whisper.rn 0.5.0-rc.9 → 0.5.0

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

@@ -23,6 +23,9 @@
 #define N_R0_Q8_0 4
 #define N_SG_Q8_0 2
 
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+
 #define N_R0_Q2_K 4
 #define N_SG_Q2_K 2
 
@@ -126,8 +129,18 @@ typedef struct {
     uint64_t nb2;
     uint64_t nb3;
     uint64_t offs;
+    uint64_t o1[8];
 } wsp_ggml_metal_kargs_bin;
 
+typedef struct {
+    int64_t ne0;
+    int64_t ne1;
+    size_t nb01;
+    size_t nb02;
+    size_t nb11;
+    size_t nb21;
+} wsp_ggml_metal_kargs_add_id;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -229,14 +242,18 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne32;
+    int32_t  ne33;
     uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
     float    scale;
     float    max_bias;
     float    m0;
     float    m1;
-    uint16_t n_head_log2;
+    int32_t  n_head_log2;
     float    logit_softcap;
 } wsp_ggml_metal_kargs_flash_attn_ext;
 
@@ -373,8 +390,16 @@ typedef struct {
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
-    uint64_t nb01;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    eps;
+    int32_t  nef1[3];
+    int32_t  nef2[3];
+    int32_t  nef3[3];
+    uint64_t nbf1[3];
+    uint64_t nbf2[3];
+    uint64_t nbf3[3];
 } wsp_ggml_metal_kargs_rms_norm;
 
 typedef struct {
@@ -431,6 +456,8 @@ typedef struct{
     uint64_t nb1;
     int32_t  i00;
     int32_t  i10;
+    float    alpha;
+    float    limit;
 } wsp_ggml_metal_kargs_glu;
 
 typedef struct {
@@ -461,14 +488,26 @@ typedef struct {
 } wsp_ggml_metal_kargs_sum_rows;
 
 typedef struct {
-    int64_t  ne00;
-    int64_t  ne01;
-    int64_t  ne02;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    scale;
     float    max_bias;
     float    m0;
     float    m1;
-    uint32_t n_head_log2;
+    int32_t  n_head_log2;
 } wsp_ggml_metal_kargs_soft_max;
 
 typedef struct {
@@ -499,26 +538,26 @@ typedef struct {
 typedef struct {
     int64_t  d_state;
     int64_t  d_inner;
+    int64_t  n_head;
+    int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
-    uint64_t nb00;
+    int64_t  s_off;
     uint64_t nb01;
     uint64_t nb02;
-    uint64_t nb10;
+    uint64_t nb03;
     uint64_t nb11;
     uint64_t nb12;
     uint64_t nb13;
-    uint64_t nb20;
     uint64_t nb21;
     uint64_t nb22;
-    uint64_t nb30;
     uint64_t nb31;
-    uint64_t nb40;
     uint64_t nb41;
     uint64_t nb42;
-    uint64_t nb50;
+    uint64_t nb43;
     uint64_t nb51;
     uint64_t nb52;
+    uint64_t nb53;
 } wsp_ggml_metal_kargs_ssm_scan;
 
 typedef struct {

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

@@ -74,16 +74,26 @@ extern "C" {
         WSP_GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
+    enum wsp_ggml_opt_optimizer_type {
+        WSP_GGML_OPT_OPTIMIZER_TYPE_ADAMW,
+        WSP_GGML_OPT_OPTIMIZER_TYPE_SGD,
+
+        WSP_GGML_OPT_OPTIMIZER_TYPE_COUNT
+    };
+
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
     struct wsp_ggml_opt_optimizer_params {
-        // AdamW optimizer parameters
         struct {
             float alpha; // learning rate
-            float beta1;
-            float beta2;
+            float beta1; // first AdamW momentum
+            float beta2; // second AdamW momentum
             float eps;   // epsilon for numerical stability
-            float wd;    // weight decay for AdamW, use 0.0f to disable
+            float wd;    // weight decay - 0.0f to disable
         } adamw;
+        struct {
+            float alpha; // learning rate
+            float wd;    // weight decay
+        } sgd;
     };
 
     // callback to calculate optimizer parameters prior to a backward pass
@@ -112,8 +122,11 @@ extern "C" {
 
         int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
 
-        wsp_ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
-        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+        wsp_ggml_opt_get_optimizer_params get_opt_pars;    // callback for calculating optimizer parameters
+        void *                        get_opt_pars_ud; // userdata for calculating optimizer parameters
+
+        // only WSP_GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
+        enum wsp_ggml_opt_optimizer_type optimizer;
     };
 
     // get parameters for an optimization context with defaults set where possible
@@ -142,6 +155,10 @@ extern "C" {
     // get the gradient accumulator for a node from the forward graph
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_opt_grad_acc(wsp_ggml_opt_context_t opt_ctx, struct wsp_ggml_tensor * node);
 
+    WSP_GGML_API enum wsp_ggml_opt_optimizer_type wsp_ggml_opt_context_optimizer_type(wsp_ggml_opt_context_t); //TODO consistent naming scheme
+
+    WSP_GGML_API const char * wsp_ggml_opt_optimizer_name(enum wsp_ggml_opt_optimizer_type);
+
     // ====== Optimization Result ======
 
     WSP_GGML_API wsp_ggml_opt_result_t wsp_ggml_opt_result_init(void);
@@ -226,12 +243,14 @@ extern "C" {
             struct wsp_ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             wsp_ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum wsp_ggml_opt_loss_type         loss_type,      // loss to minimize
+            enum wsp_ggml_opt_optimizer_type    optimizer,      // sgd or adamw
             wsp_ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
             int64_t                         nepoch,         // how many times the dataset should be iterated over
             int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
             float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
             bool                            silent);        // whether or not info prints to stderr should be suppressed
 
+
 #ifdef  __cplusplus
 }
 #endif

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

@@ -21,6 +21,8 @@ WSP_GGML_API void wsp_quantize_row_q5_1_ref(const float * WSP_GGML_RESTRICT x, b
 WSP_GGML_API void wsp_quantize_row_q8_0_ref(const float * WSP_GGML_RESTRICT x, block_q8_0 * WSP_GGML_RESTRICT y, int64_t k);
 WSP_GGML_API void wsp_quantize_row_q8_1_ref(const float * WSP_GGML_RESTRICT x, block_q8_1 * WSP_GGML_RESTRICT y, int64_t k);
 
+WSP_GGML_API void wsp_quantize_row_mxfp4_ref(const float * WSP_GGML_RESTRICT x, block_mxfp4 * WSP_GGML_RESTRICT y, int64_t k);
+
 WSP_GGML_API void wsp_quantize_row_q2_K_ref(const float * WSP_GGML_RESTRICT x, block_q2_K * WSP_GGML_RESTRICT y, int64_t k);
 WSP_GGML_API void wsp_quantize_row_q3_K_ref(const float * WSP_GGML_RESTRICT x, block_q3_K * WSP_GGML_RESTRICT y, int64_t k);
 WSP_GGML_API void wsp_quantize_row_q4_K_ref(const float * WSP_GGML_RESTRICT x, block_q4_K * WSP_GGML_RESTRICT y, int64_t k);
@@ -45,6 +47,8 @@ WSP_GGML_API void wsp_dewsp_quantize_row_q5_1(const block_q5_1 * WSP_GGML_RESTRI
 WSP_GGML_API void wsp_dewsp_quantize_row_q8_0(const block_q8_0 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
 //WSP_GGML_API void wsp_dewsp_quantize_row_q8_1(const block_q8_1 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
 
+WSP_GGML_API void wsp_dewsp_quantize_row_mxfp4(const block_mxfp4 * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
+
 WSP_GGML_API void wsp_dewsp_quantize_row_q2_K(const block_q2_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
 WSP_GGML_API void wsp_dewsp_quantize_row_q3_K(const block_q3_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
 WSP_GGML_API void wsp_dewsp_quantize_row_q4_K(const block_q4_K * WSP_GGML_RESTRICT x, float * WSP_GGML_RESTRICT y, int64_t k);
@@ -90,6 +94,8 @@ WSP_GGML_API size_t wsp_quantize_q5_0(const float * WSP_GGML_RESTRICT src, void
 WSP_GGML_API size_t wsp_quantize_q5_1(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 WSP_GGML_API size_t wsp_quantize_q8_0(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
+WSP_GGML_API size_t wsp_quantize_mxfp4(const float * WSP_GGML_RESTRICT src, void * WSP_GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
 WSP_GGML_API void wsp_iq2xs_init_impl(enum wsp_ggml_type type);
 WSP_GGML_API void wsp_iq2xs_free_impl(enum wsp_ggml_type type);
 WSP_GGML_API void wsp_iq3xs_init_impl(int grid_size);

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

@@ -241,6 +241,8 @@
 #define WSP_GGML_ROPE_TYPE_MROPE  8
 #define WSP_GGML_ROPE_TYPE_VISION 24
 
+#define WSP_GGML_MROPE_SECTIONS   4
+
 #define WSP_GGML_UNUSED(x) (void)(x)
 
 #define WSP_GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
@@ -304,6 +306,16 @@
     WSP_GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
     WSP_GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
+#define WSP_GGML_TENSOR_TERNARY_OP_LOCALS \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne2, src2, ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb2, src2, nb) \
+    WSP_GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    WSP_GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
 #define WSP_GGML_TENSOR_BINARY_OP_LOCALS01 \
     WSP_GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
     WSP_GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
@@ -314,6 +326,13 @@
 extern "C" {
 #endif
 
+    // Function type used in fatal error callbacks
+    typedef void (*wsp_ggml_abort_callback_t)(const char * error_message);
+
+    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+    // Returns the old callback for chaining
+    WSP_GGML_API wsp_ggml_abort_callback_t wsp_ggml_set_abort_callback(wsp_ggml_abort_callback_t callback);
+
     WSP_GGML_NORETURN WSP_GGML_ATTRIBUTE_FORMAT(3, 4)
     WSP_GGML_API void wsp_ggml_abort(const char * file, int line, const char * fmt, ...);
 
@@ -388,7 +407,8 @@ extern "C" {
         // WSP_GGML_TYPE_IQ4_NL_4_4 = 36,
         // WSP_GGML_TYPE_IQ4_NL_4_8 = 37,
         // WSP_GGML_TYPE_IQ4_NL_8_8 = 38,
-        WSP_GGML_TYPE_COUNT   = 39,
+        WSP_GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
+        WSP_GGML_TYPE_COUNT   = 40,
     };
 
     // precision
@@ -423,6 +443,7 @@ extern "C" {
         WSP_GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         WSP_GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         WSP_GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
+        WSP_GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
     };
 
     // available tensor operations:
@@ -431,6 +452,7 @@ extern "C" {
 
         WSP_GGML_OP_DUP,
         WSP_GGML_OP_ADD,
+        WSP_GGML_OP_ADD_ID,
         WSP_GGML_OP_ADD1,
         WSP_GGML_OP_ACC,
         WSP_GGML_OP_SUB,
@@ -488,7 +510,7 @@ extern "C" {
         WSP_GGML_OP_POOL_1D,
         WSP_GGML_OP_POOL_2D,
         WSP_GGML_OP_POOL_2D_BACK,
-        WSP_GGML_OP_UPSCALE, // nearest interpolate
+        WSP_GGML_OP_UPSCALE,
         WSP_GGML_OP_PAD,
         WSP_GGML_OP_PAD_REFLECT_1D,
         WSP_GGML_OP_ROLL,
@@ -520,6 +542,7 @@ extern "C" {
         WSP_GGML_OP_CROSS_ENTROPY_LOSS,
         WSP_GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         WSP_GGML_OP_OPT_STEP_ADAMW,
+        WSP_GGML_OP_OPT_STEP_SGD,
 
         WSP_GGML_OP_GLU,
 
@@ -550,6 +573,9 @@ extern "C" {
         WSP_GGML_GLU_OP_REGLU,
         WSP_GGML_GLU_OP_GEGLU,
         WSP_GGML_GLU_OP_SWIGLU,
+        WSP_GGML_GLU_OP_SWIGLU_OAI,
+        WSP_GGML_GLU_OP_GEGLU_ERF,
+        WSP_GGML_GLU_OP_GEGLU_QUICK,
 
         WSP_GGML_GLU_OP_COUNT,
     };
@@ -639,6 +665,9 @@ extern "C" {
 
     // misc
 
+    WSP_GGML_API const char * wsp_ggml_version(void);
+    WSP_GGML_API const char * wsp_ggml_commit(void);
+
     WSP_GGML_API void    wsp_ggml_time_init(void); // call this once at the beginning of the program
     WSP_GGML_API int64_t wsp_ggml_time_ms(void);
     WSP_GGML_API int64_t wsp_ggml_time_us(void);
@@ -819,6 +848,13 @@ extern "C" {
             struct wsp_ggml_tensor  * b,
             enum   wsp_ggml_type      type);
 
+    // dst[i0, i1, i2] = a[i0, i1, i2] + b[i0, ids[i1, i2]]
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_add_id(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * ids);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_add1(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -1137,6 +1173,22 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_erf(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_erf_swapped(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_quick(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_quick_swapped(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a);
+
     // A: n columns, r rows,
     // B: n columns, r rows,
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_glu_split(
@@ -1160,6 +1212,23 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             struct wsp_ggml_tensor  * b);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_erf_split(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_geglu_quick_split(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_swiglu_oai(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            float                 alpha,
+            float                 limit);
+
     // normalize along rows
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_norm(
             struct wsp_ggml_context * ctx,
@@ -1259,6 +1328,19 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             float                 s);
 
+    // x = s * a + b
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_scale_bias(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_scale_bias_inplace(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
     // b -> view(a,offset,nb1,nb2,3), return modified a
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set(
             struct wsp_ggml_context * ctx,
@@ -1503,8 +1585,14 @@ extern "C" {
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a);
 
+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
     // max_bias = 0.0f for no ALiBi
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_ext(
             struct wsp_ggml_context * ctx,
@@ -1513,6 +1601,10 @@ extern "C" {
             float                 scale,
             float                 max_bias);
 
+    WSP_GGML_API void wsp_ggml_soft_max_add_sinks(
+            struct wsp_ggml_tensor * a,
+            struct wsp_ggml_tensor * sinks);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_soft_max_ext_back(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -1571,7 +1663,7 @@ extern "C" {
             struct wsp_ggml_tensor  * b,
             struct wsp_ggml_tensor  * c,
             int                   n_dims,
-            int                   sections[4],
+            int                   sections[WSP_GGML_MROPE_SECTIONS],
             int                   mode,
             int                   n_ctx_orig,
             float                 freq_base,
@@ -1597,6 +1689,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_multi_inplace(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            struct wsp_ggml_tensor  * b,
+            struct wsp_ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[WSP_GGML_MROPE_SECTIONS],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     WSP_GGML_DEPRECATED(WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_rope_custom(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -1967,11 +2075,17 @@ extern "C" {
 
 #define WSP_GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    1]
-    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
-    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = WSP_GGML_PAD(n_batch, WSP_GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = WSP_GGML_PAD(n_batch, WSP_GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_attn_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * q,
@@ -1989,6 +2103,10 @@ extern "C" {
     WSP_GGML_API enum wsp_ggml_prec wsp_ggml_flash_attn_ext_get_prec(
             const struct wsp_ggml_tensor * a);
 
+    WSP_GGML_API void wsp_ggml_flash_attn_ext_add_sinks(
+            struct wsp_ggml_tensor * a,
+            struct wsp_ggml_tensor * sinks);
+
     // TODO: needs to be adapted to wsp_ggml_flash_attn_ext
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_flash_attn_back(
            struct wsp_ggml_context * ctx,
@@ -2010,7 +2128,8 @@ extern "C" {
             struct wsp_ggml_tensor  * dt,
             struct wsp_ggml_tensor  * A,
             struct wsp_ggml_tensor  * B,
-            struct wsp_ggml_tensor  * C);
+            struct wsp_ggml_tensor  * C,
+            struct wsp_ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -2193,7 +2312,14 @@ extern "C" {
             struct wsp_ggml_tensor  * grad,
             struct wsp_ggml_tensor  * m,
             struct wsp_ggml_tensor  * v,
-            struct wsp_ggml_tensor  * adamw_params); // parameters such a the learning rate
+            struct wsp_ggml_tensor  * adamw_params); // parameters such as the learning rate
+
+    // stochastic gradient descent step (with weight decay)
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_opt_step_sgd(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor *  a,
+        struct wsp_ggml_tensor *  grad,
+        struct wsp_ggml_tensor *  sgd_params); // alpha, weight decay
 
     //
     // automatic differentiation

package/ios/rnwhisper.xcframework/tvos-arm64_x86_64-simulator/rnwhisper.framework/Headers/ggml-common.h

@@ -99,6 +99,9 @@ typedef sycl::half2 wsp_ggml_half2;
 #define QI4_1 (QK4_1 / (4 * QR4_1))
 #define QR4_1 2
 
+#define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
+#define QR_MXFP4 2
+
 #define QI5_0 (QK5_0 / (4 * QR5_0))
 #define QR5_0 2
 
@@ -184,6 +187,13 @@ typedef struct {
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(wsp_ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
 
+#define QK_MXFP4 32
+typedef struct {
+    uint8_t e; // E8M0
+    uint8_t qs[QK_MXFP4/2];
+} block_mxfp4;
+static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
+
 #define QK5_0 32
 typedef struct {
     wsp_ggml_half d;           // delta
@@ -1074,10 +1084,17 @@ WSP_GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 WSP_GGML_TABLE_END()
 
+// TODO: fix name to kvalues_iq4_nl
 WSP_GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
     -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
 WSP_GGML_TABLE_END()
 
+// e2m1 values (doubled)
+// ref: https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
+WSP_GGML_TABLE_BEGIN(int8_t, kvalues_mxfp4, 16)
+    0, 1, 2, 3, 4, 6, 8, 12, 0, -1, -2, -3, -4, -6, -8, -12,
+WSP_GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

package/ios/rnwhisper.xcframework/tvos-arm64_x86_64-simulator/rnwhisper.framework/Headers/ggml-impl.h

@@ -73,6 +73,22 @@ static inline int wsp_ggml_up(int n, int m) {
     return (n + m - 1) & ~(m - 1);
 }
 
+// TODO: move to ggml.h?
+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;
+    }
+    for (int i = 0; i < WSP_GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
 //
 // logging
 //
@@ -394,6 +410,67 @@ static inline wsp_ggml_fp16_t wsp_ggml_compute_fp32_to_fp16(float f) {
 #define WSP_GGML_FP16_TO_FP32(x) WSP_GGML_COMPUTE_FP16_TO_FP32(x)
 #define WSP_GGML_FP32_TO_FP16(x) WSP_GGML_COMPUTE_FP32_TO_FP16(x)
 
+static inline float wsp_ggml_e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;  // Stores the raw bit representation of the float
+
+    // Handle special case for minimum exponent (denormalized float)
+    if (x == 0) {
+        // Bit pattern for 2^(-127):
+        // - Sign bit: 0 (positive)
+        // - Exponent: 0 (denormalized number)
+        // - Mantissa: 0x400000 (0.5 in fractional form)
+        // Value = 0.5 * 2^(-126) = 2^(-127)
+        bits = 0x00400000;
+    }
+    // note: disabled as we don't need to handle NaNs
+    //// Handle special case for NaN (all bits set)
+    //else if (x == 0xFF) {
+    //    // Standard quiet NaN pattern:
+    //    // - Sign bit: 0
+    //    // - Exponent: all 1s (0xFF)
+    //    // - Mantissa: 0x400000 (quiet NaN flag)
+    //    bits = 0x7FC00000;
+    //}
+    // Normalized values (most common case)
+    else {
+        // Construct normalized float by shifting exponent into position:
+        // - Exponent field: 8 bits (positions 30-23)
+        // - Mantissa: 0 (implicit leading 1)
+        // Value = 2^(x - 127)
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;  // Final float value
+                   // Safely reinterpret bit pattern as float without type-punning issues
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+// Equal to wsp_ggml_e8m0_to_fp32/2
+// Useful with MXFP4 quantization since the E0M2 values are doubled
+static inline float wsp_ggml_e8m0_to_fp32_half(uint8_t x) {
+    uint32_t bits;
+
+    // For x < 2: use precomputed denormal patterns
+    if (x < 2) {
+        // 0x00200000 = 2^(-128), 0x00400000 = 2^(-127)
+        bits = 0x00200000 << x;
+    }
+    // For x >= 2: normalized exponent adjustment
+    else {
+        // 0.5 * 2^(x-127) = 2^(x-128) = normalized with exponent (x-1)
+        bits = (uint32_t)(x - 1) << 23;
+    }
+    // Note: NaNs are not handled here
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
+#define WSP_GGML_E8M0_TO_FP32(x) wsp_ggml_e8m0_to_fp32(x)
+#define WSP_GGML_E8M0_TO_FP32_HALF(x) wsp_ggml_e8m0_to_fp32_half(x)
+
 /**
  * Converts brain16 to float32.
  *