple of 4 to work with vectorized paths, and must divide mul_mat_vec wg size
+#define WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG 64
+#define WEBGPU_MUL_MAT_VEC_TILE_K         256
+
 /* End Constants */
 
 // This is a "fake" base pointer, since WebGPU buffers do not have pointers to their locations.
@@ -62,6 +138,11 @@ struct webgpu_pool_bufs {
     wgpu::Buffer dev_buf;
 };
 
+// The futures to wait on for a single queue submission
+struct webgpu_submission_futures {
+    std::vector<wgpu::FutureWaitInfo> futures;
+};
+
 // Holds a pool of parameter buffers for WebGPU operations
 struct webgpu_buf_pool {
     std::vector<webgpu_pool_bufs> free;
@@ -108,6 +189,124 @@ struct webgpu_buf_pool {
     }
 };
 
+#ifdef GGML_WEBGPU_GPU_PROFILE
+struct webgpu_gpu_profile_bufs {
+    wgpu::Buffer   host_buf;
+    wgpu::Buffer   dev_buf;
+    wgpu::QuerySet query_set;
+};
+
+// Holds a pool of parameter buffers for WebGPU operations
+struct webgpu_gpu_profile_buf_pool {
+    std::vector<webgpu_gpu_profile_bufs> free;
+
+    std::mutex mutex;
+
+    std::condition_variable cv;
+
+    void init(wgpu::Device      device,
+              int               num_bufs,
+              size_t            buf_size,
+              wgpu::BufferUsage dev_buf_usage,
+              wgpu::BufferUsage host_buf_usage) {
+        for (int i = 0; i < num_bufs; i++) {
+            wgpu::Buffer host_buf;
+            wgpu::Buffer dev_buf;
+            ggml_webgpu_create_buffer(device, host_buf, buf_size, host_buf_usage, "ggml_webgpu_host_profile_buf");
+            ggml_webgpu_create_buffer(device, dev_buf, buf_size, dev_buf_usage, "ggml_webgpu_dev_profile_buf");
+            // Create a query set for 2 timestamps
+            wgpu::QuerySetDescriptor ts_query_set_desc = {};
+
+            ts_query_set_desc.type      = wgpu::QueryType::Timestamp;
+            ts_query_set_desc.count     = 2;
+            wgpu::QuerySet ts_query_set = device.CreateQuerySet(&ts_query_set_desc);
+
+            free.push_back({ host_buf, dev_buf, ts_query_set });
+        }
+    }
+
+    webgpu_gpu_profile_bufs alloc_bufs() {
+        std::unique_lock<std::mutex> lock(mutex);
+        cv.wait(lock, [this] { return !free.empty(); });
+        webgpu_gpu_profile_bufs bufs = free.back();
+        free.pop_back();
+        return bufs;
+    }
+
+    void free_bufs(std::vector<webgpu_gpu_profile_bufs> bufs) {
+        std::lock_guard<std::mutex> lock(mutex);
+        free.insert(free.end(), bufs.begin(), bufs.end());
+        cv.notify_all();
+    }
+
+    void cleanup() {
+        std::lock_guard<std::mutex> lock(mutex);
+        for (auto & bufs : free) {
+            bufs.host_buf.Destroy();
+            bufs.dev_buf.Destroy();
+            bufs.query_set.Destroy();
+        }
+        free.clear();
+    }
+};
+#endif
+
+struct webgpu_pipeline {
+    wgpu::ComputePipeline pipeline;
+    std::string           name;
+    void *                context = nullptr;
+};
+
+struct webgpu_command {
+    wgpu::CommandBuffer             commands;
+    webgpu_pool_bufs                params_bufs;
+    std::optional<webgpu_pool_bufs> set_rows_error_bufs;
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    webgpu_gpu_profile_bufs timestamp_query_bufs;
+    std::string             pipeline_name;
+#endif
+};
+
+struct flash_attn_pipeline_key {
+    int      q_type;
+    int      kv_type;
+    int      dst_type;
+    uint32_t head_dim_qk;
+    uint32_t head_dim_v;
+    bool     kv_direct;
+    bool     has_mask;
+    bool     has_sinks;
+    bool     uses_logit_softcap;
+
+    bool operator==(const flash_attn_pipeline_key & other) const {
+        return q_type == other.q_type && kv_type == other.kv_type && dst_type == other.dst_type &&
+               head_dim_qk == other.head_dim_qk && head_dim_v == other.head_dim_v && kv_direct == other.kv_direct &&
+               has_mask == other.has_mask && has_sinks == other.has_sinks &&
+               uses_logit_softcap == other.uses_logit_softcap;
+    }
+};
+
+// Same hash combine function as in boost
+template <typename T> inline void ggml_webgpu_hash_combine(size_t & seed, const T & value) {
+    seed ^= std::hash<T>{}(value) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+}
+
+struct flash_attn_pipeline_key_hash {
+    size_t operator()(const flash_attn_pipeline_key & key) const {
+        size_t seed = 0;
+        ggml_webgpu_hash_combine(seed, key.q_type);
+        ggml_webgpu_hash_combine(seed, key.kv_type);
+        ggml_webgpu_hash_combine(seed, key.dst_type);
+        ggml_webgpu_hash_combine(seed, key.head_dim_qk);
+        ggml_webgpu_hash_combine(seed, key.head_dim_v);
+        ggml_webgpu_hash_combine(seed, key.kv_direct);
+        ggml_webgpu_hash_combine(seed, key.has_mask);
+        ggml_webgpu_hash_combine(seed, key.has_sinks);
+        ggml_webgpu_hash_combine(seed, key.uses_logit_softcap);
+        return seed;
+    }
+};
+
 // All the base objects needed to run operations on a WebGPU device
 struct webgpu_context_struct {
     wgpu::Instance instance;
@@ -116,47 +315,68 @@ struct webgpu_context_struct {
     wgpu::Queue    queue;
     wgpu::Limits   limits;
 
-    // Separate this out from limits since on some Metal systems, the limit returned by
-    // querying the limits is higher than the actual allowed maximum.
-    uint32_t max_wg_size_x;
+    uint32_t max_subgroup_size;
+
+    bool     supports_subgroup_matrix = false;
+    uint32_t sg_mat_m;
+    uint32_t sg_mat_n;
+    uint32_t sg_mat_k;
 
     std::recursive_mutex mutex;
+    std::atomic_uint     inflight_threads = 0;
 
     webgpu_buf_pool param_buf_pool;
     webgpu_buf_pool set_rows_error_buf_pool;
 
-    wgpu::ComputePipeline memset_pipeline;
-    wgpu::ComputePipeline mul_mat_pipeline[30][2];
-    wgpu::ComputePipeline set_rows_pipeline;
-    wgpu::ComputePipeline get_rows_pipeline[30];
-    wgpu::ComputePipeline get_rows_f32_no_vec_pipeline;
-    wgpu::ComputePipeline cpy_pipeline;
-    wgpu::ComputePipeline add_pipeline[2];
-    wgpu::ComputePipeline add_ip_pipeline[2];
-    wgpu::ComputePipeline mul_pipeline[2];
-    wgpu::ComputePipeline mul_ip_pipeline[2];
-    wgpu::ComputePipeline rms_norm_pipeline;
-    wgpu::ComputePipeline rms_norm_ip_pipeline;
+    pre_wgsl::Preprocessor p;
 
-    size_t memset_bytes_per_thread;
+    std::map<int, webgpu_pipeline> memset_pipelines;                                 // variant or type index
 
-    // Staging buffer for reading data from the GPU
-    wgpu::Buffer get_tensor_staging_buf;
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> mul_mat_pipelines;  // src0_type, src1_type, vectorized
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>>
+        mul_mat_vec_pipelines;                                                       // src0_type, src1_type, vectorized
+
+    std::unordered_map<flash_attn_pipeline_key, webgpu_pipeline, flash_attn_pipeline_key_hash> flash_attn_pipelines;
 
-    // Command buffers which need to be submitted
-    std::vector<wgpu::CommandBuffer> staged_command_bufs;
+    std::map<int, std::map<int, webgpu_pipeline>> set_rows_pipelines;                 // dst_type, vectorized
+    std::map<int, std::map<int, webgpu_pipeline>> get_rows_pipelines;                 // src_type, vectorized
 
-    // Parameter buffers associated with the staged command buffers
-    std::vector<webgpu_pool_bufs> staged_param_bufs;
-    // Buffers associated with set_rows operations, used to store potential errors
-    std::vector<webgpu_pool_bufs> staged_set_row_error_bufs;
+    std::map<int, std::map<int, webgpu_pipeline>> cpy_pipelines;                      // src_type, dst_type
+    std::map<int, std::map<int, webgpu_pipeline>> add_pipelines;                      // type, inplace
+    std::map<int, std::map<int, webgpu_pipeline>> sub_pipelines;                      // type, inplace
+    std::map<int, std::map<int, webgpu_pipeline>> mul_pipelines;                      // type, inplace
+    std::map<int, std::map<int, webgpu_pipeline>> div_pipelines;                      // type, inplace
 
-    std::vector<wgpu::FutureWaitInfo> callback_futures;
+    std::map<int, webgpu_pipeline>                               rms_norm_pipelines;  // inplace
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> rope_pipelines;      // type, ff, inplace
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> glu_pipelines;       // glu_op, type, split
+    std::map<int, webgpu_pipeline>                               scale_pipelines;     // inplace
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> soft_max_pipelines;  // mask_type, has_sink, inplace
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> unary_pipelines;     // unary_op, type, inplace
+
+    size_t memset_bytes_per_thread;
+
+    // Staging buffer for reading data from the GPU
+    wgpu::Buffer get_tensor_staging_buf;
 
 #ifdef GGML_WEBGPU_DEBUG
     wgpu::Buffer debug_host_buf;
     wgpu::Buffer debug_dev_buf;
 #endif
+
+#ifdef GGML_WEBGPU_CPU_PROFILE
+    // Profiling: labeled CPU time in ms (total)
+    std::unordered_map<std::string, double> cpu_time_ms;
+    // Profiling: detailed CPU time in ms
+    std::unordered_map<std::string, double> cpu_detail_ms;
+#endif
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // Profiling: per-shader GPU time in ms
+    std::unordered_map<std::string, double> shader_gpu_time_ms;
+    // Profiling: pool of timestamp query buffers (one per operation)
+    webgpu_gpu_profile_buf_pool             timestamp_query_buf_pool;
+#endif
 };
 
 typedef std::shared_ptr<webgpu_context_struct> webgpu_context;
@@ -181,23 +401,39 @@ struct ggml_backend_webgpu_context {
 struct ggml_backend_webgpu_buffer_context {
     webgpu_context webgpu_ctx;
     wgpu::Buffer   buffer;
+    std::string    label;
 
-    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf) :
+    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf, std::string lbl) :
         webgpu_ctx(std::move(ctx)),
-        buffer(std::move(buf)) {}
+        buffer(std::move(buf)),
+        label(std::move(lbl)) {}
 };
 
-/* End struct definitions */
-
 /* WebGPU object initializations */
 
-static void ggml_webgpu_create_pipeline(wgpu::Device &                           device,
-                                        wgpu::ComputePipeline &                  pipeline,
-                                        const char *                             shader_code,
-                                        const char *                             label,
-                                        const std::vector<wgpu::ConstantEntry> & constants = {}) {
-    WEBGPU_LOG_DEBUG("ggml_webgpu_create_pipeline()");
+// Process a WGSL shader string, replacing tokens of the form {{KEY}} with
+// the corresponding values provided in `repls`.
+static std::string ggml_webgpu_process_shader_repls(const char *                               src,
+                                                    const std::map<std::string, std::string> & repls) {
+    if (!src) {
+        return std::string();
+    }
+    std::string s = src;
+    for (const auto & kv : repls) {
+        std::string token = "{{" + kv.first + "}}";
+        size_t      pos   = 0;
+        while ((pos = s.find(token, pos)) != std::string::npos) {
+            s.replace(pos, token.length(), kv.second);
+            pos += kv.second.length();
+        }
+    }
+    return s;
+}
 
+static webgpu_pipeline ggml_webgpu_create_pipeline(wgpu::Device &                           device,
+                                                   const char *                             shader_code,
+                                                   const char *                             label,
+                                                   const std::vector<wgpu::ConstantEntry> & constants = {}) {
     wgpu::ShaderSourceWGSL shader_source;
     shader_source.code = shader_code;
 
@@ -215,7 +451,7 @@ static void ggml_webgpu_create_pipeline(wgpu::Device &
         pipeline_desc.compute.constants     = constants.data();
         pipeline_desc.compute.constantCount = constants.size();
     }
-    pipeline = device.CreateComputePipeline(&pipeline_desc);
+    return { device.CreateComputePipeline(&pipeline_desc), label };
 }
 
 static void ggml_webgpu_create_buffer(wgpu::Device &    device,
@@ -223,8 +459,6 @@ static void ggml_webgpu_create_buffer(wgpu::Device &    device,
                                       size_t            size,
                                       wgpu::BufferUsage usage,
                                       const char *      label) {
-    WEBGPU_LOG_DEBUG("ggml_webgpu_create_buffer()");
-
     wgpu::BufferDescriptor buffer_desc;
     buffer_desc.size             = size;
     buffer_desc.usage            = usage;
@@ -240,79 +474,35 @@ static void ggml_webgpu_create_buffer(wgpu::Device &    device,
 /** WebGPU Actions */
 
 // Wait for the queue to finish processing all submitted work
-static void ggml_backend_webgpu_wait_on_submission(webgpu_context & ctx) {
-    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
-    if (ctx->callback_futures.empty()) {
-        // no existing callbacks, wait on queue submission
-        ctx->instance.WaitAny(
-            ctx->queue.OnSubmittedWorkDone(wgpu::CallbackMode::AllowSpontaneous,
-                                           [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
-                                               if (status != wgpu::QueueWorkDoneStatus::Success) {
-                                                   GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n",
-                                                                  std::string(message).c_str());
-                                               }
-                                           }),
-            UINT64_MAX);
-    } else {
-        // existing callbacks, wait on them
-        ctx->instance.WaitAny(ctx->callback_futures.size(), ctx->callback_futures.data(), UINT64_MAX);
-        ctx->callback_futures.clear();
+static void ggml_backend_webgpu_wait(webgpu_context &                         ctx,
+                                     std::vector<webgpu_submission_futures> & futures,
+                                     bool                                     block = true) {
+    // If we have too many in-flight submissions, wait on the oldest one first. If there are many threads,
+    // inflight_max may be 0, meaning that we must wait on all futures.
+    uint64_t timeout_ms       = block ? UINT64_MAX : 0;
+    uint32_t inflight_threads = ctx->inflight_threads;
+    uint32_t inflight_max     = WEBGPU_MAX_INFLIGHT_SUBS_PER_THREAD / std::max(inflight_threads, 1u);
+    while (futures.size() >= inflight_max && futures.size() > 0) {
+        ctx->instance.WaitAny(futures[0].futures.size(), futures[0].futures.data(), UINT64_MAX);
+        futures.erase(futures.begin());
     }
-}
-
-static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
-    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_submit_queue()");
-    if (ctx->staged_command_bufs.empty()) {
-        // Nothing to submit
-        return;
-    }
-    ctx->queue.Submit(ctx->staged_command_bufs.size(), ctx->staged_command_bufs.data());
-
-    // If there are SET_ROWS operations in this submission, copy their error buffers to the host.
-    if (ctx->staged_set_row_error_bufs.size() > 0) {
-        wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
-        for (auto & error_bufs : ctx->staged_set_row_error_bufs) {
-            // Copy the error buffer to the host buffer
-            encoder.CopyBufferToBuffer(error_bufs.dev_buf, 0, error_bufs.host_buf, 0, error_bufs.host_buf.GetSize());
+    size_t i = 0;
+    while (i < futures.size()) {
+        auto waitStatus = ctx->instance.WaitAny(futures[i].futures.size(), futures[i].futures.data(), timeout_ms);
+        switch (waitStatus) {
+            case wgpu::WaitStatus::Success:
+                futures.erase(futures.begin() + i);
+                break;
+            case wgpu::WaitStatus::TimedOut:
+                i++;
+                break;
+            case wgpu::WaitStatus::Error:
+                GGML_LOG_ERROR("ggml_webgpu: WaitAny returned an error\n");
+                break;
+            default:
+                GGML_LOG_ERROR("ggml_webgpu: WaitAny returned an unknown status\n");
+                break;
         }
-        wgpu::CommandBuffer commands = encoder.Finish();
-        ctx->queue.Submit(1, &commands);
-    }
-
-    ctx->staged_command_bufs.clear();
-    std::vector<webgpu_pool_bufs> staged_param_bufs         = std::move(ctx->staged_param_bufs);
-    std::vector<webgpu_pool_bufs> staged_set_row_error_bufs = std::move(ctx->staged_set_row_error_bufs);
-
-    // Free the staged parameter buffers once the submission completes
-    wgpu::Future p_f = ctx->queue.OnSubmittedWorkDone(
-        wgpu::CallbackMode::AllowSpontaneous,
-        [ctx, staged_param_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
-            if (status != wgpu::QueueWorkDoneStatus::Success) {
-                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
-            }
-            // Free the staged buffers
-            ctx->param_buf_pool.free_bufs(staged_param_bufs);
-        });
-    ctx->callback_futures.push_back({ p_f });
-
-    // Check for errrors in SET_ROWS operations
-    for (auto & error_bufs : staged_set_row_error_bufs) {
-        wgpu::Future f = error_bufs.host_buf.MapAsync(
-            wgpu::MapMode::Read, 0, error_bufs.host_buf.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
-            [ctx, error_bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
-                if (status != wgpu::MapAsyncStatus::Success) {
-                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str());
-                } else {
-                    const uint32_t * error_data = (const uint32_t *) error_bufs.host_buf.GetConstMappedRange();
-                    if (*error_data) {
-                        GGML_ABORT("ggml_webgpu: SET_ROWS index > 2^32, unsupported.");
-                    }
-                    // We can't unmap in here due to WebGPU reentrancy limitations.
-                    ctx->set_rows_error_buf_pool.free_bufs({ error_bufs });
-                }
-            });
-        ctx->callback_futures.push_back({ f });
     }
 }
 
@@ -336,30 +526,97 @@ static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
 // To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and
 // debug statements in the shader, and then call this function after encoding the commands and submitting them.
 static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
-    ggml_backend_webgpu_submit_queue(ctx);
     wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
     encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize());
     wgpu::CommandBuffer commands = encoder.Finish();
     ctx->queue.Submit(1, &commands);
-
     ggml_backend_webgpu_map_buffer(ctx, ctx->debug_host_buf, wgpu::MapMode::Read, 0, ctx->debug_host_buf.GetSize());
-    const uint32_t * debug_data = (const uint32_t *) ctx->debug_host_buf.GetConstMappedRange();
-    std::cout << "debug data:";
-    for (size_t i = 0; i < WEBGPU_DEBUG_BUF_ELEMS; i++) {
-        std::cout << "  " << i << ": " << debug_data[i];
-    }
-    std::cout << "\n";
+    const float * debug_data = (const float *) ctx->debug_host_buf.GetConstMappedRange();
+    std::cout << "debug[0]: " << debug_data[0] << "\n";
     ctx->debug_host_buf.Unmap();
 }
 #endif
 
-static void ggml_backend_webgpu_build_and_enqueue(webgpu_context &                  ctx,
-                                                  wgpu::ComputePipeline &           pipeline,
-                                                  std::vector<uint32_t>             params,
-                                                  std::vector<wgpu::BindGroupEntry> bind_group_entries,
-                                                  uint32_t                          wg_x,
-                                                  const char *                      bind_group_label = nullptr,
-                                                  bool                              submit_and_wait  = false) {
+static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, std::vector<webgpu_command> commands) {
+    std::vector<wgpu::CommandBuffer> command_buffers;
+    std::vector<webgpu_pool_bufs>    params_bufs;
+    std::vector<webgpu_pool_bufs>    set_rows_error_bufs;
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    std::vector<std::pair<std::string, webgpu_gpu_profile_bufs>> pipeline_name_and_ts_bufs;
+#endif
+
+    for (const auto & command : commands) {
+        command_buffers.push_back(command.commands);
+        params_bufs.push_back(command.params_bufs);
+        if (command.set_rows_error_bufs) {
+            set_rows_error_bufs.push_back(command.set_rows_error_bufs.value());
+        }
+    }
+    ctx->queue.Submit(command_buffers.size(), command_buffers.data());
+
+    std::vector<wgpu::FutureWaitInfo> futures;
+
+    wgpu::Future p_f = ctx->queue.OnSubmittedWorkDone(
+        wgpu::CallbackMode::AllowSpontaneous,
+        [ctx, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+            if (status != wgpu::QueueWorkDoneStatus::Success) {
+                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
+            }
+            // Free the staged buffers
+            ctx->param_buf_pool.free_bufs({ params_bufs });
+        });
+    futures.push_back({ p_f });
+
+    for (const auto & bufs : set_rows_error_bufs) {
+        wgpu::Future f = bufs.host_buf.MapAsync(
+            wgpu::MapMode::Read, 0, bufs.host_buf.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
+            [ctx, bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+                if (status != wgpu::MapAsyncStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str());
+                } else {
+                    const uint32_t * error_data = (const uint32_t *) bufs.host_buf.GetConstMappedRange();
+                    if (*error_data) {
+                        GGML_ABORT("ggml_webgpu: SET_ROWS index > 2^32, unsupported.");
+                    }
+                    // We can't unmap in here due to WebGPU reentrancy limitations.
+                    ctx->set_rows_error_buf_pool.free_bufs({ bufs });
+                }
+            });
+        futures.push_back({ f });
+    }
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    for (const auto & command : commands) {
+        auto label   = command.pipeline_name;
+        auto ts_bufs = command.timestamp_query_bufs;
+
+        wgpu::Future f = ts_bufs.host_buf.MapAsync(
+            wgpu::MapMode::Read, 0, ts_bufs.host_buf.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
+            [ctx, ts_bufs, label](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+                if (status != wgpu::MapAsyncStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to map timestamp buffer: %s\n", std::string(message).c_str());
+                } else {
+                    const uint64_t * ts_data    = (const uint64_t *) ts_bufs.host_buf.GetConstMappedRange();
+                    // WebGPU timestamps are in ns; convert to ms
+                    double           elapsed_ms = double(ts_data[1] - ts_data[0]) * 1e-6;
+                    ctx->shader_gpu_time_ms[label] += elapsed_ms;
+                    // We can't unmap in here due to WebGPU reentrancy limitations.
+                    ctx->timestamp_query_buf_pool.free_bufs({ ts_bufs });
+                }
+            });
+        futures.push_back({ f });
+    }
+#endif
+    return { futures };
+}
+
+static webgpu_command ggml_backend_webgpu_build(webgpu_context &                  ctx,
+                                                webgpu_pipeline &                 pipeline,
+                                                std::vector<uint32_t>             params,
+                                                std::vector<wgpu::BindGroupEntry> bind_group_entries,
+                                                uint32_t                          wg_x,
+                                                uint32_t                          wg_y                = 1,
+                                                std::optional<webgpu_pool_bufs>   set_rows_error_bufs = std::nullopt) {
     webgpu_pool_bufs params_bufs = ctx->param_buf_pool.alloc_bufs();
 
     ggml_backend_webgpu_map_buffer(ctx, params_bufs.host_buf, wgpu::MapMode::Write, 0, params_bufs.host_buf.GetSize());
@@ -377,44 +634,58 @@ static void ggml_backend_webgpu_build_and_enqueue(webgpu_context &
                                    .size    = params_bufs.dev_buf.GetSize() });
 
     wgpu::BindGroupDescriptor bind_group_desc;
-    bind_group_desc.layout     = pipeline.GetBindGroupLayout(0);
+    bind_group_desc.layout     = pipeline.pipeline.GetBindGroupLayout(0);
     bind_group_desc.entryCount = bind_group_entries.size();
     bind_group_desc.entries    = bind_group_entries.data();
-    if (bind_group_label) {
-        bind_group_desc.label = bind_group_label;
-    }
+    bind_group_desc.label      = pipeline.name.c_str();
     wgpu::BindGroup bind_group = ctx->device.CreateBindGroup(&bind_group_desc);
 
     wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
     encoder.CopyBufferToBuffer(params_bufs.host_buf, 0, params_bufs.dev_buf, 0, params_bufs.dev_buf.GetSize());
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // --- Profiling: GPU timestamp queries ---
+    // Allocate a timestamp query buffer (2 timestamps: start/end)
+    webgpu_gpu_profile_bufs ts_bufs = ctx->timestamp_query_buf_pool.alloc_bufs();
+    if (ts_bufs.host_buf.GetMapState() == wgpu::BufferMapState::Mapped) {
+        ts_bufs.host_buf.Unmap();
+    }
+
+    wgpu::PassTimestampWrites   ts_writes = { .querySet                  = ts_bufs.query_set,
+                                              .beginningOfPassWriteIndex = 0,
+                                              .endOfPassWriteIndex       = 1 };
+    wgpu::ComputePassDescriptor pass_desc = { .timestampWrites = &ts_writes };
+    wgpu::ComputePassEncoder    pass      = encoder.BeginComputePass(&pass_desc);
+#else
     wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
-    pass.SetPipeline(pipeline);
+#endif
+    pass.SetPipeline(pipeline.pipeline);
     pass.SetBindGroup(0, bind_group);
-    pass.DispatchWorkgroups(wg_x, 1, 1);
+    pass.DispatchWorkgroups(wg_x, wg_y, 1);
     pass.End();
-    wgpu::CommandBuffer commands = encoder.Finish();
-    if (submit_and_wait) {
-        // Submit and wait immediately
-        ctx->queue.Submit(1, &commands);
-        ctx->instance.WaitAny(ctx->queue.OnSubmittedWorkDone(
-                                  wgpu::CallbackMode::AllowSpontaneous,
-                                  [ctx, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
-                                      if (status != wgpu::QueueWorkDoneStatus::Success) {
-                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
-                                      }
-                                      ctx->param_buf_pool.free_bufs({ params_bufs });
-                                  }),
-                              UINT64_MAX);
-    } else {
-        // Lock the context mutex when pushing to the staging vectors.
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
-        // Enqueue commands and only submit if we have enough staged commands
-        ctx->staged_command_bufs.push_back(commands);
-        ctx->staged_param_bufs.push_back(params_bufs);
-        if (ctx->staged_command_bufs.size() == WEBGPU_COMMAND_SUBMIT_BATCH_SIZE) {
-            ggml_backend_webgpu_submit_queue(ctx);
-        }
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // Resolve the query set into the device buffer
+    encoder.ResolveQuerySet(ts_bufs.query_set, 0, 2, ts_bufs.dev_buf, 0);
+    encoder.CopyBufferToBuffer(ts_bufs.dev_buf, 0, ts_bufs.host_buf, 0, ts_bufs.host_buf.GetSize());
+#endif
+
+    // If there are SET_ROWS operations in this submission, copy their error buffers to the host.
+    if (set_rows_error_bufs) {
+        encoder.CopyBufferToBuffer(set_rows_error_bufs->dev_buf, 0, set_rows_error_bufs->host_buf, 0,
+                                   set_rows_error_bufs->host_buf.GetSize());
     }
+
+    wgpu::CommandBuffer commands = encoder.Finish();
+    webgpu_command      result   = {};
+    result.commands              = commands;
+    result.params_bufs           = params_bufs;
+    result.set_rows_error_bufs   = set_rows_error_bufs;
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    result.timestamp_query_bufs = ts_bufs;
+    result.pipeline_name        = pipeline.name;
+#endif
+    return result;
 }
 
 static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
@@ -426,9 +697,12 @@ static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0, .buffer = buf, .offset = 0, .size = buf.GetSize() }
     };
-    size_t   bytes_per_wg = ctx->max_wg_size_x * ctx->memset_bytes_per_thread;
-    uint32_t wg_x         = ((size + 3) + bytes_per_wg - 1) / bytes_per_wg;
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->memset_pipeline, params, entries, wg_x, "MEMSET", true);
+    size_t   bytes_per_wg = WEBGPU_MAX_WG_SIZE * ctx->memset_bytes_per_thread;
+    uint32_t wg_x         = CEIL_DIV(size + 3, bytes_per_wg);
+
+    webgpu_command command = ggml_backend_webgpu_build(ctx, ctx->memset_pipelines[0], params, entries, wg_x);
+    std::vector<webgpu_submission_futures> futures = { ggml_backend_webgpu_submit(ctx, { command }) };
+    ggml_backend_webgpu_wait(ctx, futures);
 }
 
 /** End WebGPU Actions */
@@ -440,12 +714,53 @@ static const char * ggml_backend_webgpu_name(ggml_backend_t backend) {
     return ctx->name.c_str();
 }
 
+// TODO: implement proper cleanup
 static void ggml_backend_webgpu_free(ggml_backend_t backend) {
     ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *) backend->context;
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_free(" << ctx->name << ")");
 
-    // TODO: cleanup
+#ifdef GGML_WEBGPU_CPU_PROFILE
+    std::cout << "\n[ggml_webgpu cpu profiling summary]\n";
+    double total_cpu = 0.0;
+    for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) {
+        total_cpu += kv.second;
+    }
+    std::cout << "ggml_webgpu: total cpu time: " << total_cpu << " ms\n";
+    std::cout << "ggml_webgpu: cpu breakdown:\n";
+    for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) {
+        double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0;
+        std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
+    }
+    if (ctx->webgpu_ctx->cpu_detail_ms.size() > 0) {
+        std::cout << "ggml_webgpu: cpu detailed breakdown:\n";
+    }
+    for (const auto & kv : ctx->webgpu_ctx->cpu_detail_ms) {
+        double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0;
+        std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
+    }
+#endif
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    std::cout << "\n[ggml_webgpu gpu profiling summary]\n";
+    double total_gpu = 0.0;
+    for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) {
+        total_gpu += kv.second;
+    }
+    std::cout << "ggml_webgpu: total gpu time (all shaders): " << total_gpu << " ms\n";
+    std::cout << "\nggml_webgpu: gpu breakdown:\n";
+    for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) {
+        double pct = (total_gpu > 0.0) ? (kv.second / total_gpu * 100.0) : 0.0;
+        std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
+    }
+#endif
+
+#if defined(GGML_WEBGPU_CPU_PROFILE) && defined(GGML_WEBGPU_GPU_PROFILE)
+    std::cout << "ggml_webgpu: gpu/cpu ratio: " << (total_cpu > 0.0 ? total_gpu / total_cpu : 0.0) << "\n";
+#endif
+
+#if !defined(GGML_WEBGPU_CPU_PROFILE) && !defined(GGML_WEBGPU_GPU_PROFILE)
     GGML_UNUSED(ctx);
+#endif
 }
 
 static size_t ggml_webgpu_tensor_offset(const ggml_tensor * tensor) {
@@ -457,19 +772,18 @@ static wgpu::Buffer ggml_webgpu_tensor_buf(const ggml_tensor * tensor) {
     return ctx->buffer;
 }
 
-static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, ggml_tensor * t) {
+static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, const ggml_tensor * t) {
     size_t offset = ggml_webgpu_tensor_offset(t);
     return offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
 }
 
-static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, ggml_tensor * t) {
+static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, const ggml_tensor * t) {
     size_t offset = ggml_webgpu_tensor_offset(t);
     return offset & ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
 }
 
 static size_t ggml_webgpu_tensor_binding_size(webgpu_context & ctx, ggml_tensor * t) {
-    return (ggml_nbytes(t) + ggml_webgpu_tensor_misalignment(ctx, t) + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
-           ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1);
+    return ROUNDUP_POW2(ggml_nbytes(t) + ggml_webgpu_tensor_misalignment(ctx, t), WEBGPU_STORAGE_BUF_BINDING_MULT);
 }
 
 // Used to determine if two tensors are the same for in-place operations
@@ -478,7 +792,7 @@ static bool ggml_webgpu_tensor_equal(ggml_tensor * a, ggml_tensor * b) {
            (ggml_webgpu_tensor_offset(a) == ggml_webgpu_tensor_offset(b));
 }
 
-static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
+static webgpu_command ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
     uint32_t ne = (uint32_t) ggml_nelements(dst);
 
     std::vector<uint32_t> params = {
@@ -489,8 +803,9 @@ static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor
         (uint32_t) (src->nb[2] / ggml_type_size(src->type)), (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
         (uint32_t) (dst->nb[0] / ggml_type_size(dst->type)), (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
         (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)), (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
-        // Logical shape — same for both tensors even if permuted
-        (uint32_t) src->ne[0], (uint32_t) src->ne[1], (uint32_t) src->ne[2], (uint32_t) src->ne[3]
+        // Logical shapes
+        (uint32_t) src->ne[0], (uint32_t) src->ne[1], (uint32_t) src->ne[2], (uint32_t) dst->ne[0],
+        (uint32_t) dst->ne[1], (uint32_t) dst->ne[2]
     };
 
     std::vector<wgpu::BindGroupEntry> entries = {
@@ -504,15 +819,17 @@ static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor
          .size    = ggml_webgpu_tensor_binding_size(ctx, dst) }
     };
 
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (ne + max_wg_size - 1) / max_wg_size;
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->cpy_pipeline, params, entries, wg_x, ggml_op_name(dst->op));
+    uint32_t wg_x = CEIL_DIV(ne, WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, ctx->cpy_pipelines[src->type][dst->type], params, entries, wg_x);
 }
 
-static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * idx, ggml_tensor * dst) {
+static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
+                                                          ggml_tensor *    src,
+                                                          ggml_tensor *    idx,
+                                                          ggml_tensor *    dst) {
     // For set rows specifically, we need to check if src and idx are empty tensors.
     if (ggml_is_empty(src) || ggml_is_empty(idx)) {
-        return;
+        return std::nullopt;
     }
 
     webgpu_pool_bufs error_bufs = ctx->set_rows_error_buf_pool.alloc_bufs();
@@ -552,16 +869,24 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
         { .binding = 3, .buffer = error_bufs.dev_buf, .offset = 0, .size = error_bufs.dev_buf.GetSize() }
     };
 
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (src->ne[1] * src->ne[2] * src->ne[3] + max_wg_size - 1) / max_wg_size;
+    int             vectorized = src->ne[0] % 4 == 0;
+    webgpu_pipeline pipeline   = ctx->set_rows_pipelines[0][vectorized];
+    uint32_t        threads;
+    if (vectorized) {
+        threads = (src->ne[1] * src->ne[2] * src->ne[3]) * (src->ne[0] / 4);
+    } else {
+        threads = src->ne[0] * src->ne[1] * src->ne[2] * src->ne[3];
+    }
 
-    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
-    ctx->staged_set_row_error_bufs.push_back(error_bufs);
+    uint32_t wg_x = CEIL_DIV(threads, WEBGPU_MAX_WG_SIZE);
 
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->set_rows_pipeline, params, entries, wg_x, ggml_op_name(dst->op));
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, 1, error_bufs);
 }
 
-static void ggml_webgpu_get_rows(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * idx, ggml_tensor * dst) {
+static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx,
+                                           ggml_tensor *    src,
+                                           ggml_tensor *    idx,
+                                           ggml_tensor *    dst) {
     std::vector<uint32_t> params = {
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, idx) / ggml_type_size(idx->type)),
@@ -593,23 +918,23 @@ static void ggml_webgpu_get_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
          .size    = ggml_webgpu_tensor_binding_size(ctx, dst) }
     };
 
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (dst->ne[1] * dst->ne[2] * dst->ne[3] + max_wg_size - 1) / max_wg_size;
+    uint32_t wg_x = CEIL_DIV(dst->ne[1] * dst->ne[2] * dst->ne[3], WEBGPU_MAX_WG_SIZE);
 
-    wgpu::ComputePipeline pipeline = ctx->get_rows_pipeline[src->type];
-    if (src->type == GGML_TYPE_F32 && dst->ne[0] % 4 != 0) {
-        pipeline = ctx->get_rows_f32_no_vec_pipeline;
-    }
-    ggml_backend_webgpu_build_and_enqueue(ctx, pipeline, params, entries, wg_x, ggml_op_name(dst->op));
+    uint32_t        vectorized = src->type == GGML_TYPE_F32 && dst->ne[0] % 4 == 0;
+    webgpu_pipeline pipeline   = ctx->get_rows_pipelines[src->type][vectorized];
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
 }
 
-static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
+static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
+                                          ggml_tensor *    src0,
+                                          ggml_tensor *    src1,
+                                          ggml_tensor *    dst) {
     std::vector<uint32_t> params = {
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-        (uint32_t) dst->ne[1],                                  // number of rows in result (M)
-        (uint32_t) dst->ne[0],                                  // number of columns in result (N)
+        (uint32_t) dst->ne[0],                                  // number of rows in result (M, transposed)
+        (uint32_t) dst->ne[1],                                  // number of columns in result (N)
         (uint32_t) src0->ne[0],                                 // number of columns in src0/src1 (K)
         (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 1
         (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 1
@@ -638,18 +963,269 @@ static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_t
          .size    = ggml_webgpu_tensor_binding_size(ctx, dst)  },
     };
 
-    uint32_t wg_x =
-        (dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE;
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline[src0->type][src1->type], params, entries, wg_x,
-                                          ggml_op_name(dst->op));
+    webgpu_pipeline pipeline = ctx->mul_mat_pipelines[src0->type][src1->type][0];
+
+    uint32_t wg_x = CEIL_DIV(dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3], WEBGPU_MUL_MAT_WG_SIZE);
+    uint32_t wg_y = 1;
+
+    bool use_fast = false;
+    switch (src1->type) {
+        case GGML_TYPE_F16:
+            use_fast = (src0->type == GGML_TYPE_F16);
+            break;
+        case GGML_TYPE_F32:
+            switch (src0->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                    use_fast = true;
+                    break;
+                default:
+                    break;
+            }
+            break;
+        default:
+            break;
+    }
+
+    if (use_fast) {
+        int vectorized = src0->ne[0] % 4 == 0 && dst->ne[0] % 4 == 0 && dst->ne[1] % 4 == 0;
+        if (dst->ne[1] == 1) {
+            // We don't support vectorized mul_mat_vec for quantized types
+            vectorized             = vectorized && (src0->type < 2);
+            pipeline               = ctx->mul_mat_vec_pipelines[src0->type][src1->type][vectorized];
+            uint32_t batches       = dst->ne[2] * dst->ne[3];
+            uint32_t output_groups = CEIL_DIV(dst->ne[0], WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG);
+            uint32_t total_wg      = output_groups * batches;
+            wg_x                   = total_wg % ctx->limits.maxComputeWorkgroupsPerDimension;
+            wg_y                   = CEIL_DIV(total_wg, ctx->limits.maxComputeWorkgroupsPerDimension);
+        } else {
+            pipeline = ctx->mul_mat_pipelines[src0->type][src1->type][vectorized];
+            uint32_t wg_m;
+            uint32_t wg_n;
+#ifndef __EMSCRIPTEN__
+            if (ctx->supports_subgroup_matrix) {
+                // The total number of subgroups/workgroups needed per matrix.
+                uint32_t wg_m_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M * ctx->sg_mat_m;
+                wg_m                  = CEIL_DIV(dst->ne[0], wg_m_sg_tile);
+                uint32_t wg_n_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N * ctx->sg_mat_n;
+                wg_n                  = CEIL_DIV(dst->ne[1], wg_n_sg_tile);
+            } else {
+#endif
+                uint32_t tile_m_s = WEBGPU_MUL_MAT_TILE_M * WEBGPU_MUL_MAT_WG_SIZE_M;
+                uint32_t tile_n_s = WEBGPU_MUL_MAT_TILE_N * WEBGPU_MUL_MAT_WG_SIZE_N;
+                wg_m              = CEIL_DIV(dst->ne[0], tile_m_s);
+                wg_n              = CEIL_DIV(dst->ne[1], tile_n_s);
+#ifndef __EMSCRIPTEN__
+            }
+#endif
+
+            wg_x = wg_m * wg_n * dst->ne[2] * dst->ne[3];
+        }
+    }
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
+}
+
+static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
+                                             ggml_tensor *    Q,
+                                             ggml_tensor *    K,
+                                             ggml_tensor *    V,
+                                             ggml_tensor *    mask,
+                                             ggml_tensor *    sinks,
+                                             ggml_tensor *    dst) {
+    float scale = *(float *) dst->op_params;
+    float max_bias;
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+    float logit_softcap;
+    memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float));
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
+    float n_head_log2 = float(1u << (uint32_t) floor(log2(Q->ne[2])));
+    float m0          = powf(2.0f, -(max_bias) / n_head_log2);
+    float m1          = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    const int has_mask  = (mask != nullptr);
+    const int has_sinks = (sinks != nullptr);
+
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, Q) / ggml_type_size(Q->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, K) / ggml_type_size(K->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, V) / ggml_type_size(V->type)),
+        has_mask ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, mask) / ggml_type_size(mask->type)) : 0,
+        has_sinks ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, sinks) / ggml_type_size(sinks->type)) : 0,
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) Q->ne[2],                              // number of heads
+        (uint32_t) Q->ne[1],                              // sequence length (Q)
+        (uint32_t) K->ne[1],                              // sequence length (K/V)
+        (uint32_t) (Q->nb[1] / ggml_type_size(Q->type)),  // stride (elements/blocks) of Q in dimension 1
+        (uint32_t) (Q->nb[2] / ggml_type_size(Q->type)),  // stride (elements/blocks) of Q in dimension 2
+        (uint32_t) (Q->nb[3] / ggml_type_size(Q->type)),  // stride (elements/blocks) of Q in dimension 3
+        (uint32_t) (K->nb[1] / ggml_type_size(K->type)),  // stride (elements/blocks) of K in dimension 1
+        (uint32_t) (K->nb[2] / ggml_type_size(K->type)),  // stride (elements/blocks) of K in dimension 2
+        (uint32_t) (K->nb[3] / ggml_type_size(K->type)),  // stride (elements/blocks) of K in dimension 3
+        (uint32_t) (V->nb[1] / ggml_type_size(V->type)),  // stride (elements/blocks) of V in dimension 1
+        (uint32_t) (V->nb[2] / ggml_type_size(V->type)),  // stride (elements/blocks) of V in dimension 2
+        (uint32_t) (V->nb[3] / ggml_type_size(V->type)),  // stride (elements/blocks) of V in dimension 3
+        has_mask ? (uint32_t) (mask->nb[3] / ggml_type_size(mask->type)) : 0,  // stride of mask dim 3
+        (uint32_t) (Q->ne[2] / K->ne[2]),  // repeat factor for K/V in dim 2 (MHA/MQA/GQA)
+        *(uint32_t *) &scale,              // scale (possibly adjusted for logit softcap)
+        *(uint32_t *) &max_bias,
+        *(uint32_t *) &logit_softcap,
+        *(uint32_t *) &n_head_log2,
+        *(uint32_t *) &m0,
+        *(uint32_t *) &m1
+
+    };
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(Q),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, Q),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, Q) },
+        { .binding = 1,
+         .buffer  = ggml_webgpu_tensor_buf(K),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, K),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, K) },
+        { .binding = 2,
+         .buffer  = ggml_webgpu_tensor_buf(V),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, V),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, V) }
+    };
+    uint32_t binding_index = 3;
+    if (has_mask) {
+        entries.push_back({ .binding = binding_index++,
+                            .buffer  = ggml_webgpu_tensor_buf(mask),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, mask),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, mask) });
+    }
+    if (has_sinks) {
+        entries.push_back({ .binding = binding_index++,
+                            .buffer  = ggml_webgpu_tensor_buf(sinks),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, sinks),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, sinks) });
+    }
+    entries.push_back({ .binding = binding_index++,
+                        .buffer  = ggml_webgpu_tensor_buf(dst),
+                        .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                        .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
+
+    bool kv_direct =
+        (K->type == GGML_TYPE_F16) && (Q->ne[0] % ctx->sg_mat_k == 0) && (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0);
+
+    flash_attn_pipeline_key key = {
+        .q_type             = Q->type,
+        .kv_type            = K->type,
+        .dst_type           = dst->type,
+        .head_dim_qk        = (uint32_t) Q->ne[0],
+        .head_dim_v         = (uint32_t) V->ne[0],
+        .kv_direct          = kv_direct,
+        .has_mask           = static_cast<bool>(has_mask),
+        .has_sinks          = static_cast<bool>(has_sinks),
+        .uses_logit_softcap = logit_softcap != 0.0f,
+    };
+
+    webgpu_pipeline                         pipeline;
+    ggml_webgpu_flash_attn_shader_decisions decisions = {};
+
+    auto it = ctx->flash_attn_pipelines.find(key);
+    if (it != ctx->flash_attn_pipelines.end()) {
+        pipeline  = it->second;
+        decisions = *static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context);
+    } else {
+        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        it = ctx->flash_attn_pipelines.find(key);
+        if (it != ctx->flash_attn_pipelines.end()) {
+            pipeline  = it->second;
+            decisions = *static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context);
+        } else {
+            ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = { .kv_type     = K->type,
+                                                                         .head_dim_qk = (uint32_t) Q->ne[0],
+                                                                         .head_dim_v  = (uint32_t) V->ne[0],
+                                                                         .kv_direct   = kv_direct,
+                                                                         .has_mask    = static_cast<bool>(has_mask),
+                                                                         .has_sinks   = static_cast<bool>(has_sinks),
+                                                                         .uses_logit_softcap = logit_softcap != 0.0f,
+                                                                         .sg_mat_m           = ctx->sg_mat_m,
+                                                                         .sg_mat_n           = ctx->sg_mat_n,
+                                                                         .sg_mat_k           = ctx->sg_mat_k,
+                                                                         .wg_mem_limit_bytes =
+                                                                             ctx->limits.maxComputeWorkgroupStorageSize,
+                                                                         .max_subgroup_size = ctx->max_subgroup_size };
+
+            ggml_webgpu_processed_shader processed =
+                ggml_webgpu_preprocess_flash_attn_shader(ctx->p, wgsl_flash_attn, shader_lib_ctx);
+            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline.context = new ggml_webgpu_flash_attn_shader_decisions(processed.decisions);
+            ctx->flash_attn_pipelines.emplace(key, pipeline);
+            decisions = processed.decisions;
+        }
+    }
+
+    uint32_t wg_per_head = CEIL_DIV(Q->ne[1], decisions.q_tile);
+    uint32_t wg_x        = wg_per_head * Q->ne[2] * Q->ne[3];  // wg per head * number of heads * number of batches
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
 }
 
-static void ggml_webgpu_binary_op(webgpu_context &        ctx,
-                                  ggml_tensor *           src0,
-                                  ggml_tensor *           src1,
-                                  ggml_tensor *           dst,
-                                  wgpu::ComputePipeline & pipeline,
-                                  bool                    in_place) {
+static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
+    uint32_t      ne       = (uint32_t) ggml_nelements(dst);
+    ggml_unary_op unary_op = ggml_get_unary_op(dst);
+    uint32_t      inplace  = ggml_webgpu_tensor_equal(src, dst);
+
+    std::vector<uint32_t> params = {
+        ne, (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        // Convert byte-strides to element-strides
+        (uint32_t) (src->nb[0] / ggml_type_size(src->type)), (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
+        (uint32_t) (src->nb[2] / ggml_type_size(src->type)), (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+        (uint32_t) (dst->nb[0] / ggml_type_size(dst->type)), (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)), (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        // Logical shapes
+        (uint32_t) src->ne[0], (uint32_t) src->ne[1], (uint32_t) src->ne[2], (uint32_t) dst->ne[0],
+        (uint32_t) dst->ne[1], (uint32_t) dst->ne[2]
+    };
+
+    switch (unary_op) {
+        case GGML_UNARY_OP_XIELU:
+            {
+                // Get float parameters and reinterpret their bit patterns as uint32_t
+                // for passing through the params buffer
+                float alpha_n = ggml_get_op_params_f32(dst, 1);
+                float alpha_p = ggml_get_op_params_f32(dst, 2);
+                float beta    = ggml_get_op_params_f32(dst, 3);
+                float eps     = ggml_get_op_params_f32(dst, 4);
+                params.push_back(*reinterpret_cast<const uint32_t *>(&alpha_n));
+                params.push_back(*reinterpret_cast<const uint32_t *>(&alpha_p));
+                params.push_back(*reinterpret_cast<const uint32_t *>(&beta));
+                params.push_back(*reinterpret_cast<const uint32_t *>(&eps));
+                break;
+            }
+        default:
+            break;
+    }
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) },
+    };
+    if (!inplace) {
+        entries.push_back({ .binding = 1,
+                            .buffer  = ggml_webgpu_tensor_buf(dst),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
+    }
+
+    uint32_t wg_x = CEIL_DIV(ne, WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, ctx->unary_pipelines[unary_op][dst->type][inplace], params, entries, wg_x);
+}
+
+static webgpu_command ggml_webgpu_binary_op(webgpu_context &  ctx,
+                                            ggml_tensor *     src0,
+                                            ggml_tensor *     src1,
+                                            ggml_tensor *     dst,
+                                            webgpu_pipeline & pipeline,
+                                            bool              inplace) {
     std::vector<uint32_t> params = {
         (uint32_t) ggml_nelements(dst),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
@@ -678,43 +1254,210 @@ static void ggml_webgpu_binary_op(webgpu_context &        ctx,
          .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
          .size    = ggml_webgpu_tensor_binding_size(ctx, src1) }
     };
-    if (!in_place) {
+    if (!inplace) {
         entries.push_back({ .binding = 2,
                             .buffer  = ggml_webgpu_tensor_buf(dst),
                             .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
                             .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
     }
 
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (ggml_nelements(dst) + max_wg_size - 1) / max_wg_size;
-    ggml_backend_webgpu_build_and_enqueue(ctx, pipeline, params, entries, wg_x, ggml_op_name(dst->op));
+    uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
 }
 
-static void ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
-    bool in_place = ggml_webgpu_tensor_equal(src, dst);
+static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
+    int inplace = ggml_webgpu_tensor_equal(src, dst);
 
-    uint32_t eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
+        (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
+        (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        (uint32_t) src->ne[0],
+        (uint32_t) src->ne[1],
+        (uint32_t) src->ne[2],
+        (uint32_t) src->ne[3],
+        *(uint32_t *) dst->op_params  // epsilon, treated as f32 in the shader
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) }
+    };
+    if (!inplace) {
+        entries.push_back({ .binding = 1,
+                            .buffer  = ggml_webgpu_tensor_buf(dst),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
+    }
+
+    return ggml_backend_webgpu_build(ctx, ctx->rms_norm_pipelines[inplace], params, entries, ggml_nrows(src));
+}
+
+static webgpu_command ggml_webgpu_rope(webgpu_context & ctx,
+                                       ggml_tensor *    src0,
+                                       ggml_tensor *    src1,
+                                       ggml_tensor *    src2,
+                                       ggml_tensor *    dst) {
+    const int inplace         = ggml_webgpu_tensor_equal(src0, dst);
+    const int has_freq_factor = (src2 != nullptr);
+
+    const int n_dims     = ((int32_t *) dst->op_params)[1];
+    const int mode       = ((int32_t *) dst->op_params)[2];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+
+    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
+    memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
+    memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
+    memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
+
+    int sections[4];
+    memcpy(sections, (int32_t *) dst->op_params + 11, 4 * sizeof(int));
+
+    float theta_scale = powf(freq_base, -2.0f / n_dims);
+
+    float corr_dims[2];
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
+        src2 != nullptr ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src2) / ggml_type_size(src2->type)) : 0,
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        (uint32_t) ggml_nelements(src0) / 2,
+        (uint32_t) src0->ne[0],
+        (uint32_t) src0->ne[1],
+        (uint32_t) src0->ne[2],
+        (uint32_t) n_dims,
+        (uint32_t) mode,
+        *(uint32_t *) &theta_scale,
+        *(uint32_t *) &attn_factor,
+        *(uint32_t *) &freq_scale,
+        *(uint32_t *) &ext_factor,
+        *(uint32_t *) &corr_dims[0],
+        *(uint32_t *) &corr_dims[1],
+        (uint32_t) sections[0],
+        (uint32_t) sections[1],
+        (uint32_t) sections[2],
+        (uint32_t) sections[3]
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) },
+        { .binding = 1,
+         .buffer  = ggml_webgpu_tensor_buf(src1),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src1) }
+    };
+    uint32_t dst_binding = 2;
+    if (has_freq_factor) {
+        dst_binding = 3;
+        entries.push_back({ .binding = 2,
+                            .buffer  = ggml_webgpu_tensor_buf(src2),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, src2),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, src2) });
+    }
+    if (!inplace) {
+        entries.push_back({ .binding = dst_binding,
+                            .buffer  = ggml_webgpu_tensor_buf(dst),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
+    }
+
+    webgpu_pipeline pipeline = ctx->rope_pipelines[dst->type][has_freq_factor][inplace];
+    uint32_t        wg_x     = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+}
+
+static webgpu_command ggml_webgpu_glu(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
+    const int split = (src1 != nullptr);
+
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        src1 != nullptr ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)) : 0,
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+        src1 != nullptr ? (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)) :
+                          (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+        src1 != nullptr ? (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)) :
+                          (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+        src1 != nullptr ? (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)) :
+                          (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        (uint32_t) ggml_nelements(dst),
+        (uint32_t) dst->ne[0],
+        (uint32_t) dst->ne[1],
+        (uint32_t) dst->ne[2],
+        (uint32_t) ((int32_t *) dst->op_params)[1],  // swapped
+        *(uint32_t *) &dst->op_params[2],            // alpha, for swiglu_oai
+        *(uint32_t *) &dst->op_params[3],            // limit, for swiglu_oai
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) },
+    };
+    uint32_t dst_binding = 1;
+    if (split) {
+        dst_binding = 2;
+        entries.push_back({ .binding = 1,
+                            .buffer  = ggml_webgpu_tensor_buf(src1),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, src1) });
+    }
+    entries.push_back({ .binding = dst_binding,
+                        .buffer  = ggml_webgpu_tensor_buf(dst),
+                        .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                        .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
+
+    webgpu_pipeline pipeline = ctx->glu_pipelines[ggml_get_glu_op(dst)][dst->type][split];
+    uint32_t        wg_x     = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+}
+
+static webgpu_command ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
+    int inplace = ggml_webgpu_tensor_equal(src, dst);
 
     std::vector<uint32_t> params = {
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
+        (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
+        (uint32_t) (src->nb[3] / ggml_type_size(src->type)),
+        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        (uint32_t) ggml_nelements(dst),
+        (uint32_t) src->ne[0],
+        (uint32_t) src->ne[1],
+        (uint32_t) src->ne[2],
+        *(uint32_t *) dst->op_params,     // scale
+        *(uint32_t *) &dst->op_params[1]  // bias
     };
-    if (!in_place) {
-        params.push_back((uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)));
-    }
-    params.push_back((uint32_t) (src->nb[1] / ggml_type_size(src->type)));
-    params.push_back((uint32_t) (src->nb[2] / ggml_type_size(src->type)));
-    params.push_back((uint32_t) (src->nb[3] / ggml_type_size(src->type)));
-    if (!in_place) {
-        params.push_back((uint32_t) (dst->nb[1] / ggml_type_size(dst->type)));
-        params.push_back((uint32_t) (dst->nb[2] / ggml_type_size(dst->type)));
-        params.push_back((uint32_t) (dst->nb[3] / ggml_type_size(dst->type)));
-    }
-    params.push_back((uint32_t) src->ne[0]);
-    params.push_back((uint32_t) src->ne[1]);
-    params.push_back((uint32_t) src->ne[2]);
-    params.push_back((uint32_t) src->ne[3]);
-    params.push_back(eps);  // epsilon, will be bitcast to float in shader
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
@@ -722,33 +1465,100 @@ static void ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_t
          .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
          .size    = ggml_webgpu_tensor_binding_size(ctx, src) }
     };
-    if (!in_place) {
+    if (!inplace) {
         entries.push_back({ .binding = 1,
                             .buffer  = ggml_webgpu_tensor_buf(dst),
                             .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
                             .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
     }
 
-    wgpu::ComputePipeline pipeline;
-    if (in_place) {
-        pipeline = ctx->rms_norm_ip_pipeline;
-    } else {
-        pipeline = ctx->rms_norm_pipeline;
+    uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
+    return ggml_backend_webgpu_build(ctx, ctx->scale_pipelines[inplace], params, entries, wg_x);
+}
+
+static webgpu_command ggml_webgpu_soft_max(webgpu_context & ctx,
+                                           ggml_tensor *    src0,
+                                           ggml_tensor *    src1,
+                                           ggml_tensor *    src2,
+                                           ggml_tensor *    dst) {
+    const int inplace   = ggml_webgpu_tensor_equal(src0, dst);
+    const int mask_type = (src1 != nullptr) ? src1->type : 2;  // use 2 for no mask here
+    const int has_sink  = (src2 != nullptr);
+    float     max_bias;
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+    float n_head_log2 = float(1u << (uint32_t) floor(log2(src0->ne[2])));
+    float m0          = powf(2.0f, -(max_bias) / n_head_log2);
+    float m1          = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        mask_type < 2 ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)) : 0,
+        has_sink ? (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src2) / ggml_type_size(src2->type)) : 0,
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+        mask_type < 2 ? (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)) : 0,
+        mask_type < 2 ? (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)) : 0,
+        mask_type < 2 ? (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)) : 0,
+        (uint32_t) (dst->nb[1] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[2] / ggml_type_size(dst->type)),
+        (uint32_t) (dst->nb[3] / ggml_type_size(dst->type)),
+        (uint32_t) ggml_nelements(dst),
+        (uint32_t) src0->ne[0],
+        (uint32_t) src0->ne[1],
+        (uint32_t) src0->ne[2],
+        mask_type < 2 ? (uint32_t) src1->ne[2] : 0,
+        mask_type < 2 ? (uint32_t) src1->ne[3] : 0,
+        *(uint32_t *) dst->op_params,  // scale
+        *(uint32_t *) &max_bias,
+        *(uint32_t *) &n_head_log2,
+        *(uint32_t *) &m0,
+        *(uint32_t *) &m1
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) }
+    };
+    uint32_t binding_num = 1;
+    if (mask_type < 2) {
+        entries.push_back({ .binding = binding_num,
+                            .buffer  = ggml_webgpu_tensor_buf(src1),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, src1) });
+        binding_num++;
+    }
+    if (has_sink) {
+        entries.push_back({ .binding = binding_num,
+                            .buffer  = ggml_webgpu_tensor_buf(src2),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, src2),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, src2) });
+        binding_num++;
+    }
+    if (!inplace) {
+        entries.push_back({ .binding = binding_num,
+                            .buffer  = ggml_webgpu_tensor_buf(dst),
+                            .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+                            .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
     }
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (src->ne[1] * src->ne[2] * src->ne[3] + max_wg_size - 1) / max_wg_size;
-    ggml_backend_webgpu_build_and_enqueue(ctx, pipeline, params, entries, wg_x, ggml_op_name(dst->op));
+
+    return ggml_backend_webgpu_build(ctx, ctx->soft_max_pipelines[mask_type][has_sink][inplace], params, entries,
+                                     ggml_nrows(dst));
 }
 
-// Returns true if node has enqueued work into the queue, false otherwise
-static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
+// Returns the encoded command, or std::nullopt if the operation is a no-op
+static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
     if (ggml_is_empty(node)) {
-        return false;
+        return std::nullopt;
     }
     WEBGPU_LOG_DEBUG("ggml_webgpu_encode_node(" << node << ", " << ggml_op_name(node->op) << ")");
 
     ggml_tensor * src0 = node->src[0];
     ggml_tensor * src1 = node->src[1];
+    ggml_tensor * src2 = node->src[2];
 
     switch (node->op) {
             // no-ops
@@ -757,40 +1567,53 @@ static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
         case GGML_OP_RESHAPE:
-            return false;
+            return std::nullopt;
         case GGML_OP_CPY:
-            ggml_webgpu_cpy(ctx, src0, node);
-            break;
+        case GGML_OP_CONT:
+            return ggml_webgpu_cpy(ctx, src0, node);
         case GGML_OP_SET_ROWS:
-            ggml_webgpu_set_rows(ctx, src0, src1, node);
-            break;
+            return ggml_webgpu_set_rows(ctx, src0, src1, node);
         case GGML_OP_GET_ROWS:
-            ggml_webgpu_get_rows(ctx, src0, src1, node);
-            break;
+            return ggml_webgpu_get_rows(ctx, src0, src1, node);
         case GGML_OP_MUL_MAT:
-            ggml_webgpu_mul_mat(ctx, src0, src1, node);
-            break;
+            return ggml_webgpu_mul_mat(ctx, src0, src1, node);
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ggml_webgpu_flash_attn(ctx, src0, src1, src2, node->src[3], node->src[4], node);
         case GGML_OP_ADD:
-            if (ggml_webgpu_tensor_equal(src0, node)) {
-                ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->add_ip_pipeline[node->type], true);
-            } else {
-                ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->add_pipeline[node->type], false);
+            {
+                int inplace = ggml_webgpu_tensor_equal(src0, node);
+                return ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->add_pipelines[node->type][inplace], inplace);
+            }
+        case GGML_OP_SUB:
+            {
+                int inplace = ggml_webgpu_tensor_equal(src0, node);
+                return ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->sub_pipelines[node->type][inplace], inplace);
             }
-            break;
         case GGML_OP_MUL:
-            if (ggml_webgpu_tensor_equal(src0, node)) {
-                ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->mul_ip_pipeline[node->type], true);
-            } else {
-                ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->mul_pipeline[node->type], false);
+            {
+                int inplace = ggml_webgpu_tensor_equal(src0, node);
+                return ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->mul_pipelines[node->type][inplace], inplace);
+            }
+        case GGML_OP_DIV:
+            {
+                int inplace = ggml_webgpu_tensor_equal(src0, node);
+                return ggml_webgpu_binary_op(ctx, src0, src1, node, ctx->div_pipelines[node->type][inplace], inplace);
             }
-            break;
         case GGML_OP_RMS_NORM:
-            ggml_webgpu_rms_norm(ctx, src0, node);
-            break;
+            return ggml_webgpu_rms_norm(ctx, src0, node);
+        case GGML_OP_ROPE:
+            return ggml_webgpu_rope(ctx, src0, src1, src2, node);
+        case GGML_OP_GLU:
+            return ggml_webgpu_glu(ctx, src0, src1, node);
+        case GGML_OP_SCALE:
+            return ggml_webgpu_scale(ctx, src0, node);
+        case GGML_OP_SOFT_MAX:
+            return ggml_webgpu_soft_max(ctx, src0, src1, src2, node);
+        case GGML_OP_UNARY:
+            return ggml_webgpu_unary_op(ctx, src0, node);
         default:
-            return false;
+            return std::nullopt;
     }
-    return true;
 }
 
 static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
@@ -799,13 +1622,36 @@ static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, str
     ggml_backend_webgpu_context * backend_ctx = static_cast<ggml_backend_webgpu_context *>(backend->context);
     webgpu_context                ctx         = backend_ctx->webgpu_ctx;
 
+    WEBGPU_CPU_PROFILE_TOTAL_START(graph_compute);
+
+    ctx->inflight_threads++;
+
+    std::vector<webgpu_command>            commands;
+    std::vector<webgpu_submission_futures> futures;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_webgpu_encode_node(ctx, cgraph->nodes[i]);
+        if (auto cmd = ggml_webgpu_encode_node(ctx, cgraph->nodes[i])) {
+            commands.push_back(*cmd);
+        }
+        // compute the batch size based on the number of inflight threads
+        uint32_t inflight_threads = ctx->inflight_threads;
+        uint32_t batch_size       = std::min(std::max(1u, WEBGPU_NUM_PARAM_BUFS / std::max(inflight_threads, 1u)),
+                                             WEBGPU_COMMAND_SUBMIT_BATCH_SIZE);
+        if (commands.size() >= batch_size) {
+            futures.push_back(ggml_backend_webgpu_submit(ctx, commands));
+            // Process events and check for completed submissions
+            ctx->instance.ProcessEvents();
+            ggml_backend_webgpu_wait(ctx, futures, false);
+            commands.clear();
+        }
+    }
+    if (!commands.empty()) {
+        webgpu_submission_futures new_futures = ggml_backend_webgpu_submit(ctx, commands);
+        futures.push_back(new_futures);
     }
 
-    ggml_backend_webgpu_submit_queue(ctx);
-    ggml_backend_webgpu_wait_on_submission(ctx);
-
+    ggml_backend_webgpu_wait(ctx, futures);
+    ctx->inflight_threads--;
+    WEBGPU_CPU_PROFILE_TOTAL_END(graph_compute, ctx);
     return GGML_STATUS_SUCCESS;
 }
 
@@ -831,7 +1677,6 @@ static ggml_backend_i ggml_backend_webgpu_i = {
 /* GGML Backend Buffer Interface */
 
 static void ggml_backend_webgpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_free_buffer()");
     ggml_backend_webgpu_buffer_context * ctx = static_cast<ggml_backend_webgpu_buffer_context *>(buffer->context);
     ctx->buffer.Destroy();
 }
@@ -852,16 +1697,19 @@ static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffe
         return;
     }
 
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", "
-                                                                 << offset << ", " << size << ")");
+    WEBGPU_CPU_PROFILE_TOTAL_START(memset_tensor);
 
     ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
 
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buf_ctx->label << ", " << tensor << ", " << value
+                                                                 << ", " << offset << ", " << size << ")");
+
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
     // This is a trick to set all bytes of a u32 to the same 1 byte value.
     uint32_t val32 = (uint32_t) value * 0x01010101;
     ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, val32, total_offset, size);
+    WEBGPU_CPU_PROFILE_TOTAL_END(memset_tensor, buf_ctx->webgpu_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
@@ -869,11 +1717,13 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                   const void *          data,
                                                   size_t                offset,
                                                   size_t                size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", "
-                                                              << offset << ", " << size << ")");
+    WEBGPU_CPU_PROFILE_TOTAL_START(set_tensor);
     ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
     webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
 
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
+                                                              << ", " << offset << ", " << size << ")");
+
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
     webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
@@ -893,8 +1743,17 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           remaining_size);
     } else {
         // wait for WriteBuffer to complete
-        ggml_backend_webgpu_wait_on_submission(webgpu_ctx);
+        webgpu_ctx->instance.WaitAny(
+            webgpu_ctx->queue.OnSubmittedWorkDone(wgpu::CallbackMode::AllowSpontaneous,
+                                                  [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+                                                      if (status != wgpu::QueueWorkDoneStatus::Success) {
+                                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n",
+                                                                         std::string(message).c_str());
+                                                      }
+                                                  }),
+            UINT64_MAX);
     }
+    WEBGPU_CPU_PROFILE_TOTAL_END(set_tensor, webgpu_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
@@ -902,12 +1761,12 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                   void *                data,
                                                   size_t                offset,
                                                   size_t                size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", "
-                                                              << offset << ", " << size << ")");
-
-    ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
-    webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
-    wgpu::Device                         device     = webgpu_ctx->device;
+    WEBGPU_CPU_PROFILE_TOTAL_START(get_tensor);
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
+                                                              << ", " << offset << ", " << size << ")");
+    webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
+    wgpu::Device   device     = webgpu_ctx->device;
 
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
@@ -944,12 +1803,15 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
     // Copy the data from the mapped range to the output buffer
     std::memcpy(data, mapped_range, size);
     webgpu_ctx->get_tensor_staging_buf.Unmap();
+    WEBGPU_CPU_PROFILE_TOTAL_END(get_tensor, webgpu_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_clear(" << buffer << ", " << (uint32_t) value << ")");
+    WEBGPU_CPU_PROFILE_TOTAL_START(clear);
     ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
     ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, value, 0, buffer->size);
+    WEBGPU_CPU_PROFILE_TOTAL_END(clear, buf_ctx->webgpu_ctx);
 }
 
 static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = {
@@ -975,16 +1837,19 @@ static const char * ggml_backend_webgpu_buffer_type_get_name(ggml_backend_buffer
 
 static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                                                           size_t                     size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer(" << size << ")");
+    static std::atomic<int> buffer_count;
+    int                     buffer_id = buffer_count++;
+    std::string             buf_name  = "tensor_buf" + std::to_string(buffer_id);
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer_" << buffer_id << ": " << size << " bytes");
     ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
 
     wgpu::Buffer buf;
-    ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf,
-                              (size + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1),
+    ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf, ROUNDUP_POW2(size, WEBGPU_STORAGE_BUF_BINDING_MULT),
                               wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
-                              "allocated_buffer");
+                              buf_name.c_str());
 
-    ggml_backend_webgpu_buffer_context * buf_ctx = new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf);
+    ggml_backend_webgpu_buffer_context * buf_ctx =
+        new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf, buf_name);
 
     return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
 }
@@ -1016,9 +1881,18 @@ static const char * ggml_backend_webgpu_device_get_description(ggml_backend_dev_
 
 static void ggml_backend_webgpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
-    // TODO: what do we actually want to return here? maxBufferSize might not be the full available memory.
-    *free                                    = ctx->webgpu_ctx->limits.maxBufferSize;
-    *total                                   = ctx->webgpu_ctx->limits.maxBufferSize;
+    // TODO: for now, return maxBufferSize as both free and total memory
+    // Track https://github.com/gpuweb/gpuweb/issues/5505 for updates.
+    uint64_t max_buffer_size = ctx->webgpu_ctx->limits.maxBufferSize;
+    // If we're on a 32-bit system, clamp to UINTPTR_MAX
+#if UINTPTR_MAX < UINT64_MAX
+    uint64_t max_ptr_size = static_cast<uint64_t>(UINTPTR_MAX);
+    if (max_buffer_size > max_ptr_size) {
+        max_buffer_size = max_ptr_size;
+    }
+#endif
+    *free  = static_cast<size_t>(max_buffer_size);
+    *total = static_cast<size_t>(max_buffer_size);
 }
 
 static enum ggml_backend_dev_type ggml_backend_webgpu_device_get_type(ggml_backend_dev_t dev) {
@@ -1044,168 +1918,609 @@ static ggml_guid_t ggml_backend_webgpu_guid(void) {
     return reinterpret_cast<ggml_guid_t>((void *) guid_str);
 }
 
-// The max workgroup size is a common constant
-static std::vector<wgpu::ConstantEntry> ggml_webgpu_max_wg_size_entry(webgpu_context & webgpu_ctx) {
+// Workgroup size is a common constant
+static std::vector<wgpu::ConstantEntry> ggml_webgpu_wg_size_entry(uint32_t wg_size) {
     std::vector<wgpu::ConstantEntry> constants(1);
     constants[0].key   = "wg_size";
-    constants[0].value = webgpu_ctx->max_wg_size_x;
+    constants[0].value = wg_size;
     return constants;
 }
 
 static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
     // we use the maximum workgroup size for the memset pipeline
-    size_t max_wg_size = webgpu_ctx->max_wg_size_x;
-    size_t max_threads = max_wg_size * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension;
+    size_t max_threads                  = WEBGPU_MAX_WG_SIZE * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension;
     // Size the bytes_per_thread so that the largest buffer size can be handled
-    webgpu_ctx->memset_bytes_per_thread =
-        (webgpu_ctx->limits.maxStorageBufferBindingSize + max_threads - 1) / max_threads;
+    webgpu_ctx->memset_bytes_per_thread = CEIL_DIV(webgpu_ctx->limits.maxStorageBufferBindingSize, max_threads);
     std::vector<wgpu::ConstantEntry> constants(2);
-    constants[0].key   = "wg_size";
-    constants[0].value = max_wg_size;
-    constants[1].key   = "bytes_per_thread";
-    constants[1].value = webgpu_ctx->memset_bytes_per_thread;
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->memset_pipeline, wgsl_memset, "memset", constants);
+    constants[0].key                = "wg_size";
+    constants[0].value              = WEBGPU_MAX_WG_SIZE;
+    constants[1].key                = "bytes_per_thread";
+    constants[1].value              = webgpu_ctx->memset_bytes_per_thread;
+    webgpu_ctx->memset_pipelines[0] = ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_memset, "memset", constants);
 }
 
 static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F32][GGML_TYPE_F32],
-                                wgsl_mul_mat_f32_f32, "mul_mat_f32_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F16],
-                                wgsl_mul_mat_f16_f16, "mul_mat_f16_f16");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F32],
-                                wgsl_mul_mat_f16_f32, "mul_mat_f16_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_0][GGML_TYPE_F32],
-                                wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_1][GGML_TYPE_F32],
-                                wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_0][GGML_TYPE_F32],
-                                wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_1][GGML_TYPE_F32],
-                                wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q8_0][GGML_TYPE_F32],
-                                wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q2_K][GGML_TYPE_F32],
-                                wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q3_K][GGML_TYPE_F32],
-                                wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_K][GGML_TYPE_F32],
-                                wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_K][GGML_TYPE_F32],
-                                wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q6_K][GGML_TYPE_F32],
-                                wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XS][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_S][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_S][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_S][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_M][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_NL][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_XS][GGML_TYPE_F32],
-                                wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32");
+    // Q4/Q5/Q8 classic quantizations
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_1][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_0][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_1][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q8_0][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32");
+
+    // K-quantizations
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q2_K][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q3_K][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_K][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_K][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q6_K][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32");
+
+    // IQ quantizations (2-, 3-, 4-bit variants)
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XS][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_S][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32");
+
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_S][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32");
+
+    // 1-bit and 4-bit IQ variants
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_S][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_M][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_NL][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32");
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_XS][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32");
+
+    std::string proc_mul_mat_f32_f32;
+    std::string proc_mul_mat_f32_f32_vec;
+    std::string proc_mul_mat_f16_f32;
+    std::string proc_mul_mat_f16_f32_vec;
+    std::string proc_mul_mat_f16_f16;
+    std::string proc_mul_mat_f16_f16_vec;
+    std::string proc_mul_mat_q4_0_f32;
+    std::string proc_mul_mat_q4_0_f32_vec;
+
+    std::vector<wgpu::ConstantEntry> mul_mat_constants;
+#ifndef __EMSCRIPTEN__
+    if (webgpu_ctx->supports_subgroup_matrix) {
+        std::map<std::string, std::string> sg_matrix_repls;
+        sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] = std::to_string(webgpu_ctx->max_subgroup_size);
+        sg_matrix_repls["WEBGPU_TILE_K"]            = std::to_string(WEBGPU_MUL_MAT_TILE_K);
+        sg_matrix_repls["WEBGPU_SUBGROUP_M"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_M);
+        sg_matrix_repls["WEBGPU_SUBGROUP_N"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_N);
+        sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_M"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M);
+        sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_N"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N);
+        sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_m);
+        sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_n);
+        sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_k);
+
+        proc_mul_mat_f32_f32 = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32, sg_matrix_repls);
+        proc_mul_mat_f32_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32_vec, sg_matrix_repls);
+        proc_mul_mat_f16_f32 = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f32, sg_matrix_repls);
+        proc_mul_mat_f16_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f32_vec, sg_matrix_repls);
+        proc_mul_mat_f16_f16 = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f16, sg_matrix_repls);
+        proc_mul_mat_f16_f16_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f16_vec, sg_matrix_repls);
+        proc_mul_mat_q4_0_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_q4_0_f32, sg_matrix_repls);
+        proc_mul_mat_q4_0_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_q4_0_f32_vec, sg_matrix_repls);
+    } else {
+#endif
+        mul_mat_constants.push_back({ .key = "TILE_K", .value = WEBGPU_MUL_MAT_TILE_K });
+        mul_mat_constants.push_back({ .key = "WORKGROUP_SIZE_M", .value = WEBGPU_MUL_MAT_WG_SIZE_M });
+        mul_mat_constants.push_back({ .key = "WORKGROUP_SIZE_N", .value = WEBGPU_MUL_MAT_WG_SIZE_N });
+
+        std::map<std::string, std::string> reg_repls;
+        reg_repls["WEBGPU_TILE_M"] = std::to_string(WEBGPU_MUL_MAT_TILE_M);
+        reg_repls["WEBGPU_TILE_N"] = std::to_string(WEBGPU_MUL_MAT_TILE_N);
+
+        proc_mul_mat_f32_f32      = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f32_f32, reg_repls);
+        proc_mul_mat_f32_f32_vec  = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f32_f32_vec, reg_repls);
+        proc_mul_mat_f16_f32      = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f32, reg_repls);
+        proc_mul_mat_f16_f32_vec  = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f32_vec, reg_repls);
+        proc_mul_mat_f16_f16      = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f16, reg_repls);
+        proc_mul_mat_f16_f16_vec  = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f16_vec, reg_repls);
+        proc_mul_mat_q4_0_f32     = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_q4_0_f32, reg_repls);
+        proc_mul_mat_q4_0_f32_vec = ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_q4_0_f32_vec, reg_repls);
+#ifndef __EMSCRIPTEN__
+    }
+#endif
+
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f32_f32.c_str(), "mul_mat_f32_f32", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f32_f32_vec.c_str(), "mul_mat_f32_f32_vec", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f16_f32.c_str(), "mul_mat_f16_f32", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f16_f32_vec.c_str(), "mul_mat_f16_f32_vec", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f16_f16.c_str(), "mul_mat_f16_f16", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_f16_f16_vec.c_str(), "mul_mat_f16_f16_vec", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_q4_0_f32.c_str(), "mul_mat_q4_0_f32", mul_mat_constants);
+    webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, proc_mul_mat_q4_0_f32_vec.c_str(), "mul_mat_q4_0_f32_vec", mul_mat_constants);
+
+    std::vector<wgpu::ConstantEntry> mul_mat_vec_constants(3);
+    mul_mat_vec_constants[0].key   = "WORKGROUP_SIZE";
+    mul_mat_vec_constants[0].value = WEBGPU_MUL_MAT_VEC_WG_SIZE;
+    mul_mat_vec_constants[1].key   = "TILE_K";
+    mul_mat_vec_constants[1].value = WEBGPU_MUL_MAT_VEC_TILE_K;
+    mul_mat_vec_constants[2].key   = "OUTPUTS_PER_WG";
+    mul_mat_vec_constants[2].value = WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG;
+
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants);
 }
 
 static void ggml_webgpu_init_set_rows_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->set_rows_pipeline, wgsl_set_rows, "set_rows",
-                                ggml_webgpu_max_wg_size_entry(webgpu_ctx));
+    webgpu_ctx->set_rows_pipelines[0][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_set_rows_f16, "set_rows_f16", ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE));
+    webgpu_ctx->set_rows_pipelines[0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_set_rows_f16_vec, "set_rows_f16_vec", ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE));
 }
 
 static void ggml_webgpu_init_get_rows_pipeline(webgpu_context & webgpu_ctx) {
-    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_max_wg_size_entry(webgpu_ctx);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_F32], wgsl_get_rows_f32_vec,
-                                "get_rows_f32_vec", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_f32_no_vec_pipeline, wgsl_get_rows_f32,
-                                "get_rows_f32", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_F16], wgsl_get_rows_f16,
-                                "get_rows_f16", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_I32], wgsl_get_rows_i32,
-                                "get_rows_i32", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q4_0], wgsl_get_rows_q4_0,
-                                "get_rows_q4_0", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q4_1], wgsl_get_rows_q4_1,
-                                "get_rows_q4_1", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q5_0], wgsl_get_rows_q5_0,
-                                "get_rows_q5_0", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q5_1], wgsl_get_rows_q5_1,
-                                "get_rows_q5_1", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q8_0], wgsl_get_rows_q8_0,
-                                "get_rows_q8_0", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q2_K], wgsl_get_rows_q2_k,
-                                "get_rows_q2_k", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q3_K], wgsl_get_rows_q3_k,
-                                "get_rows_q3_k", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q4_K], wgsl_get_rows_q4_k,
-                                "get_rows_q4_k", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q5_K], wgsl_get_rows_q5_k,
-                                "get_rows_q5_k", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_Q6_K], wgsl_get_rows_q6_k,
-                                "get_rows_q6_k", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ2_XXS],
-                                wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ2_XS],
-                                wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ2_S], wgsl_get_rows_iq2_s,
-                                "get_rows_iq2_s", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ3_XXS],
-                                wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ3_S], wgsl_get_rows_iq3_s,
-                                "get_rows_iq3_s", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ1_S], wgsl_get_rows_iq1_s,
-                                "get_rows_iq1_s", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ1_M], wgsl_get_rows_iq1_m,
-                                "get_rows_iq1_m", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ4_NL],
-                                wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->get_rows_pipeline[GGML_TYPE_IQ4_XS],
-                                wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants);
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32, "get_rows_f32", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32_vec, "get_rows_f32_vec", constants);
+
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f16, "get_rows_f16", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_I32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_i32, "get_rows_i32", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_0, "get_rows_q4_0", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_1][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_1, "get_rows_q4_1", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_0, "get_rows_q5_0", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_1][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_1, "get_rows_q5_1", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q8_0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q8_0, "get_rows_q8_0", constants);
+
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q2_K][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q2_k, "get_rows_q2_k", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q3_K][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q3_k, "get_rows_q3_k", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_K][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_k, "get_rows_q4_k", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_K][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_k, "get_rows_q5_k", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q6_K][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q6_k, "get_rows_q6_k", constants);
+
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XXS][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XS][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_S][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_s, "get_rows_iq2_s", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_XXS][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_S][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_s, "get_rows_iq3_s", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_S][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_s, "get_rows_iq1_s", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_M][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_m, "get_rows_iq1_m", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_NL][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_XS][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants);
 }
 
 static void ggml_webgpu_init_cpy_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->cpy_pipeline, wgsl_cpy, "cpy",
-                                ggml_webgpu_max_wg_size_entry(webgpu_ctx));
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F32] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f32, "cpy_f32_f32", constants);
+    webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F16] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f16, "cpy_f32_f16", constants);
+    webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F32] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f32, "cpy_f16_f32", constants);
+    webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F16] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f16, "cpy_f16_f16", constants);
 }
 
 static void ggml_webgpu_init_add_pipeline(webgpu_context & webgpu_ctx) {
-    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_max_wg_size_entry(webgpu_ctx);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->add_pipeline[GGML_TYPE_F32], wgsl_add_f32, "add_f32",
-                                constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->add_pipeline[GGML_TYPE_F16], wgsl_add_f16, "add_f16",
-                                constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->add_ip_pipeline[GGML_TYPE_F32], wgsl_add_in_place_f32,
-                                "add_in_place_f32", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->add_ip_pipeline[GGML_TYPE_F16], wgsl_add_in_place_f16,
-                                "add_in_place_f16", constants);
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->add_pipelines[GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32, "add_f32", constants);
+    webgpu_ctx->add_pipelines[GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16, "add_f16", constants);
+    webgpu_ctx->add_pipelines[GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32_inplace, "add_f32_inplace", constants);
+    webgpu_ctx->add_pipelines[GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16_inplace, "add_f16_inplace", constants);
+}
+
+static void ggml_webgpu_init_sub_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->sub_pipelines[GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32, "sub_f32", constants);
+    webgpu_ctx->sub_pipelines[GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16, "sub_f16", constants);
+    webgpu_ctx->sub_pipelines[GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32_inplace, "sub_f32_inplace", constants);
+    webgpu_ctx->sub_pipelines[GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16_inplace, "sub_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_mul_pipeline(webgpu_context & webgpu_ctx) {
-    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_max_wg_size_entry(webgpu_ctx);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_pipeline[GGML_TYPE_F32], wgsl_mul_f32, "mul_f32",
-                                constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_pipeline[GGML_TYPE_F16], wgsl_mul_f16, "mul_f16",
-                                constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_ip_pipeline[GGML_TYPE_F32], wgsl_mul_in_place_f32,
-                                "mul_in_place_f32", constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_ip_pipeline[GGML_TYPE_F16], wgsl_mul_in_place_f16,
-                                "mul_in_place_f16", constants);
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->mul_pipelines[GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32, "mul_f32", constants);
+    webgpu_ctx->mul_pipelines[GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16, "mul_f16", constants);
+    webgpu_ctx->mul_pipelines[GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32_inplace, "mul_f32_inplace", constants);
+    webgpu_ctx->mul_pipelines[GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16_inplace, "mul_f16_inplace", constants);
+}
+
+static void ggml_webgpu_init_div_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->div_pipelines[GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32, "div_f32", constants);
+    webgpu_ctx->div_pipelines[GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16, "div_f16", constants);
+    webgpu_ctx->div_pipelines[GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32_inplace, "div_f32_inplace", constants);
+    webgpu_ctx->div_pipelines[GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16_inplace, "div_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_rms_norm_pipeline(webgpu_context & webgpu_ctx) {
-    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_max_wg_size_entry(webgpu_ctx);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->rms_norm_pipeline, wgsl_rms_norm, "rms_norm",
-                                constants);
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->rms_norm_ip_pipeline, wgsl_rms_norm_in_place,
-                                "rms_norm_in_place", constants);
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_ROW_SPLIT_WG_SIZE);
+
+    webgpu_ctx->rms_norm_pipelines[0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm, "rms_norm", constants);
+    webgpu_ctx->rms_norm_pipelines[1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm_inplace, "rms_norm_inplace", constants);
+}
+
+static void ggml_webgpu_init_rope_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32, "rope_f32", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_inplace, "rope_f32_inplace", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff, "rope_f32_ff", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff_inplace, "rope_f32_ff_inplace", constants);
+
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16, "rope_f16", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_inplace, "rope_f16_inplace", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff, "rope_f16_ff", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff_inplace, "rope_f16_ff_inplace", constants);
+}
+
+static void ggml_webgpu_init_glu_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    // REGLU
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32, "reglu_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16, "reglu_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32_split, "reglu_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16_split, "reglu_f16_split", constants);
+
+    // GEGLU
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32, "geglu_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16, "geglu_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32_split, "geglu_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16_split, "geglu_f16_split", constants);
+
+    // SWIGLU
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32, "swiglu_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16, "swiglu_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32_split, "swiglu_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16_split, "swiglu_f16_split", constants);
+
+    // SWIGLU_OAI
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32, "swiglu_oai_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32_split, "swiglu_oai_f32_split", constants);
+
+    // GEGLU_ERF
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32, "geglu_erf_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16, "geglu_erf_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32_split, "geglu_erf_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16_split, "geglu_erf_f16_split", constants);
+
+    // GEGLU_QUICK
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32, "geglu_quick_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16, "geglu_quick_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32_split, "geglu_quick_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16_split, "geglu_quick_f16_split", constants);
+}
+
+static void ggml_webgpu_init_unary_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    // ABS
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ABS][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_abs_f32, "abs_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ABS][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_abs_f16, "abs_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ABS][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_abs_inplace_f32, "abs_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ABS][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_abs_inplace_f16, "abs_inplace_f16", constants);
+
+    // SGN
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SGN][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sgn_f32, "sgn_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SGN][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sgn_f16, "sgn_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SGN][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sgn_inplace_f32, "sgn_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SGN][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sgn_inplace_f16, "sgn_inplace_f16", constants);
+
+    // NEG
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_NEG][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_neg_f32, "neg_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_NEG][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_neg_f16, "neg_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_NEG][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_neg_inplace_f32, "neg_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_NEG][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_neg_inplace_f16, "neg_inplace_f16", constants);
+
+    // STEP
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_STEP][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_step_f32, "step_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_STEP][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_step_f16, "step_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_STEP][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_step_inplace_f32, "step_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_STEP][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_step_inplace_f16, "step_inplace_f16", constants);
+
+    // TANH
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_TANH][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_tanh_f32, "tanh_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_TANH][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_tanh_f16, "tanh_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_TANH][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_tanh_inplace_f32, "tanh_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_TANH][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_tanh_inplace_f16, "tanh_inplace_f16", constants);
+
+    // ELU
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ELU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_elu_f32, "elu_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ELU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_elu_f16, "elu_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ELU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_elu_inplace_f32, "elu_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_ELU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_elu_inplace_f16, "elu_inplace_f16", constants);
+
+    // RELU
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_RELU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_relu_f32, "relu_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_RELU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_relu_f16, "relu_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_RELU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_relu_inplace_f32, "relu_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_RELU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_relu_inplace_f16, "relu_inplace_f16", constants);
+
+    // SIGMOID
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SIGMOID][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sigmoid_f32, "sigmoid_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SIGMOID][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sigmoid_f16, "sigmoid_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SIGMOID][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sigmoid_inplace_f32, "sigmoid_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SIGMOID][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sigmoid_inplace_f16, "sigmoid_inplace_f16", constants);
+
+    // GELU
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_f32, "gelu_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_f16, "gelu_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_inplace_f32, "gelu_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_inplace_f16, "gelu_inplace_f16", constants);
+
+    // GELU_QUICK
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_QUICK][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_quick_f32, "gelu_quick_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_QUICK][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_quick_f16, "gelu_quick_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_QUICK][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_gelu_quick_inplace_f32, "gelu_quick_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_QUICK][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_gelu_quick_inplace_f16, "gelu_quick_inplace_f16", constants);
+
+    // SILU
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SILU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_silu_f32, "silu_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SILU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_silu_f16, "silu_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SILU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_silu_inplace_f32, "silu_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_SILU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_silu_inplace_f16, "silu_inplace_f16", constants);
+
+    // HARDSWISH
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSWISH][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardswish_f32, "hardswish_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSWISH][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardswish_f16, "hardswish_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSWISH][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardswish_inplace_f32, "hardswish_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSWISH][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardswish_inplace_f16, "hardswish_inplace_f16", constants);
+
+    // HARDSIGMOID
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSIGMOID][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardsigmoid_f32, "hardsigmoid_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSIGMOID][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_hardsigmoid_f16, "hardsigmoid_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSIGMOID][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_hardsigmoid_inplace_f32, "hardsigmoid_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_HARDSIGMOID][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_hardsigmoid_inplace_f16, "hardsigmoid_inplace_f16", constants);
+
+    // EXP
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_EXP][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_exp_f32, "exp_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_EXP][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_exp_f16, "exp_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_EXP][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_exp_inplace_f32, "exp_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_EXP][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_exp_inplace_f16, "exp_inplace_f16", constants);
+
+    // GELU_ERF
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_ERF][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_erf_f32, "gelu_erf_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_ERF][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_erf_f16, "gelu_erf_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_ERF][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_erf_inplace_f32, "gelu_erf_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_GELU_ERF][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_gelu_erf_inplace_f16, "gelu_erf_inplace_f16", constants);
+
+    // XIELU
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_XIELU][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_xielu_f32, "xielu_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_XIELU][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_xielu_f16, "xielu_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_XIELU][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_xielu_inplace_f32, "xielu_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_XIELU][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_xielu_inplace_f16, "xielu_inplace_f16", constants);
+
+    // CEIL
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_CEIL][GGML_TYPE_F32][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_ceil_f32, "ceil_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_CEIL][GGML_TYPE_F16][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_ceil_f16, "ceil_f16", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_CEIL][GGML_TYPE_F32][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_ceil_inplace_f32, "ceil_inplace_f32", constants);
+    webgpu_ctx->unary_pipelines[GGML_UNARY_OP_CEIL][GGML_TYPE_F16][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_ceil_inplace_f16, "ceil_inplace_f16", constants);
 }
 
+static void ggml_webgpu_init_scale_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
+
+    webgpu_ctx->scale_pipelines[0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32, "scale_f32", constants);
+    webgpu_ctx->scale_pipelines[1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32_inplace, "scale_f32_inplace", constants);
+}
+
+static void ggml_webgpu_init_soft_max_pipeline(webgpu_context & webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_ROW_SPLIT_WG_SIZE);
+
+    // f32 (no mask)
+    webgpu_ctx->soft_max_pipelines[2][0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32, "soft_max_f32", constants);
+    webgpu_ctx->soft_max_pipelines[2][0][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_inplace, "soft_max_f32_inplace", constants);
+    webgpu_ctx->soft_max_pipelines[2][1][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_sink, "soft_max_f32_sink", constants);
+    webgpu_ctx->soft_max_pipelines[2][1][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_sink_inplace, "soft_max_f32_sink_inplace", constants);
+
+    // f32 mask (mask_type = 0)
+    webgpu_ctx->soft_max_pipelines[0][0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f32, "soft_max_f32_mask_f32", constants);
+    webgpu_ctx->soft_max_pipelines[0][0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_inplace, "soft_max_f32_mask_f32_inplace", constants);
+    webgpu_ctx->soft_max_pipelines[0][1][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink, "soft_max_f32_mask_f32_sink", constants);
+    webgpu_ctx->soft_max_pipelines[0][1][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink_inplace, "soft_max_f32_mask_f32_sink_inplace", constants);
+
+    // f16 mask (mask_type = 1)
+    webgpu_ctx->soft_max_pipelines[1][0][0] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f16, "soft_max_f32_mask_f16", constants);
+    webgpu_ctx->soft_max_pipelines[1][0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_inplace, "soft_max_f32_mask_f16_inplace", constants);
+    webgpu_ctx->soft_max_pipelines[1][1][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink, "soft_max_f32_mask_f16_sink", constants);
+    webgpu_ctx->soft_max_pipelines[1][1][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink_inplace, "soft_max_f32_mask_f16_sink_inplace", constants);
+}
+
+// TODO: move most initialization logic here
 static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, const char * params) {
     GGML_UNUSED(params);
 
@@ -1225,12 +2540,12 @@ static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, co
         /* .device    = */ dev,
         /* .context   = */ &backend_ctx,
     };
-
     return &backend;
 }
 
 static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggml_backend_dev_t dev) {
     // See GGML Backend Buffer Type Interface section
+
     static struct ggml_backend_buffer_type ggml_backend_webgpu_buffer_type = {
         /* .iface = */ {
                         /* .get_name         = */ ggml_backend_webgpu_buffer_type_get_name,
@@ -1287,6 +2602,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
 
     ggml_tensor * src0 = op->src[0];
     ggml_tensor * src1 = op->src[1];
+    ggml_tensor * src2 = op->src[2];
+
     // on smaller devices (or CI), tensors may be larger than the max storage buffer size
     if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize ||
         (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
@@ -1304,28 +2621,36 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
             supports_op = true;
             break;
         case GGML_OP_ADD:
+        case GGML_OP_SUB:
         case GGML_OP_MUL:
-            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (op->src[0]->type == op->type) &&
-                          (op->src[1]->type == op->type);
+        case GGML_OP_DIV:
+            // TODO: support non-contiguous tensors, e.g. for MOE_EXPERT_REDUCE
+            // see https://github.com/ggml-org/llama.cpp/pull/16857
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type) &&
+                          (src1->type == op->type) && ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
             break;
         case GGML_OP_CPY:
+        case GGML_OP_CONT:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
+                          (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+            break;
         case GGML_OP_SET_ROWS:
-            supports_op = (op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_I64);
+            supports_op = (op->type == GGML_TYPE_F16 && src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I64);
             break;
         case GGML_OP_GET_ROWS:
-            if (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 ||
-                op->src[0]->type == GGML_TYPE_I32 || ggml_webgpu_supported_qtype(op->src[0]->type)) {
+            if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_I32 ||
+                ggml_webgpu_supported_qtype(src0->type)) {
                 supports_op = (op->type == GGML_TYPE_F32);
             }
             break;
         case GGML_OP_MUL_MAT:
             {
-                switch (op->src[1]->type) {
+                switch (src1->type) {
                     case GGML_TYPE_F16:
-                        supports_op = (op->src[0]->type == GGML_TYPE_F16);
+                        supports_op |= (src0->type == GGML_TYPE_F16);
                         break;
                     case GGML_TYPE_F32:
-                        switch (op->src[0]->type) {
+                        switch (src0->type) {
                             case GGML_TYPE_F32:
                             case GGML_TYPE_F16:
                             case GGML_TYPE_Q4_0:
@@ -1357,19 +2682,110 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                 }
                 break;
             }
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                if (!webgpu_ctx->supports_subgroup_matrix) {
+                    break;
+                }
+                // Head dimensions must fit in workgroup memory with minimum tile sizes
+                size_t     limit_bytes = webgpu_ctx->limits.maxComputeWorkgroupStorageSize;
+                const bool has_mask    = op->src[3] != nullptr;
+                const bool kv_direct   = src1->type == GGML_TYPE_F16 && (src0->ne[0] % webgpu_ctx->sg_mat_k) == 0 &&
+                                       (src1->ne[1] % GGML_WEBGPU_KV_SEQ_PAD) == 0;
+                const size_t min_bytes = ggml_webgpu_flash_attn_wg_mem_bytes(
+                    webgpu_ctx->sg_mat_m, webgpu_ctx->sg_mat_n, (uint32_t) src0->ne[0], (uint32_t) src2->ne[0],
+                    has_mask, kv_direct);
+                if (min_bytes > limit_bytes) {
+                    break;
+                }
+
+                supports_op = src0->type == GGML_TYPE_F32 &&
+                              (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 ||
+                               src1->type == GGML_TYPE_Q4_0 || src1->type == GGML_TYPE_Q8_0) &&
+                              src2->type == src1->type && op->type == GGML_TYPE_F32;
+                break;
+            }
         case GGML_OP_RMS_NORM:
-            supports_op = op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
+            supports_op = op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32;
+            break;
+        case GGML_OP_ROPE:
+            supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
+            break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
+                    break;
+                case GGML_GLU_OP_SWIGLU_OAI:
+                    supports_op = op->type == GGML_TYPE_F32;
+                    break;
+                default:
+                    break;
+            }
+            break;
+        case GGML_OP_SCALE:
+            supports_op = op->type == GGML_TYPE_F32;
+            break;
+        case GGML_OP_SOFT_MAX:
+            supports_op = op->type == GGML_TYPE_F32;
+            break;
+        case GGML_OP_UNARY:
+            {
+                const ggml_unary_op UNARY_OP = ggml_get_unary_op(op);
+
+                switch (UNARY_OP) {
+                    case GGML_UNARY_OP_ABS:
+                    case GGML_UNARY_OP_SGN:
+                    case GGML_UNARY_OP_NEG:
+                    case GGML_UNARY_OP_STEP:
+                    case GGML_UNARY_OP_TANH:
+                    case GGML_UNARY_OP_ELU:
+                    case GGML_UNARY_OP_RELU:
+                    case GGML_UNARY_OP_SIGMOID:
+                    case GGML_UNARY_OP_GELU:
+                    case GGML_UNARY_OP_GELU_QUICK:
+                    case GGML_UNARY_OP_SILU:
+                    case GGML_UNARY_OP_HARDSWISH:
+                    case GGML_UNARY_OP_HARDSIGMOID:
+                    case GGML_UNARY_OP_EXP:
+                    case GGML_UNARY_OP_GELU_ERF:
+                    case GGML_UNARY_OP_XIELU:
+                    case GGML_UNARY_OP_CEIL:
+                        supports_op = supports_op =
+                            (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
+                        break;
+                    default:
+                        break;
+                }
+            }
             break;
+
         default:
             break;
     }
-#ifdef GGML_WEBGPU_DEBUG
+    if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize ||
+        (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
+        (src1 != nullptr && ggml_nbytes(src1) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
+        (src2 != nullptr && ggml_nbytes(src2) > webgpu_ctx->limits.maxStorageBufferBindingSize)) {
+        supports_op = false;
+        WEBGPU_LOG_DEBUG("ggml_webgpu op not supported due to size: ");
+    }
+
     if (!supports_op) {
-        WEBGPU_LOG_DEBUG("not supported: " << ggml_op_name(op->op) << " with types dst: " << ggml_type_name(op->type)
-                                           << ", src0: " << (op->src[0] ? ggml_type_name(op->src[0]->type) : "null")
-                                           << ", src1: " << (op->src[1] ? ggml_type_name(op->src[1]->type) : "null"));
+        WEBGPU_LOG_DEBUG("ggml_webgpu op not supported: "
+                         << ggml_op_name(op->op) << " with types dst: " << ggml_type_name(op->type)
+                         << ", src0: " << (op->src[0] ? ggml_type_name(op->src[0]->type) : "null")
+                         << ", src1: " << (op->src[1] ? ggml_type_name(op->src[1]->type) : "null"));
+    } else {
+        WEBGPU_LOG_DEBUG("ggml_webgpu op supported: "
+                         << ggml_op_name(op->op) << " with types dst: " << ggml_type_name(op->type)
+                         << ", src0: " << (op->src[0] ? ggml_type_name(op->src[0]->type) : "null")
+                         << ", src1: " << (op->src[1] ? ggml_type_name(op->src[1]->type) : "null"));
     }
-#endif
     return supports_op;
 }
 
@@ -1406,16 +2822,29 @@ static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) {
 }
 
 // TODO: Does this need to be thread safe? Is it only called once?
+// TODO: move most logic to device_init function so backend can be freed/initialized properly
 // Only one device is supported for now
 static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
     GGML_ASSERT(index == 0);
     WEBGPU_LOG_DEBUG("ggml_backend_reg_get_device()");
 
+    WEBGPU_CPU_PROFILE_TOTAL_START(reg_get_device);
+
     ggml_backend_webgpu_reg_context * reg_ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
 
     webgpu_context ctx = reg_ctx->webgpu_ctx;
 
     wgpu::RequestAdapterOptions options = {};
+
+#ifndef __EMSCRIPTEN__
+    // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215
+    const char * const          adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" };
+    wgpu::DawnTogglesDescriptor adapterTogglesDesc;
+    adapterTogglesDesc.enabledToggles     = adapterEnabledToggles;
+    adapterTogglesDesc.enabledToggleCount = 2;
+    options.nextInChain                   = &adapterTogglesDesc;
+#endif
+
     ctx->instance.WaitAny(ctx->instance.RequestAdapter(
                               &options, wgpu::CallbackMode::AllowSpontaneous,
                               [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
@@ -1429,15 +2858,61 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     GGML_ASSERT(ctx->adapter != nullptr);
 
     ctx->adapter.GetLimits(&ctx->limits);
-    ctx->max_wg_size_x = 288;  // default value
 
     wgpu::AdapterInfo info{};
+#ifndef __EMSCRIPTEN__
+    wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{};
+    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        info.nextInChain = &subgroup_matrix_configs;
+    }
+#endif
     ctx->adapter.GetInfo(&info);
 
+    wgpu::SupportedFeatures features;
+    ctx->adapter.GetFeatures(&features);
+    // we require f16 support
+    GGML_ASSERT(ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16));
+
+#ifndef __EMSCRIPTEN__
+    // Only support square f16 matrices of size 8 or 16 for now
+    bool valid_subgroup_matrix_config = false;
+    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) {
+            const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i];
+            if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) &&
+                config.componentType == wgpu::SubgroupMatrixComponentType::F16 &&
+                config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) {
+                ctx->sg_mat_m                = config.M;
+                ctx->sg_mat_n                = config.N;
+                ctx->sg_mat_k                = config.K;
+                valid_subgroup_matrix_config = true;
+                break;
+            }
+        }
+    }
+
+    ctx->supports_subgroup_matrix = valid_subgroup_matrix_config;
+#endif
+    // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate.
+    // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter.
+    ctx->max_subgroup_size = info.subgroupMaxSize;
+
     // Initialize device
-    std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
-                                                         wgpu::FeatureName::ImplicitDeviceSynchronization };
-    wgpu::DeviceDescriptor         dev_desc;
+    std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16 };
+
+#ifndef __EMSCRIPTEN__
+    required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization);
+    if (ctx->supports_subgroup_matrix) {
+        required_features.push_back(wgpu::FeatureName::Subgroups);
+        required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix);
+    }
+#endif
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    required_features.push_back(wgpu::FeatureName::TimestampQuery);
+#endif
+
+    wgpu::DeviceDescriptor dev_desc;
     dev_desc.requiredLimits       = &ctx->limits;
     dev_desc.requiredFeatures     = required_features.data();
     dev_desc.requiredFeatureCount = required_features.size();
@@ -1445,15 +2920,35 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
         wgpu::CallbackMode::AllowSpontaneous,
         [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
             GGML_UNUSED(device);
-            GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason),
-                           std::string(message).c_str());
+            GGML_UNUSED(reason);
+            GGML_UNUSED(message);
+            //TODO: uncomment once proper free logic is in place
+            //GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason),
+            //std::string(message).c_str());
         });
     dev_desc.SetUncapturedErrorCallback(
         [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
             GGML_UNUSED(device);
-            GGML_LOG_ERROR("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
-                           std::string(message).c_str());
+            GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
+                       std::string(message).c_str());
         });
+
+#ifndef __EMSCRIPTEN__
+    // Enable Dawn-specific toggles to increase native performance
+    // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these,
+    //       only for native performance?
+    const char * const deviceEnabledToggles[]  = { "skip_validation", "disable_robustness", "disable_workgroup_init",
+                                                   "disable_polyfills_on_integer_div_and_mod" };
+    const char * const deviceDisabledToggles[] = { "timestamp_quantization" };
+    wgpu::DawnTogglesDescriptor deviceTogglesDesc;
+    deviceTogglesDesc.enabledToggles      = deviceEnabledToggles;
+    deviceTogglesDesc.enabledToggleCount  = 4;
+    deviceTogglesDesc.disabledToggles     = deviceDisabledToggles;
+    deviceTogglesDesc.disabledToggleCount = 1;
+
+    dev_desc.nextInChain = &deviceTogglesDesc;
+#endif
+
     ctx->instance.WaitAny(ctx->adapter.RequestDevice(
                               &dev_desc, wgpu::CallbackMode::AllowSpontaneous,
                               [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
@@ -1474,6 +2969,15 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     ctx->param_buf_pool.init(ctx->device, WEBGPU_NUM_PARAM_BUFS, WEBGPU_PARAMS_BUF_SIZE_BYTES,
                              wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
                              wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // Initialize buffer pool for timestamp queries (profiling)
+    ctx->timestamp_query_buf_pool.init(ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS,
+                                       WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES,
+                                       wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc,
+                                       wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst);
+#endif
+
     ctx->set_rows_error_buf_pool.init(ctx->device, WEBGPU_NUM_SET_ROWS_ERROR_BUFS, WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
                                       wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
                                       wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
@@ -1484,8 +2988,15 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     ggml_webgpu_init_get_rows_pipeline(ctx);
     ggml_webgpu_init_cpy_pipeline(ctx);
     ggml_webgpu_init_add_pipeline(ctx);
+    ggml_webgpu_init_sub_pipeline(ctx);
     ggml_webgpu_init_mul_pipeline(ctx);
+    ggml_webgpu_init_div_pipeline(ctx);
     ggml_webgpu_init_rms_norm_pipeline(ctx);
+    ggml_webgpu_init_rope_pipeline(ctx);
+    ggml_webgpu_init_glu_pipeline(ctx);
+    ggml_webgpu_init_scale_pipeline(ctx);
+    ggml_webgpu_init_soft_max_pipeline(ctx);
+    ggml_webgpu_init_unary_pipeline(ctx);
 
 #ifdef GGML_WEBGPU_DEBUG
     // Initialize debug buffers
@@ -1512,6 +3023,8 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
         /* .reg     = */ reg,
         /* .context = */ &device_ctx,
     };
+
+    WEBGPU_CPU_PROFILE_TOTAL_END(reg_get_device, ctx);
     return &device;
 }
 
@@ -1538,7 +3051,23 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
     std::vector<wgpu::InstanceFeatureName> instance_features = { wgpu::InstanceFeatureName::TimedWaitAny };
     instance_descriptor.requiredFeatures                     = instance_features.data();
     instance_descriptor.requiredFeatureCount                 = instance_features.size();
-    webgpu_ctx->instance                                     = wgpu::CreateInstance(&instance_descriptor);
+
+#ifndef __EMSCRIPTEN__
+    const char * const          instanceEnabledToggles[] = { "allow_unsafe_apis" };
+    wgpu::DawnTogglesDescriptor instanceTogglesDesc;
+    instanceTogglesDesc.enabledToggles     = instanceEnabledToggles;
+    instanceTogglesDesc.enabledToggleCount = 1;
+    instance_descriptor.nextInChain        = &instanceTogglesDesc;
+#endif
+
+    webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
+
+#ifdef __EMSCRIPTEN__
+    if (webgpu_ctx->instance == nullptr) {
+        GGML_LOG_ERROR("ggml_webgpu: Failed to create WebGPU instance. Make sure either -sASYNCIFY or -sJSPI is set\n");
+        return nullptr;
+    }
+#endif
     GGML_ASSERT(webgpu_ctx->instance != nullptr);
 
     static ggml_backend_reg reg = {