whisper.rn 0.3.0-rc.1 → 0.3.0-rc.3

package/cpp/ggml.h

@@ -190,9 +190,12 @@
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
 #define GGML_FILE_VERSION 1
 
+#define GGML_QNT_VERSION        2    // bump this on quantization format changes
+#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
+
 #define GGML_MAX_DIMS          4
 #define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        16
+#define GGML_MAX_PARAMS        256
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_OPT           4
 #define GGML_DEFAULT_N_THREADS 4
@@ -231,7 +234,7 @@ extern "C" {
         GGML_TYPE_F16  = 1,
         GGML_TYPE_Q4_0 = 2,
         GGML_TYPE_Q4_1 = 3,
-        GGML_TYPE_Q4_2 = 4,
+        // GGML_TYPE_Q4_2 = 4, support has been removed
         // GGML_TYPE_Q4_3 (5) support has been removed
         GGML_TYPE_Q5_0 = 6,
         GGML_TYPE_Q5_1 = 7,
@@ -243,6 +246,11 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
+    enum ggml_backend {
+        GGML_BACKEND_CPU = 0,
+        GGML_BACKEND_CUDA = 1,
+    };
+
     // model file types
     enum ggml_ftype {
         GGML_FTYPE_UNKNOWN     = -1,
@@ -251,7 +259,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        GGML_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
@@ -263,12 +270,16 @@ extern "C" {
 
         GGML_OP_DUP,
         GGML_OP_ADD,
+        GGML_OP_ADD1,
+        GGML_OP_ACC,
         GGML_OP_SUB,
         GGML_OP_MUL,
         GGML_OP_DIV,
         GGML_OP_SQR,
         GGML_OP_SQRT,
+        GGML_OP_LOG,
         GGML_OP_SUM,
+        GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
         GGML_OP_ABS,
@@ -278,12 +289,15 @@ extern "C" {
         GGML_OP_RELU,
         GGML_OP_GELU,
         GGML_OP_SILU,
+        GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
+        GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
 
         GGML_OP_SCALE,
+        GGML_OP_SET,
         GGML_OP_CPY,
         GGML_OP_CONT,
         GGML_OP_RESHAPE,
@@ -291,10 +305,15 @@ extern "C" {
         GGML_OP_PERMUTE,
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
+        GGML_OP_GET_ROWS_BACK,
+        GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
+        GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
         GGML_OP_ROPE,
+        GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
+        GGML_OP_CLAMP,
         GGML_OP_CONV_1D_1S,
         GGML_OP_CONV_1D_2S,
 
@@ -322,7 +341,8 @@ extern "C" {
 
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type type;
+        enum ggml_type    type;
+        enum ggml_backend backend;
 
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
@@ -353,7 +373,7 @@ extern "C" {
 
         char name[32];
 
-        char padding[8]; // TODO: remove and add padding to name?
+        char padding[16];
     };
 
     // computation graph
@@ -497,6 +517,29 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_acc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_acc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
     GGML_API struct ggml_tensor * ggml_sub(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -520,12 +563,24 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_log(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_log_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return scalar
-    // TODO: compute sum along rows
     GGML_API struct ggml_tensor * ggml_sum(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
+    GGML_API struct ggml_tensor * ggml_sum_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // mean along rows
     GGML_API struct ggml_tensor * ggml_mean(
             struct ggml_context * ctx,
@@ -567,6 +622,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_silu_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     // TODO: eps is hardcoded to 1e-5 for now
     GGML_API struct ggml_tensor * ggml_norm(
@@ -577,6 +639,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_rms_norm_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // A: m rows, n columns
     // B: p rows, n columns (i.e. we transpose it internally)
     // result is m columns, p rows
@@ -589,12 +658,66 @@ extern "C" {
     // operations on tensors without backpropagation
     //
 
-    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_scale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_scale_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_2d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
             struct ggml_context * ctx,
@@ -615,6 +738,11 @@ extern "C" {
 
     // return view(a)
     // TODO: when we start computing gradient, make a copy instead of view
+    GGML_API struct ggml_tensor * ggml_reshape_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
     GGML_API struct ggml_tensor * ggml_reshape_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -630,6 +758,14 @@ extern "C" {
             int64_t               ne1,
             int64_t               ne2);
 
+    GGML_API struct ggml_tensor * ggml_reshape_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
     // offset in bytes
     GGML_API struct ggml_tensor * ggml_view_1d(
             struct ggml_context * ctx,
@@ -655,6 +791,18 @@ extern "C" {
             size_t                nb2, // slice stride in bytes
             size_t                offset);
 
+    GGML_API struct ggml_tensor * ggml_view_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            size_t                nb1, // row   stride in bytes
+            size_t                nb2, // slice stride in bytes
+            size_t                nb3,
+            size_t                offset);
+
     GGML_API struct ggml_tensor * ggml_permute(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -673,20 +821,50 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_get_rows_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c);
+
+    GGML_API struct ggml_tensor * ggml_diag(
+        struct ggml_context     * ctx,
+        struct ggml_tensor      * a);
+
     // set elements above the diagonal to -INF
-    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_diag_mask_inf(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past);
 
     // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // set elements above the diagonal to 0
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
     GGML_API struct ggml_tensor * ggml_soft_max(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // rotary position embedding
     // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
     // TODO: avoid creating a new tensor every time
@@ -697,13 +875,39 @@ extern "C" {
             int                   n_dims,
             int                   mode);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode);
+
+    // rotary position embedding backward, i.e compute dx from dy
+    // a - dy
+    GGML_API struct ggml_tensor * ggml_rope_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode);
+
     // alibi position embedding
     // in-place, returns view(a)
     struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
-            int                   n_head);
+            int                   n_head,
+            float                 bias_max);
+
+    // clamp
+    // in-place, returns view(a)
+    struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
 
     // padding = 1
     // TODO: we don't support extra parameters for now
@@ -741,13 +945,13 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-            const  ggml_unary_op_f32_t fun);
+                   ggml_unary_op_f32_t   fun);
 
     GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-            const  ggml_binary_op_f32_t fun);
+                   ggml_binary_op_f32_t   fun);
 
     //
     // automatic differentiation
@@ -876,7 +1080,6 @@ extern "C" {
 
     GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);

package/cpp/whisper.cpp

@@ -139,7 +139,7 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
     { "hi",  { 17,  "hindi",          } },
     { "fi",  { 18,  "finnish",        } },
     { "vi",  { 19,  "vietnamese",     } },
-    { "iw",  { 20,  "hebrew",         } },
+    { "he",  { 20,  "hebrew",         } },
     { "uk",  { 21,  "ukrainian",      } },
     { "el",  { 22,  "greek",          } },
     { "ms",  { 23,  "malay",          } },
@@ -291,15 +291,6 @@ static const std::map<ggml_type, std::map<e_model, size_t>> MEM_REQ_MODEL = {
             { MODEL_LARGE,  1124ull*MB },
         },
     },
-    { GGML_TYPE_Q4_2,
-        {
-            { MODEL_TINY,     26ull*MB },
-            { MODEL_BASE,     50ull*MB },
-            { MODEL_SMALL,   154ull*MB },
-            { MODEL_MEDIUM,  470ull*MB },
-            { MODEL_LARGE,   940ull*MB },
-        },
-    },
     { GGML_TYPE_Q5_0,
         {
             { MODEL_TINY,     30ull*MB },
@@ -861,6 +852,10 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             model.type = e_model::MODEL_LARGE;
         }
 
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+
         // for the big tensors, we have the option to store the data in 16-bit floats or quantized
         // in order to save memory and also to speed up the computation
         wctx.wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
@@ -882,6 +877,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         fprintf(stderr, "%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
         fprintf(stderr, "%s: n_mels        = %d\n", __func__, hparams.n_mels);
         fprintf(stderr, "%s: ftype         = %d\n", __func__, model.hparams.ftype);
+        fprintf(stderr, "%s: qntvr         = %d\n", __func__, qntvr);
         fprintf(stderr, "%s: type          = %d\n", __func__, model.type);
 
         // print memory requirements
@@ -1106,7 +1102,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_1_b
         }
 
-        ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*256; // object overhead
+        ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*512; // object overhead
 
         fprintf(stderr, "%s: model ctx     = %7.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
     }
@@ -1554,14 +1550,14 @@ static bool whisper_encode_internal(
                             Qcur),
                         Qcur);
 
-                //Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+                //Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
                 // note: no bias for Key
                 struct ggml_tensor * Kcur = ggml_mul_mat(ctx0,
                         layer.attn_k_w,
                         cur);
 
-                //Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+                //Kcur = ggml_scale_inplace(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
                         layer.attn_v_w,
@@ -1621,12 +1617,12 @@ static bool whisper_encode_internal(
                 struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
                 struct ggml_tensor * KQ_scaled =
-                    ggml_scale(ctx0,
+                    ggml_scale_inplace(ctx0,
                             KQ,
                             ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
                             );
 
-                struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_scaled);
+                struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_scaled);
 
                 struct ggml_tensor * V =
                     ggml_cpy(ctx0,
@@ -1809,7 +1805,7 @@ static bool whisper_encode_internal(
                 layer.cross_attn_k_w,
                 cur);
 
-            Kcross = ggml_scale(ctx0, Kcross, ggml_new_f32(ctx0, pow(float(n_state) / n_head, -0.25)));
+            Kcross = ggml_scale_inplace(ctx0, Kcross, ggml_new_f32(ctx0, pow(float(n_state) / n_head, -0.25)));
 
             wstate.use_buf(ctx0, 1);
 
@@ -1956,14 +1952,14 @@ static bool whisper_decode_internal(
                         Qcur),
                     Qcur);
 
-            Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
             // note: no bias for Key
             struct ggml_tensor * Kcur = ggml_mul_mat(ctx0,
                     layer.attn_k_w,
                     cur);
 
-            Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Kcur = ggml_scale_inplace(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
             // store key and value to memory
             {
@@ -2012,14 +2008,14 @@ static bool whisper_decode_internal(
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
             //struct ggml_tensor * KQ_scaled =
-            //    ggml_scale(ctx0,
+            //    ggml_scale_inplace(ctx0,
             //            KQ,
             //            ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
             //            );
 
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ, n_past);
 
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
 
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
@@ -2083,7 +2079,7 @@ static bool whisper_decode_internal(
                         Qcur),
                     Qcur);
 
-            Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
 
             // Kcross is already scaled
             struct ggml_tensor * Kcross =
@@ -2123,15 +2119,15 @@ static bool whisper_decode_internal(
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
             //struct ggml_tensor * KQ_scaled =
-            //    ggml_scale(ctx0,
+            //    ggml_scale_inplace(ctx0,
             //            KQ,
             //            ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
             //            );
 
             // no masking for cross-attention
-            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
 
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ);
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
@@ -2602,6 +2598,15 @@ static std::string whisper_get_coreml_path_encoder(std::string path_bin) {
         path_bin = path_bin.substr(0, pos);
     }
 
+    // match "-qx_x"
+    pos = path_bin.rfind('-');
+    if (pos != std::string::npos) {
+        auto sub = path_bin.substr(pos);
+        if (sub.size() == 5 && sub[1] == 'q' && sub[3] == '_') {
+            path_bin = path_bin.substr(0, pos);
+        }
+    }
+
     path_bin += "-encoder.mlmodelc";
 
     return path_bin;
@@ -2847,6 +2852,12 @@ void whisper_free(struct whisper_context * ctx) {
     }
 }
 
+void whisper_free_params(struct whisper_full_params * params) {
+    if (params) {
+        delete params;
+    }
+}
+
 int whisper_pcm_to_mel_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
     if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, WHISPER_N_FFT, WHISPER_HOP_LENGTH, WHISPER_N_MEL, n_threads, ctx->model.filters, false, state->mel)) {
         fprintf(stderr, "%s: failed to compute mel spectrogram\n", __func__);
@@ -3280,6 +3291,14 @@ const char * whisper_print_system_info(void) {
 
 ////////////////////////////////////////////////////////////////////////////
 
+struct whisper_full_params * whisper_full_default_params_by_ref(enum whisper_sampling_strategy strategy) {
+    struct whisper_full_params params = whisper_full_default_params(strategy);
+
+    struct whisper_full_params* result = new whisper_full_params();
+    *result = params;
+    return result;
+}
+
 struct whisper_full_params whisper_full_default_params(enum whisper_sampling_strategy strategy) {
     struct whisper_full_params result = {
         /*.strategy         =*/ strategy,
@@ -4903,7 +4922,7 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
     // b: N*N*sizeof(float)
     // c: N*N*sizeof(float)
     // when F16 is used, there is an extra work buffer of size N*N*sizeof(float)
-    std::vector<char> buf(4llu*N_max*N_max*sizeof(float) + 4*256);
+    std::vector<char> buf(4llu*N_max*N_max*sizeof(float) + 4*512);
 
     // put a bunch of random data in the buffer
     for (size_t i = 0; i < buf.size(); i++) buf[i] = i;
@@ -4911,7 +4930,6 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
     for (int j = 0; j < (int) sizes.size(); j++) {
         int n_q4_0 = 0;
         int n_q4_1 = 0;
-        int n_q4_2 = 0;
         int n_q5_0 = 0;
         int n_q5_1 = 0;
         int n_q8_0 = 0;
@@ -4921,7 +4939,6 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
         // GFLOPS/s
         double s_q4_0 = 0.0;
         double s_q4_1 = 0.0;
-        double s_q4_2 = 0.0;
         double s_q5_0 = 0.0;
         double s_q5_1 = 0.0;
         double s_q8_0 = 0.0;
@@ -4930,18 +4947,17 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
 
         const size_t N = sizes[j];
 
-        for (int k = 0; k < 8; ++k) {
+        for (int k = 0; k < 7; ++k) {
             const ggml_type wtype =
                 k == 0 ? GGML_TYPE_Q4_0 :
                 k == 1 ? GGML_TYPE_Q4_1 :
-                k == 2 ? GGML_TYPE_Q4_2 :
-                k == 3 ? GGML_TYPE_Q5_0 :
-                k == 4 ? GGML_TYPE_Q5_1 :
-                k == 5 ? GGML_TYPE_Q8_0 :
-                k == 6 ? GGML_TYPE_F16  : GGML_TYPE_F32;
+                k == 2 ? GGML_TYPE_Q5_0 :
+                k == 3 ? GGML_TYPE_Q5_1 :
+                k == 4 ? GGML_TYPE_Q8_0 :
+                k == 5 ? GGML_TYPE_F16  : GGML_TYPE_F32;
 
-            double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_q4_2 : k == 3 ? s_q5_0 : k == 4 ? s_q5_1 : k == 5 ? s_q8_0 : k == 6 ? s_fp16 : /*k == 7*/ s_fp32;
-            int    & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_q4_2 : k == 3 ? n_q5_0 : k == 4 ? n_q5_1 : k == 5 ? n_q8_0 : k == 6 ? n_fp16 : /*k == 7*/ n_fp32;
+            double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_q5_0 : k == 3 ? s_q5_1 : k == 4 ? s_q8_0 : k == 5 ? s_fp16 : /*k == 6*/ s_fp32;
+            int    & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_q5_0 : k == 3 ? n_q5_1 : k == 4 ? n_q8_0 : k == 5 ? n_fp16 : /*k == 6*/ n_fp32;
 
             struct ggml_init_params gparams = {
                 /*.mem_size   =*/ buf.size(),
@@ -4985,9 +5001,9 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
             s = ((2.0*N*N*N*n)/tsum)*1e-9;
         }
 
-        // Q4_0 | Q4_1 | Q4_2
-        snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) | Q4_1 %7.1f GFLOPS (%3d runs) | Q4_2 %7.1f GFLOPS (%3d runs)\n",
-                N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1, s_q4_2, n_q4_2);
+        // Q4_0 | Q4_1
+        snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) | Q4_1 %7.1f GFLOPS (%3d runs)\n",
+                N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1);
         s += strbuf;
 
         // Q5_0 | Q5_1 | Q8_0