whisper.rn 0.5.0 → 0.5.1

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 +89,19 @@ 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;
+    }
+}
+
 //
 // logging
 //
@@ -329,6 +342,10 @@ 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);
@@ -570,27 +587,27 @@ 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 +619,22 @@ 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);
+}
+
 #ifdef __cplusplus
 }
 #endif

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

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

@@ -39,18 +39,13 @@ extern "C" {
 // user-code should use only these functions
 //
 
+// TODO: remove in the future
 WSP_GGML_BACKEND_API wsp_ggml_backend_t wsp_ggml_backend_metal_init(void);
 
 WSP_GGML_BACKEND_API bool wsp_ggml_backend_is_metal(wsp_ggml_backend_t backend);
 
-WSP_GGML_DEPRECATED(
-        WSP_GGML_BACKEND_API wsp_ggml_backend_buffer_t wsp_ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
-
 WSP_GGML_BACKEND_API void wsp_ggml_backend_metal_set_abort_callback(wsp_ggml_backend_t backend, wsp_ggml_abort_callback abort_callback, void * user_data);
 
-WSP_GGML_BACKEND_API wsp_ggml_backend_buffer_type_t wsp_ggml_backend_metal_buffer_type(void);
-
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

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

@@ -244,6 +244,13 @@
 #define WSP_GGML_MROPE_SECTIONS   4
 
 #define WSP_GGML_UNUSED(x) (void)(x)
+#ifdef __CUDACC__
+template<typename... Args>
+__host__ __device__ constexpr inline void wsp_ggml_unused_vars_impl(Args&&...) noexcept {}
+#define WSP_GGML_UNUSED_VARS(...) wsp_ggml_unused_vars_impl(__VA_ARGS__)
+#else
+#define WSP_GGML_UNUSED_VARS(...) do { (void)sizeof((__VA_ARGS__, 0)); } while(0)
+#endif // __CUDACC__
 
 #define WSP_GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
@@ -277,19 +284,19 @@
 //    WSP_GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb);
 //
 #define WSP_GGML_TENSOR_LOCALS_1(type, prefix, pointer, array) \
-    const type prefix##0 = (pointer)->array[0]; \
+    const type prefix##0 = (pointer) ? (pointer)->array[0] : 0; \
     WSP_GGML_UNUSED(prefix##0);
 #define WSP_GGML_TENSOR_LOCALS_2(type, prefix, pointer, array) \
     WSP_GGML_TENSOR_LOCALS_1    (type, prefix, pointer, array) \
-    const type prefix##1 = (pointer)->array[1]; \
+    const type prefix##1 = (pointer) ? (pointer)->array[1] : 0; \
     WSP_GGML_UNUSED(prefix##1);
 #define WSP_GGML_TENSOR_LOCALS_3(type, prefix, pointer, array) \
     WSP_GGML_TENSOR_LOCALS_2    (type, prefix, pointer, array) \
-    const type prefix##2 = (pointer)->array[2]; \
+    const type prefix##2 = (pointer) ? (pointer)->array[2] : 0; \
     WSP_GGML_UNUSED(prefix##2);
 #define WSP_GGML_TENSOR_LOCALS(type, prefix, pointer, array) \
     WSP_GGML_TENSOR_LOCALS_3  (type, prefix, pointer, array) \
-    const type prefix##3 = (pointer)->array[3]; \
+    const type prefix##3 = (pointer) ? (pointer)->array[3] : 0; \
     WSP_GGML_UNUSED(prefix##3);
 
 #define WSP_GGML_TENSOR_UNARY_OP_LOCALS \
@@ -504,7 +511,9 @@ extern "C" {
         WSP_GGML_OP_CONV_TRANSPOSE_1D,
         WSP_GGML_OP_IM2COL,
         WSP_GGML_OP_IM2COL_BACK,
+        WSP_GGML_OP_IM2COL_3D,
         WSP_GGML_OP_CONV_2D,
+        WSP_GGML_OP_CONV_3D,
         WSP_GGML_OP_CONV_2D_DW,
         WSP_GGML_OP_CONV_TRANSPOSE_2D,
         WSP_GGML_OP_POOL_1D,
@@ -1395,6 +1404,7 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b);
 
+    // note: casting from f32 to i32 will discard the fractional part
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_cast(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -1519,7 +1529,11 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
-    // supports 3D: a->ne[2] == b->ne[1]
+    // supports 4D a:
+    // a     [n_embd, ne1, ne2, ne3]
+    // b I32 [n_rows, ne2, ne3, 1]
+    //
+    // return [n_embd, n_rows, ne2, ne3]
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_get_rows(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,  // data
@@ -1862,6 +1876,41 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_im2col_3d(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            int64_t               IC,
+            int                   s0, // stride width
+            int                   s1, // stride height
+            int                   s2, // stride depth
+            int                   p0, // padding width
+            int                   p1, // padding height
+            int                   p2, // padding depth
+            int                   d0, // dilation width
+            int                   d1, // dilation height
+            int                   d2, // dilation depth
+            enum wsp_ggml_type        dst_type);
+
+    // a: [OC*IC, KD, KH, KW]
+    // b: [N*IC, ID, IH, IW]
+    // result: [N*OC, OD, OH, OW]
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_3d(
+                struct wsp_ggml_context * ctx,
+                struct wsp_ggml_tensor  * a,
+                struct wsp_ggml_tensor  * b,
+                int64_t               IC,
+                int                   s0, // stride width
+                int                   s1, // stride height
+                int                   s2, // stride depth
+                int                   p0, // padding width
+                int                   p1, // padding height
+                int                   p2, // padding depth
+                int                   d0, // dilation width
+                int                   d1, // dilation height
+                int                   d2  // dilation depth
+        );
+
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
     // padding is zero
@@ -1933,6 +1982,23 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_3d_direct(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,   // kernel [KW, KH, KD, IC * OC]
+            struct wsp_ggml_tensor  * b,   // input  [W, H, D, C * N]
+            int                   s0,  // stride
+            int                   s1,
+            int                   s2,
+            int                   p0,  // padding
+            int                   p1,
+            int                   p2,
+            int                   d0,  // dilation
+            int                   d1,
+            int                   d2,
+            int                   n_channels,
+            int                   n_batch,
+            int                   n_channels_out);
+
     enum wsp_ggml_op_pool {
         WSP_GGML_OP_POOL_MAX,
         WSP_GGML_OP_POOL_AVG,
@@ -2023,6 +2089,19 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_ext(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                  lp0,
+            int                  rp0,
+            int                  lp1,
+            int                  rp1,
+            int                  lp2,
+            int                  rp2,
+            int                  lp3,
+            int                  rp3
+            );
+
     // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_reflect_1d(
             struct wsp_ggml_context * ctx,

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/whisper.h

@@ -526,6 +526,7 @@ extern "C" {
         // use whisper_tokenize() to convert text to tokens
         // maximum of whisper_n_text_ctx()/2 tokens are used (typically 224)
         const char * initial_prompt;
+        bool carry_initial_prompt; // if true, always prepend initial_prompt to every decode window (may reduce conditioning on previous text)
         const whisper_token * prompt_tokens;
         int prompt_n_tokens;
 

package/ios/rnwhisper.xcframework/ios-arm64_x86_64-simulator/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 {

package/ios/rnwhisper.xcframework/ios-arm64_x86_64-simulator/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;
     };
 
@@ -302,11 +314,15 @@ 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_x86_64-simulator/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);