whisper.rn 0.5.0 → 0.5.2

package/ios/CMakeLists.txt

@@ -55,7 +55,12 @@ add_library(rnwhisper SHARED
     ${SOURCE_DIR}/ggml-cpu/binary-ops.cpp
     ${SOURCE_DIR}/ggml-cpu/vec.cpp
     ${SOURCE_DIR}/ggml-cpu/ops.cpp
-    ${SOURCE_DIR}/ggml-metal.m
+    ${SOURCE_DIR}/ggml-metal/ggml-metal.cpp
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-common.cpp
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-device.cpp
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-context.m
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-device.m
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-ops.cpp
     ${SOURCE_DIR}/ggml-opt.cpp
     ${SOURCE_DIR}/ggml-threading.cpp
     ${SOURCE_DIR}/ggml-quants.c

package/ios/RNWhisperContext.mm

@@ -168,6 +168,7 @@ static void* retained_log_block = nullptr;
     self->recordState.sliceNSamples.push_back(0);
 
     self->recordState.job = rnwhisper::job_new(jobId, [self createParams:options jobId:jobId]);
+    self->recordState.job->n_processors = options[@"nProcessors"] != nil ? [options[@"nProcessors"] intValue] : 1;
     self->recordState.job->set_realtime_params(
         {
             .use_vad = options[@"useVad"] != nil ? [options[@"useVad"] boolValue] : false,
@@ -471,6 +472,7 @@ struct rnwhisper_segments_callback_data {
         }
 
         rnwhisper::job* job = rnwhisper::job_new(jobId, params);
+        job->n_processors = options[@"nProcessors"] != nil ? [options[@"nProcessors"] intValue] : 1;
         self->recordState.job = job;
         int code = [self fullTranscribe:job audioData:audioData audioDataCount:audioDataCount];
         rnwhisper::job_remove(jobId);
@@ -572,7 +574,7 @@ struct rnwhisper_segments_callback_data {
   audioDataCount:(int)audioDataCount
 {
     whisper_reset_timings(self->ctx);
-    int code = whisper_full(self->ctx, job->params, audioData, audioDataCount);
+    int code = whisper_full_parallel(self->ctx, job->params, audioData, audioDataCount, job->n_processors);
     if (job && job->is_aborted()) code = -999;
     // if (code == 0) {
     //     whisper_print_timings(self->ctx);

package/ios/RNWhisperVadContext.mm

@@ -20,27 +20,28 @@
 
 #ifdef WSP_GGML_USE_METAL
     if (ctx_params.use_gpu) {
-        ctx_params.gpu_device = 0;
+        // TODO: GPU VAD is forced disabled until the performance is improved (ref: whisper.cpp/whisper_vad_init_context)
+        ctx_params.use_gpu = false;
+        // ctx_params.gpu_device = 0;
 
-        id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+        // id<MTLDevice> device = MTLCreateSystemDefaultDevice();
 
-        // Check ggml-metal availability
-        BOOL supportsGgmlMetal = [device supportsFamily:MTLGPUFamilyApple7];
-        if (@available(iOS 16.0, tvOS 16.0, *)) {
-            supportsGgmlMetal = supportsGgmlMetal && [device supportsFamily:MTLGPUFamilyMetal3];
-        }
-        if (!supportsGgmlMetal) {
-          ctx_params.use_gpu = false;
-            reasonNoMetal = @"Metal is not supported in this device";
-        }
+        // // Check ggml-metal availability
+        // BOOL supportsGgmlMetal = [device supportsFamily:MTLGPUFamilyApple7];
+        // if (@available(iOS 16.0, tvOS 16.0, *)) {
+        //     supportsGgmlMetal = supportsGgmlMetal && [device supportsFamily:MTLGPUFamilyMetal3];
+        // }
+        // if (!supportsGgmlMetal) {
+        //   ctx_params.use_gpu = false;
+        //     reasonNoMetal = @"Metal is not supported in this device";
+        // }
+        // device = nil;
 
 #if TARGET_OS_SIMULATOR
         // Use the backend, but no layers because not supported fully on simulator
         ctx_params.use_gpu = false;
         reasonNoMetal = @"Metal is not supported in simulator";
 #endif
-
-        device = nil;
     }
 #endif // WSP_GGML_USE_METAL
 

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-backend-impl.h

@@ -8,7 +8,7 @@
 extern "C" {
 #endif
 
-    #define WSP_GGML_BACKEND_API_VERSION 1
+    #define WSP_GGML_BACKEND_API_VERSION 2
 
     //
     // Backend buffer type
@@ -114,6 +114,9 @@ extern "C" {
         void (*event_record)(wsp_ggml_backend_t backend, wsp_ggml_backend_event_t event);
         // wait for an event on on a different stream
         void (*event_wait)  (wsp_ggml_backend_t backend, wsp_ggml_backend_event_t event);
+
+        // (optional) sort/optimize the nodes in the graph
+        void                      (*graph_optimize)    (wsp_ggml_backend_t backend, struct wsp_ggml_cgraph * cgraph);
     };
 
     struct wsp_ggml_backend {
@@ -206,9 +209,6 @@ extern "C" {
         void * context;
     };
 
-    // Internal backend registry API
-    WSP_GGML_API void wsp_ggml_backend_register(wsp_ggml_backend_reg_t reg);
-
     // Add backend dynamic loading support to the backend
 
     // Initialize the backend

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-backend.h

@@ -132,6 +132,8 @@ extern "C" {
         WSP_GGML_BACKEND_DEVICE_TYPE_CPU,
         // GPU device using dedicated memory
         WSP_GGML_BACKEND_DEVICE_TYPE_GPU,
+        // integrated GPU device using host memory
+        WSP_GGML_BACKEND_DEVICE_TYPE_IGPU,
         // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
         WSP_GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
@@ -150,11 +152,21 @@ extern "C" {
 
     // all the device properties
     struct wsp_ggml_backend_dev_props {
+        // device name
         const char * name;
+        // device description
         const char * description;
+        // device free memory in bytes
         size_t memory_free;
+        // device total memory in bytes
         size_t memory_total;
+        // device type
         enum wsp_ggml_backend_dev_type type;
+        // device id
+        //   for PCI devices, this should be the PCI bus id formatted as "domain:bus:device.function" (e.g. "0000:01:00.0")
+        //   if the id is unknown, this should be NULL
+        const char * device_id;
+        // device capabilities
         struct wsp_ggml_backend_dev_caps caps;
     };
 
@@ -203,6 +215,8 @@ extern "C" {
     // Backend registry
     //
 
+    WSP_GGML_API void wsp_ggml_backend_register(wsp_ggml_backend_reg_t reg);
+
     WSP_GGML_API void wsp_ggml_backend_device_register(wsp_ggml_backend_dev_t device);
 
     // Backend (reg) enumeration
@@ -302,11 +316,15 @@ extern "C" {
     WSP_GGML_API int                  wsp_ggml_backend_sched_get_n_splits(wsp_ggml_backend_sched_t sched);
     WSP_GGML_API int                  wsp_ggml_backend_sched_get_n_copies(wsp_ggml_backend_sched_t sched);
 
-    WSP_GGML_API size_t               wsp_ggml_backend_sched_get_buffer_size(wsp_ggml_backend_sched_t sched, wsp_ggml_backend_t backend);
+    WSP_GGML_API wsp_ggml_backend_buffer_type_t wsp_ggml_backend_sched_get_buffer_type(wsp_ggml_backend_sched_t sched, wsp_ggml_backend_t backend);
+    WSP_GGML_API size_t                     wsp_ggml_backend_sched_get_buffer_size(wsp_ggml_backend_sched_t sched, wsp_ggml_backend_t backend);
 
     WSP_GGML_API void                 wsp_ggml_backend_sched_set_tensor_backend(wsp_ggml_backend_sched_t sched, struct wsp_ggml_tensor * node, wsp_ggml_backend_t backend);
     WSP_GGML_API wsp_ggml_backend_t       wsp_ggml_backend_sched_get_tensor_backend(wsp_ggml_backend_sched_t sched, struct wsp_ggml_tensor * node);
 
+    // Split graph without allocating it
+    WSP_GGML_API void                 wsp_ggml_backend_sched_split_graph(wsp_ggml_backend_sched_t sched, struct wsp_ggml_cgraph * graph);
+
     // Allocate and compute graph on the backend scheduler
     WSP_GGML_API bool                 wsp_ggml_backend_sched_alloc_graph(wsp_ggml_backend_sched_t sched, struct wsp_ggml_cgraph * graph); // returns success
     WSP_GGML_API enum wsp_ggml_status     wsp_ggml_backend_sched_graph_compute(wsp_ggml_backend_sched_t sched, struct wsp_ggml_cgraph * graph);

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-cpu.h

@@ -101,7 +101,6 @@ extern "C" {
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_riscv_v    (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vsx        (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vxe        (void);
-    WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_nnpa       (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_wasm_simd  (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_llamafile  (void);
 
@@ -135,6 +134,7 @@ extern "C" {
     WSP_GGML_BACKEND_API wsp_ggml_backend_reg_t wsp_ggml_backend_cpu_reg(void);
 
     WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
+    WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_i32 (const float *,     int32_t *, int64_t);
     WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_fp16(const float *, wsp_ggml_fp16_t *, int64_t);
     WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp16_to_fp32(const wsp_ggml_fp16_t *, float *, int64_t);
     WSP_GGML_BACKEND_API void wsp_ggml_cpu_fp32_to_bf16(const float *, wsp_ggml_bf16_t *, int64_t);

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-impl.h

@@ -73,7 +73,7 @@ static inline int wsp_ggml_up(int n, int m) {
     return (n + m - 1) & ~(m - 1);
 }
 
-// TODO: move to ggml.h?
+// TODO: move to ggml.h? (won't be able to inline)
 static bool wsp_ggml_are_same_layout(const struct wsp_ggml_tensor * a, const struct wsp_ggml_tensor * b) {
     if (a->type != b->type) {
         return false;
@@ -89,6 +89,22 @@ static bool wsp_ggml_are_same_layout(const struct wsp_ggml_tensor * a, const str
     return true;
 }
 
+static bool wsp_ggml_op_is_empty(enum wsp_ggml_op op) {
+    switch (op) {
+        case WSP_GGML_OP_NONE:
+        case WSP_GGML_OP_RESHAPE:
+        case WSP_GGML_OP_TRANSPOSE:
+        case WSP_GGML_OP_VIEW:
+        case WSP_GGML_OP_PERMUTE:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static inline float wsp_ggml_softplus(float input) {
+    return (input > 20.0f) ? input : logf(1 + expf(input));
+}
 //
 // logging
 //
@@ -329,6 +345,10 @@ struct wsp_ggml_cgraph {
 // if you need the gradients, get them from the original graph
 struct wsp_ggml_cgraph wsp_ggml_graph_view(struct wsp_ggml_cgraph * cgraph, int i0, int i1);
 
+// ggml-alloc.c: true if the operation can reuse memory from its sources
+WSP_GGML_API bool wsp_ggml_op_can_inplace(enum wsp_ggml_op op);
+
+
 // Memory allocation
 
 WSP_GGML_API void * wsp_ggml_aligned_malloc(size_t size);
@@ -545,14 +565,23 @@ static inline wsp_ggml_bf16_t wsp_ggml_compute_fp32_to_bf16(float s) {
 #define WSP_GGML_FP32_TO_BF16(x) wsp_ggml_compute_fp32_to_bf16(x)
 #define WSP_GGML_BF16_TO_FP32(x) wsp_ggml_compute_bf16_to_fp32(x)
 
+static inline int32_t wsp_ggml_node_get_use_count(const struct wsp_ggml_cgraph * cgraph, int node_idx) {
+    const struct wsp_ggml_tensor * node = cgraph->nodes[node_idx];
+
+    size_t hash_pos = wsp_ggml_hash_find(&cgraph->visited_hash_set, node);
+    if (!wsp_ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        return 0;
+    }
+    return cgraph->use_counts[hash_pos];
+}
+
 // return true if the node's results are only used by N other nodes
 // and can be fused into their calculations.
 static inline bool wsp_ggml_node_has_n_uses(const struct wsp_ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
     const struct wsp_ggml_tensor * node = cgraph->nodes[node_idx];
 
     // check the use count against how many we're replacing
-    size_t hash_pos = wsp_ggml_hash_find(&cgraph->visited_hash_set, node);
-    if (!wsp_ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
+    if (wsp_ggml_node_get_use_count(cgraph, node_idx) != n_uses) {
         return false;
     }
 
@@ -570,27 +599,27 @@ static inline bool wsp_ggml_node_has_n_uses(const struct wsp_ggml_cgraph * cgrap
     return true;
 }
 
-// Returns true if nodes [i, i+ops.size()) are the sequence of wsp_ggml_ops in ops[]
+// Returns true if nodes with indices { node_idxs } are the sequence of wsp_ggml_ops in ops[]
 // and are fusable. Nodes are considered fusable according to this function if:
 // - all nodes except the last have only one use and are not views/outputs (see wsp_ggml_node_has_N_uses).
 // - all nodes except the last are a src of the following node.
 // - all nodes are the same shape.
 // TODO: Consider allowing WSP_GGML_OP_NONE nodes in between
-static inline bool wsp_ggml_can_fuse(const struct wsp_ggml_cgraph * cgraph, int node_idx, const enum wsp_ggml_op * ops, int num_ops) {
-    if (node_idx + num_ops > cgraph->n_nodes) {
-        return false;
-    }
-
+static inline bool wsp_ggml_can_fuse_ext(const struct wsp_ggml_cgraph * cgraph, const int * node_idxs, const enum wsp_ggml_op * ops, int num_ops) {
     for (int i = 0; i < num_ops; ++i) {
-        struct wsp_ggml_tensor * node = cgraph->nodes[node_idx + i];
+        if (node_idxs[i] >= cgraph->n_nodes) {
+            return false;
+        }
+
+        struct wsp_ggml_tensor * node = cgraph->nodes[node_idxs[i]];
         if (node->op != ops[i]) {
             return false;
         }
-        if (i < num_ops - 1 && !wsp_ggml_node_has_n_uses(cgraph, node_idx + i, 1)) {
+        if (i < num_ops - 1 && !wsp_ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
             return false;
         }
         if (i > 0) {
-            struct wsp_ggml_tensor * prev = cgraph->nodes[node_idx + i - 1];
+            struct wsp_ggml_tensor * prev = cgraph->nodes[node_idxs[i - 1]];
             if (node->src[0] != prev && node->src[1] != prev) {
                 return false;
             }
@@ -602,6 +631,52 @@ static inline bool wsp_ggml_can_fuse(const struct wsp_ggml_cgraph * cgraph, int
     return true;
 }
 
+// same as above, for sequential indices starting at node_idx
+static inline bool wsp_ggml_can_fuse(const struct wsp_ggml_cgraph * cgraph, int node_idx, const enum wsp_ggml_op * ops, int num_ops) {
+    assert(num_ops < 32);
+
+    if (node_idx + num_ops > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+    for (int i = 0; i < num_ops; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return wsp_ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
+}
+
+WSP_GGML_API bool wsp_ggml_can_fuse_subgraph_ext(const struct wsp_ggml_cgraph * cgraph,
+                                         const int *                node_idxs,
+                                         int                        count,
+                                         const enum wsp_ggml_op *       ops,
+                                         const int *                outputs,
+                                         int                        num_outputs);
+
+// Returns true if the subgraph formed by {node_idxs} can be fused
+// checks whethers all nodes which are not part of outputs can be elided
+// by checking if their num_uses are confined to the subgraph
+static inline bool wsp_ggml_can_fuse_subgraph(const struct wsp_ggml_cgraph * cgraph,
+                                          int                        node_idx,
+                                          int                        count,
+                                          const enum wsp_ggml_op *       ops,
+                                          const int *                outputs,
+                                          int                        num_outputs) {
+    WSP_GGML_ASSERT(count < 32);
+    if (node_idx + count > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+
+    for (int i = 0; i < count; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return wsp_ggml_can_fuse_subgraph_ext(cgraph, idxs, count, ops, outputs, num_outputs);
+}
+
 #ifdef __cplusplus
 }
 #endif
@@ -615,6 +690,13 @@ inline bool wsp_ggml_can_fuse(const struct wsp_ggml_cgraph * cgraph, int node_id
     return wsp_ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
 }
 
+inline bool wsp_ggml_can_fuse_subgraph(const struct wsp_ggml_cgraph *          cgraph,
+                                   int                                 start_idx,
+                                   std::initializer_list<enum wsp_ggml_op> ops,
+                                   std::initializer_list<int>          outputs = {}) {
+    return wsp_ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
+}
+
 // expose GGUF internals for test code
 WSP_GGML_API size_t wsp_gguf_type_size(enum wsp_gguf_type type);
 WSP_GGML_API struct wsp_gguf_context * wsp_gguf_init_from_file_impl(FILE * file, struct wsp_gguf_init_params params);

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-metal-impl.h

@@ -1,5 +1,5 @@
-#ifndef WSP_GGML_METAL_IMPL
-#define WSP_GGML_METAL_IMPL
+#ifndef WSP_WSP_WSP_GGML_METAL_IMPL
+#define WSP_WSP_WSP_GGML_METAL_IMPL
 
 // kernel parameters for mat-vec threadgroups
 //
@@ -101,7 +101,7 @@ typedef struct {
     uint64_t nb2;
     uint64_t nb3;
     int32_t  dim;
-} wsp_ggml_metal_kargs_concat;
+} wsp_wsp_wsp_ggml_metal_kargs_concat;
 
 typedef struct {
     int32_t  ne00;
@@ -130,7 +130,7 @@ typedef struct {
     uint64_t nb3;
     uint64_t offs;
     uint64_t o1[8];
-} wsp_ggml_metal_kargs_bin;
+} wsp_wsp_wsp_ggml_metal_kargs_bin;
 
 typedef struct {
     int64_t ne0;
@@ -139,7 +139,7 @@ typedef struct {
     size_t nb02;
     size_t nb11;
     size_t nb21;
-} wsp_ggml_metal_kargs_add_id;
+} wsp_wsp_wsp_ggml_metal_kargs_add_id;
 
 typedef struct {
     int32_t  ne00;
@@ -158,7 +158,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_repeat;
+} wsp_wsp_wsp_ggml_metal_kargs_repeat;
 
 typedef struct {
     int64_t  ne00;
@@ -177,7 +177,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_cpy;
+} wsp_wsp_wsp_ggml_metal_kargs_cpy;
 
 typedef struct {
     int64_t  ne10;
@@ -192,7 +192,7 @@ typedef struct {
     uint64_t nb3;
     uint64_t offs;
     bool     inplace;
-} wsp_ggml_metal_kargs_set;
+} wsp_wsp_wsp_ggml_metal_kargs_set;
 
 typedef struct {
     int32_t  ne00;
@@ -224,7 +224,7 @@ typedef struct {
     int32_t  sect_1;
     int32_t  sect_2;
     int32_t  sect_3;
-} wsp_ggml_metal_kargs_rope;
+} wsp_wsp_wsp_ggml_metal_kargs_rope;
 
 typedef struct {
     int32_t  ne01;
@@ -255,7 +255,7 @@ typedef struct {
     float    m1;
     int32_t  n_head_log2;
     float    logit_softcap;
-} wsp_ggml_metal_kargs_flash_attn_ext;
+} wsp_wsp_wsp_ggml_metal_kargs_flash_attn_ext;
 
 typedef struct {
     int32_t  ne00;
@@ -272,7 +272,7 @@ typedef struct {
     int32_t  ne1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mm;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm;
 
 typedef struct {
     int32_t  ne00;
@@ -293,7 +293,7 @@ typedef struct {
     int32_t  ne1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mv;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv;
 
 typedef struct {
     int32_t  ne00;
@@ -317,7 +317,7 @@ typedef struct {
     int16_t  nsg;
     int16_t  nxpsg;
     int16_t  r1ptg;
-} wsp_ggml_metal_kargs_mul_mv_ext;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv_ext;
 
 typedef struct {
     int32_t  ne10;
@@ -328,7 +328,7 @@ typedef struct {
     uint64_t nbh11;
     int32_t  ne20;  // n_expert_used
     uint64_t nb21;
-} wsp_ggml_metal_kargs_mul_mm_id_map0;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id_map0;
 
 typedef struct {
     int32_t  ne20; // n_expert_used
@@ -339,7 +339,7 @@ typedef struct {
     int32_t  ne0;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_mul_mm_id_map1;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id_map1;
 
 typedef struct {
     int32_t  ne00;
@@ -356,7 +356,7 @@ typedef struct {
     int32_t  neh1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mm_id;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id;
 
 typedef struct {
     int32_t  nei0;
@@ -378,14 +378,14 @@ typedef struct {
     int32_t  ne0;
     int32_t  ne1;
     uint64_t nb1;
-} wsp_ggml_metal_kargs_mul_mv_id;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv_id;
 
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
     uint64_t nb01;
     float    eps;
-} wsp_ggml_metal_kargs_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_norm;
 
 typedef struct {
     int32_t  ne00;
@@ -400,14 +400,14 @@ typedef struct {
     uint64_t nbf1[3];
     uint64_t nbf2[3];
     uint64_t nbf3[3];
-} wsp_ggml_metal_kargs_rms_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_rms_norm;
 
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
     uint64_t nb01;
     float    eps;
-} wsp_ggml_metal_kargs_l2_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_l2_norm;
 
 typedef struct {
     int64_t  ne00;
@@ -418,7 +418,7 @@ typedef struct {
     uint64_t nb02;
     int32_t  n_groups;
     float    eps;
-} wsp_ggml_metal_kargs_group_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_group_norm;
 
 typedef struct {
     int32_t  IC;
@@ -427,7 +427,7 @@ typedef struct {
     int32_t  s0;
     uint64_t nb0;
     uint64_t nb1;
-} wsp_ggml_metal_kargs_conv_transpose_1d;
+} wsp_wsp_wsp_ggml_metal_kargs_conv_transpose_1d;
 
 typedef struct {
     uint64_t  ofs0;
@@ -445,7 +445,7 @@ typedef struct {
     int32_t  KH;
     int32_t  KW;
     int32_t  KHW; // KH * KW, pre-computed on CPU to save GPU resources
-} wsp_ggml_metal_kargs_im2col;
+} wsp_wsp_wsp_ggml_metal_kargs_im2col;
 
 typedef struct{
     int32_t  ne00;
@@ -458,7 +458,7 @@ typedef struct{
     int32_t  i10;
     float    alpha;
     float    limit;
-} wsp_ggml_metal_kargs_glu;
+} wsp_wsp_wsp_ggml_metal_kargs_glu;
 
 typedef struct {
     int64_t  ne00;
@@ -485,7 +485,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_sum_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_sum_rows;
 
 typedef struct {
     int32_t  ne00;
@@ -508,13 +508,13 @@ typedef struct {
     float    m0;
     float    m1;
     int32_t  n_head_log2;
-} wsp_ggml_metal_kargs_soft_max;
+} wsp_wsp_wsp_ggml_metal_kargs_soft_max;
 
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
     int      n_past;
-} wsp_ggml_metal_kargs_diag_mask_inf;
+} wsp_wsp_wsp_ggml_metal_kargs_diag_mask_inf;
 
 typedef struct {
     int64_t  ne00;
@@ -533,7 +533,7 @@ typedef struct {
     uint64_t nb0;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_ssm_conv;
+} wsp_wsp_wsp_ggml_metal_kargs_ssm_conv;
 
 typedef struct {
     int64_t  d_state;
@@ -558,7 +558,7 @@ typedef struct {
     uint64_t nb51;
     uint64_t nb52;
     uint64_t nb53;
-} wsp_ggml_metal_kargs_ssm_scan;
+} wsp_wsp_wsp_ggml_metal_kargs_ssm_scan;
 
 typedef struct {
     int64_t  ne00;
@@ -569,7 +569,7 @@ typedef struct {
     uint64_t nb11;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_get_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_get_rows;
 
 typedef struct {
     int32_t  nk0;
@@ -585,7 +585,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_set_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_set_rows;
 
 typedef struct {
     int64_t  ne00;
@@ -608,7 +608,7 @@ typedef struct {
     float    sf1;
     float    sf2;
     float    sf3;
-} wsp_ggml_metal_kargs_upscale;
+} wsp_wsp_wsp_ggml_metal_kargs_upscale;
 
 typedef struct {
     int64_t  ne00;
@@ -627,7 +627,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_pad;
+} wsp_wsp_wsp_ggml_metal_kargs_pad;
 
 typedef struct {
     int64_t  ne00;
@@ -648,28 +648,28 @@ typedef struct {
     uint64_t nb3;
     int32_t  p0;
     int32_t  p1;
-} wsp_ggml_metal_kargs_pad_reflect_1d;
+} wsp_wsp_wsp_ggml_metal_kargs_pad_reflect_1d;
 
 typedef struct {
     uint64_t nb1;
     int      dim;
     int      max_period;
-} wsp_ggml_metal_kargs_timestep_embedding;
+} wsp_wsp_wsp_ggml_metal_kargs_timestep_embedding;
 
 typedef struct {
     float    slope;
-} wsp_ggml_metal_kargs_leaky_relu;
+} wsp_wsp_wsp_ggml_metal_kargs_leaky_relu;
 
 typedef struct {
     int64_t  ncols;
     int64_t  ncols_pad;
-} wsp_ggml_metal_kargs_argsort;
+} wsp_wsp_wsp_ggml_metal_kargs_argsort;
 
 typedef struct {
     int64_t  ne0;
     float    start;
     float    step;
-} wsp_ggml_metal_kargs_arange;
+} wsp_wsp_wsp_ggml_metal_kargs_arange;
 
 typedef struct {
     int32_t  k0;
@@ -683,6 +683,6 @@ typedef struct {
     int64_t  OH;
     int64_t  OW;
     int64_t  parallel_elements;
-} wsp_ggml_metal_kargs_pool_2d;
+} wsp_wsp_wsp_ggml_metal_kargs_pool_2d;
 
-#endif // WSP_GGML_METAL_IMPL
+#endif // WSP_WSP_WSP_GGML_METAL_IMPL