whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-cpu/ops.cpp

@@ -664,6 +664,8 @@ void ggml_compute_forward_add(
             {
                 ggml_compute_forward_add_non_quantized(params, dst);
             } break;
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1113,6 +1115,8 @@ void ggml_compute_forward_add1(
                     GGML_ABORT("fatal error");
                 }
             } break;
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1242,6 +1246,8 @@ void ggml_compute_forward_acc(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1910,7 +1916,11 @@ static void ggml_compute_forward_concat_any(
     GGML_ASSERT(dim >= 0 && dim < 4);
 
     int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = src0->ne[dim];
+    if (dim == 0) {
+        o[dim] = src0->ne[dim]/ggml_blck_size(src0->type);
+    } else {
+        o[dim] = src0->ne[dim];
+    }
 
     const char * x;
 
@@ -1918,8 +1928,8 @@ static void ggml_compute_forward_concat_any(
     for (int i3 = 0; i3 < ne3; i3++) {
         for (int i2 = ith; i2 < ne2; i2 += nth) {
             for (int i1 = 0; i1 < ne1; i1++) {
-                for (int i0 = 0; i0 < ne0; i0++) {
-                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                for (int i0 = 0; i0 < ne0/ggml_blck_size(dst->type); i0++) {
+                    if (i0 < ne00/ggml_blck_size(src0->type) && i1 < ne01 && i2 < ne02 && i3 < ne03) {
                         x = (const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03;
                     } else {
                         x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13;
@@ -2068,6 +2078,14 @@ void ggml_compute_forward_concat(
     ggml_tensor * dst) {
 
     const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    if (ggml_is_quantized(src0->type)) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+        GGML_ASSERT(ggml_is_contiguous(src1));
+        GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
+        GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
+    }
 
     switch (src0->type) {
         case GGML_TYPE_F16:
@@ -2232,8 +2250,42 @@ static void ggml_compute_forward_fill_f32(const ggml_compute_params * params, gg
     }
 }
 
+static void ggml_compute_forward_fill_f16(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_fp16_t c = GGML_CPU_FP32_TO_FP16(ggml_get_op_params_f32(dst, 0));
+
+    GGML_TENSOR_LOCALS(int64_t, ne, dst, ne);
+    GGML_TENSOR_LOCALS(size_t,  nb, dst, nb);
+
+    const auto [ir0, ir1] = get_thread_range(params, dst);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne2*ne1);
+        const int64_t i02 = (ir - i03*ne2*ne1)/ne1;
+        const int64_t i01 = (ir - i03*ne2*ne1 - i02*ne1);
+
+        ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1);
+
+        ggml_vec_set_f16(ne0, dst_ptr, c);
+    }
+}
+
 void ggml_compute_forward_fill(const ggml_compute_params * params, ggml_tensor * dst) {
-    ggml_compute_forward_fill_f32(params, dst);
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_fill_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_fill_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("unsupported type for ggml_compute_forward_fill: %s", ggml_type_name(src0->type));
+            }
+    }
 }
 
 // ggml_compute_tri
@@ -3651,8 +3703,6 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -3666,25 +3716,49 @@ static void ggml_compute_forward_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
 
-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
-                float mean = sum/ne00;
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    const float * xf = (const float *) x;
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
+                    float sum = 0.0;
+                    ggml_vec_sum_f32(ne00, &sum, xf);
+                    float mean = sum/ne00;
+
+                    float * yf = (float *) y;
+                    float variance = 0;
 
 #ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
+                    mean = -mean;
+                    vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
+                    vDSP_measqv(yf, 1, &variance, ne00);
 #else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+                    variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
 #endif //GGML_USE_ACCELERATE
 
-                const float scale = 1.0f/sqrtf(variance + eps);
-                ggml_vec_scale_f32(ne00, y, scale);
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    ggml_vec_scale_f32(ne00, yf, scale);
+                } else {
+                    float sum = 0.0;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        sum += *(const float *) (x + i00*nb00);
+                    }
+                    const float mean = sum/ne00;
+
+                    float variance = 0.0f;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float v = *(const float *) (x + i00*nb00) - mean;
+                        *(float *) (y + i00*nb0) = v;
+                        variance += v * v;
+                    }
+                    variance /= ne00;
+
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        *(float *) (y + i00*nb0) *= scale;
+                    }
+                }
             }
         }
     }
@@ -3710,11 +3784,27 @@ void ggml_compute_forward_norm(
 
 // ggml_compute_forward_group_rms_norm
 
+// fusion kinds that can be combined with the rms_norm computation in a single pass.
+// extend this enum when adding new fused variants (e.g. FUSE_ADD, FUSE_MUL_ADD, ...).
+enum ggml_rms_norm_fuse_op {
+    GGML_RMS_NORM_FUSE_OP_NONE,
+    GGML_RMS_NORM_FUSE_OP_MUL,
+};
+
+template <ggml_rms_norm_fuse_op FUSE_OP>
 static void ggml_compute_forward_rms_norm_f32(
         const ggml_compute_params * params,
-        ggml_tensor * dst) {
+        ggml_tensor * dst_rms_norm,
+        ggml_tensor * dst_fused = nullptr) {
 
-    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src0 = dst_rms_norm->src[0];
+    const ggml_tensor * src1 = nullptr;
+    ggml_tensor       * dst  = dst_rms_norm;
+
+    if constexpr (FUSE_OP == GGML_RMS_NORM_FUSE_OP_MUL) {
+        src1 = (dst_fused->src[0] == dst_rms_norm) ? dst_fused->src[1] : dst_fused->src[0];
+        dst  = dst_fused;
+    }
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
@@ -3723,11 +3813,10 @@ static void ggml_compute_forward_rms_norm_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    GGML_TENSOR_UNARY_OP_LOCALS
+    GGML_TENSOR_BINARY_OP_LOCALS
 
     float eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
-
+    memcpy(&eps, dst_rms_norm->op_params, sizeof(float));
     GGML_ASSERT(eps >= 0.0f);
 
     // TODO: optimize
@@ -3737,25 +3826,32 @@ static void ggml_compute_forward_rms_norm_f32(
                 const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
                 ggml_float sum = 0.0;
+                // worth switching to explicit SIMD?
                 for (int64_t i00 = 0; i00 < ne00; i00++) {
                     sum += (ggml_float)(x[i00] * x[i00]);
                 }
 
-                const float mean = sum/ne00;
-
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-                // for (int i00 = 0; i00 < ne00; i00++) {
-                //     y[i00] = x[i00];
-                // }
-
+                const float mean  = sum/ne00;
                 const float scale = 1.0f/sqrtf(mean + eps);
 
                 // if you hit this, likely you got an inf somewhere earlier
                 assert(scale > 0.0f);
 
-                ggml_vec_scale_f32(ne00, y, scale);
+                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
+
+                if constexpr (FUSE_OP == GGML_RMS_NORM_FUSE_OP_MUL) {
+                    const int64_t i11 = i01 % ne11;
+                    const int64_t i12 = i02 % ne12;
+                    const int64_t i13 = i03 % ne13;
+                    const float * w = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13);
+
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        y[i00] = x[i00] * scale * w[i00];
+                    }
+                } else {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, y, scale);
+                }
             }
         }
     }
@@ -3770,7 +3866,31 @@ void ggml_compute_forward_rms_norm(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rms_norm_f32(params, dst);
+                ggml_compute_forward_rms_norm_f32<GGML_RMS_NORM_FUSE_OP_NONE>(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// Fused RMS_NORM + MUL: computes dst = rms_norm(src0) * src1 in a single pass.
+// This avoids materializing the intermediate rms_norm result in memory.
+void ggml_compute_forward_rms_norm_mul_fused(
+        const ggml_compute_params * params,
+        ggml_tensor * dst_rms_norm,
+        ggml_tensor * dst_mul) {
+
+    GGML_ASSERT(dst_mul != nullptr);
+    GGML_ASSERT(dst_mul->src[0] == dst_rms_norm || dst_mul->src[1] == dst_rms_norm);
+
+    const ggml_tensor * src0 = dst_rms_norm->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_rms_norm_f32<GGML_RMS_NORM_FUSE_OP_MUL>(params, dst_rms_norm, dst_mul);
             } break;
         default:
             {
@@ -3925,12 +4045,12 @@ static void ggml_compute_forward_rms_norm_back_f32(
                 // dx := scale(dx, rrms)
                 float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
 
-                // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps)
-                ggml_vec_cpy_f32  (ne00, dx, x);
-                // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps);
-                ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps);
-                ggml_vec_acc_f32  (ne00, dx, dz);
-                ggml_vec_scale_f32(ne00, dx, rrms);
+                // dx[i00] = (dz + x*(-sum_xdz/sum_eps)) * rrms
+                // note: https://github.com/ggml-org/ggml/issues/1491
+                const float scale_x = (float) (-sum_xdz) / sum_eps;
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    dx[i00] = (dz[i00] + x[i00] * scale_x) * rrms;
+                }
             }
         }
     }
@@ -4059,8 +4179,6 @@ static void ggml_compute_forward_l2_norm_f32(
 
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -4075,20 +4193,27 @@ static void ggml_compute_forward_l2_norm_f32(
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
 
                 ggml_float sum = 0.0;
                 for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)(x[i00] * x[i00]);
+                    const float xi = *(const float *) (x + i00*nb00);
+                    sum += (ggml_float)(xi * xi);
                 }
 
-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-
                 const float scale = 1.0f/fmaxf(sqrtf(sum), eps);
 
-                ggml_vec_scale_f32(ne00, y, scale);
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
+
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, (float *) y, scale);
+                } else {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float xi = *(const float *) (x + i00*nb00);
+                        *(float *) (y + i00*nb0) = xi * scale;
+                    }
+                }
             }
         }
     }
@@ -4331,6 +4456,8 @@ void ggml_compute_forward_out_prod(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4606,6 +4733,8 @@ void ggml_compute_forward_set(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4829,6 +4958,8 @@ void ggml_compute_forward_get_rows(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4894,8 +5025,8 @@ void ggml_compute_forward_get_rows(
     //}
 }
 
-template<typename idx_t>
-static void ggml_compute_forward_set_rows_f32(
+template<typename src_t, typename idx_t>
+static void ggml_compute_forward_set_rows_impl(
         const ggml_compute_params * params,
               ggml_tensor * dst) {
 
@@ -4910,7 +5041,7 @@ static void ggml_compute_forward_set_rows_f32(
     assert(ne0  == nc);
     assert(ne2  == ne02);
     assert(ne3  == ne03);
-    assert(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
     assert(ne02 % ne11 == 0);
     assert(ne03 % ne12 == 0);
 
@@ -4924,6 +5055,8 @@ static void ggml_compute_forward_set_rows_f32(
     const int64_t ir0 = dr*ith;
     const int64_t ir1 = std::min(ir0 + dr, nr);
 
+    const size_t rs = ggml_row_size(src0->type, nc);
+
     ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
 
     for (int64_t i03 = 0; i03 < ne03; ++i03) {
@@ -4937,9 +5070,18 @@ static void ggml_compute_forward_set_rows_f32(
 
                 GGML_ASSERT(i1 >= 0 && i1 < ne1);
 
-                from_float(
-                        (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
-                                        ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
+                if constexpr (std::is_same_v<src_t, float>) {
+                    from_float(
+                            (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
+                                            ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
+                } else if constexpr (std::is_same_v<src_t, ggml_fp16_t>) {
+                    memcpy(
+                                            ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3),
+                                            ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
+                                            rs);
+                } else {
+                    GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
+                }
             }
         }
     }
@@ -4956,13 +5098,27 @@ void ggml_compute_forward_set_rows(
         case GGML_TYPE_F32:
             {
                 if (src1->type == GGML_TYPE_I64) {
-                    ggml_compute_forward_set_rows_f32<int64_t>(params, dst);
+                    ggml_compute_forward_set_rows_impl<float, int64_t>(params, dst);
                 } else if (src1->type == GGML_TYPE_I32) {
-                    ggml_compute_forward_set_rows_f32<int32_t>(params, dst);
+                    ggml_compute_forward_set_rows_impl<float, int32_t>(params, dst);
                 } else {
                     GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
                 }
             } break;
+        case GGML_TYPE_F16:
+            {
+                if (dst->type == GGML_TYPE_F16) {
+                    if (src1->type == GGML_TYPE_I64) {
+                        ggml_compute_forward_set_rows_impl<ggml_fp16_t, int64_t>(params, dst);
+                    } else if (src1->type == GGML_TYPE_I32) {
+                        ggml_compute_forward_set_rows_impl<ggml_fp16_t, int32_t>(params, dst);
+                    } else {
+                        GGML_ABORT("src1->type = %d (%s) not supported", src1->type, ggml_type_name(src1->type));
+                    }
+                } else {
+                    GGML_ABORT("dst->type = %d (%s) not supported with src0->type = %d (%s)", dst->type, ggml_type_name(dst->type), src0->type, ggml_type_name(src0->type));
+                }
+            } break;
         default:
             {
                 GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
@@ -5554,6 +5710,8 @@ void ggml_compute_forward_clamp(
                 ggml_compute_forward_clamp_f16(params, dst);
             } break;
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -6643,6 +6801,78 @@ static inline int64_t ggml_wrap_around(int64_t coord, int64_t size) {
     return (coord  + size) % size; // adding size avoids negative number weirdness
 }
 
+// ggml_compute_forward_col2im_1d
+//
+// Scatter-add columns [K*OC, T_in] -> signal [T_out, OC]
+// where T_out = (T_in - 1)*s + K - 2*p.  Gather approach: each output reads ceil(K/s) inputs.
+// Parallelized over the time axis so the split stays balanced whatever OC is.
+// Supports F32, F16, BF16 input/output (same type), F32 accumulator.
+
+template <typename elem_t>
+static void ggml_compute_forward_col2im_1d_impl(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src = dst->src[0];  // [K*OC, T_in]
+
+    GGML_ASSERT(ggml_is_contiguous(src));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t OC = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+
+    const int64_t K_OC = src->ne[0];
+    const int64_t T_in = src->ne[1];
+    const int64_t K    = K_OC / OC;
+    const int64_t T_out = dst->ne[0];
+
+    const elem_t * col_data = (const elem_t *) src->data;
+    elem_t       * dst_data = (elem_t *) dst->data;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // Parallelize over the time axis: the split stays balanced whatever OC is,
+    // down to OC = 1 for mono audio, and threads read disjoint column bands
+    const int64_t dr = (T_out + nth - 1) / nth;
+    const int64_t it0 = dr * ith;
+    const int64_t it1 = it0 + dr < T_out ? it0 + dr : T_out;
+
+    for (int64_t oc = 0; oc < OC; oc++) {
+        for (int64_t t_out = it0; t_out < it1; t_out++) {
+            const int64_t t_abs = t_out + p0;  // absolute position in uncropped signal
+            // Gather: find all (t_in, k) where t_in * s + k == t_abs, 0 <= k < K
+            int64_t t_in_min = (t_abs - K + 1 + s0 - 1) / s0;  // ceil((t_abs-K+1)/s)
+            if (t_in_min < 0) t_in_min = 0;
+            int64_t t_in_max = t_abs / s0;
+            if (t_in_max >= T_in) t_in_max = T_in - 1;
+
+            float sum = 0.0f;
+            for (int64_t t_in = t_in_min; t_in <= t_in_max; t_in++) {
+                int64_t k = t_abs - t_in * s0;
+                if (k >= 0 && k < K) {
+                    // col layout: [K*OC, T_in], element (oc*K+k, t_in)
+                    sum += type_conversion_table<elem_t>::to_f32(col_data[(oc * K + k) + t_in * K_OC]);
+                }
+            }
+            // dst layout: [T_out, OC], element (t_out, oc)
+            dst_data[t_out + oc * T_out] = type_conversion_table<elem_t>::from_f32(sum);
+        }
+    }
+}
+
+void ggml_compute_forward_col2im_1d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    switch (dst->src[0]->type) {
+        case GGML_TYPE_F32:  ggml_compute_forward_col2im_1d_impl<float>      (params, dst); break;
+        case GGML_TYPE_F16:  ggml_compute_forward_col2im_1d_impl<ggml_fp16_t>(params, dst); break;
+        case GGML_TYPE_BF16: ggml_compute_forward_col2im_1d_impl<ggml_bf16_t>(params, dst); break;
+        default: GGML_ABORT("col2im_1d: unsupported type %d", dst->src[0]->type);
+    }
+}
+
 // ggml_compute_forward_conv_2d
 
 
@@ -6923,16 +7153,15 @@ void ggml_compute_forward_conv_3d(
     ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
 }
 
-// ggml_compute_forward_conv_transpose_2d
-
-void ggml_compute_forward_conv_transpose_2d(
-        const ggml_compute_params * params,
-              ggml_tensor * dst) {
+template <typename kernel_t>
+static void ggml_compute_forward_conv_transpose_2d_impl(
+    const ggml_compute_params * params,
+          ggml_tensor * dst) {
 
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
@@ -6943,7 +7172,7 @@ void ggml_compute_forward_conv_transpose_2d(
 
     const int nk = ne00*ne01*ne02*ne03;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb00 == ggml_type_size(src0->type));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (ith == 0) {
@@ -6951,12 +7180,12 @@ void ggml_compute_forward_conv_transpose_2d(
 
         // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
         {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
+            kernel_t * const wdata = (kernel_t *) params->wdata + 0;
 
             for (int64_t i03 = 0; i03 < ne03; i03++) {
                 for (int64_t i02 = 0; i02 < ne02; i02++) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02);
-                    ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03;
+                    const kernel_t * const src = (kernel_t *)((char *) src0->data + i03*nb03 + i02*nb02);
+                    kernel_t * dst_data = wdata + i02*ne01*ne00*ne03;
                     for (int64_t i01 = 0; i01 < ne01; i01++) {
                         for (int64_t i00 = 0; i00 < ne00; i00++) {
                             dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00];
@@ -6968,13 +7197,17 @@ void ggml_compute_forward_conv_transpose_2d(
 
         // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh)
         {
-            ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk;
+            kernel_t * const wdata = (kernel_t *) params->wdata + nk;
             for (int i12 = 0; i12 < ne12; i12++) {
                 for (int i11 = 0; i11 < ne11; i11++) {
                     const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
-                    ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
+                    kernel_t * dst_data = wdata + i11*ne10*ne12;
                     for (int i10 = 0; i10 < ne10; i10++) {
-                        dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
+                        } else {
+                            dst_data[i10*ne12 + i12] = src[i10];
+                        }
                     }
                 }
             }
@@ -6996,21 +7229,27 @@ void ggml_compute_forward_conv_transpose_2d(
     const int ip0 = dp*ith;
     const int ip1 = MIN(ip0 + dp, np);
 
-    ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
-    ggml_fp16_t * const wdata_src = wdata + nk;
+    kernel_t * const wdata = (kernel_t *) params->wdata + 0;
+    kernel_t * const wdata_src = wdata + nk;
 
     for (int i2 = ip0; i2 < ip1; i2++) { // Cout
         float * dst_data = (float *)((char *) dst->data + i2*nb2);
-        ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
+        kernel_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
         for (int i11 = 0; i11 < ne11; i11++) {
             for (int i10 = 0; i10 < ne10; i10++) {
                 const int i1n = i11*ne10*ne12 + i10*ne12;
                 for (int i01 = 0; i01 < ne01; i01++) {
                     for (int i00 = 0; i00 < ne00; i00++) {
                         float v = 0;
-                        ggml_vec_dot_f16(ne03, &v, 0,
-                                wdata_src + i1n, 0,
-                                wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            ggml_vec_dot_f16(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        } else {
+                            ggml_vec_dot_f32(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        }
                         dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
@@ -7019,6 +7258,28 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
+void ggml_compute_forward_conv_transpose_2d(
+        const ggml_compute_params * params,
+              ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_conv_transpose_2d_impl<ggml_fp16_t>(params, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_conv_transpose_2d_impl<float>(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_conv_2d_dw
 
 struct ggml_conv_2d_dw_params {
@@ -7038,6 +7299,13 @@ struct ggml_conv_2d_dw_params {
     int dilation_y;
 };
 
+static inline float ggml_conv_2d_dw_knl_f32(const char * data, int64_t i, ggml_type type) {
+    if (type == GGML_TYPE_F16) {
+        return GGML_FP16_TO_FP32(((const ggml_fp16_t *)data)[i]);
+    }
+    return ((const float *)data)[i];
+}
+
 static void ggml_compute_forward_conv_2d_dw_cwhn(
         const ggml_compute_params * params,
         const ggml_tensor * src,
@@ -7046,7 +7314,8 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
         const ggml_conv_2d_dw_params & p) {
 
     const int64_t c = p.channels;
-    const float * knl_data = (const float *)kernel->data;
+    const char * knl_data = (const char *)kernel->data;
+    const ggml_type knl_type = kernel->type;
 
     const int64_t rows_total = p.dst_h * p.batch;
     const int64_t rows_per_thread = (rows_total + params->nth - 1) / params->nth;
@@ -7054,13 +7323,16 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
     const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
 
 #ifdef GGML_SIMD
+    int64_t c_pkg_end = 0;
+    int64_t pkg_size = GGML_F32_EPR;
+    if (knl_type == GGML_TYPE_F32) {
     #if defined(__ARM_FEATURE_SVE)
-        const int64_t pkg_size = svcntw();
+        pkg_size = svcntw();
     #else
-        const int64_t pkg_size = GGML_F32_EPR;
+        pkg_size = GGML_F32_EPR;
     #endif
-    const int64_t pkg_count = c / pkg_size;
-    const int64_t c_pkg_end = pkg_count * pkg_size;
+        c_pkg_end = (c / pkg_size) * pkg_size;
+    }
 #else
     const int64_t c_pkg_end = 0;
 #endif
@@ -7074,7 +7346,6 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
             const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;
 
 #ifdef GGML_SIMD
-            // Vectorized loop
             for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {
                 GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
                 for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
@@ -7087,7 +7358,8 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
                         if (src_x < 0 || src_x >= p.src_w) {
                             continue;
                         }
-                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(knl_data + (knl_y * p.knl_w + knl_x) * c + c_i);
+                        const float * kp = (const float *)knl_data + (knl_y * p.knl_w + knl_x) * c + c_i;
+                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(kp);
                         GGML_F32_VEC s = GGML_F32_VEC_LOAD(src_data + (src_y * p.src_w + src_x) * c + c_i);
                         sum = GGML_F32_VEC_FMA(sum, k, s);
                     }
@@ -7095,7 +7367,6 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
                 GGML_F32_VEC_STORE(dst_data + c_i, sum);
             }
 #endif
-            // Scalar loop
             for (int64_t c_i = c_pkg_end; c_i < c; ++c_i) {
                 float sum = 0.0f;
                 for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
@@ -7108,7 +7379,7 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
                         if (src_x < 0 || src_x >= p.src_w) {
                             continue;
                         }
-                        sum += knl_data[(knl_y * p.knl_w + knl_x) * c + c_i]
+                        sum += ggml_conv_2d_dw_knl_f32(knl_data, (knl_y * p.knl_w + knl_x) * c + c_i, knl_type)
                              * src_data[(src_y * p.src_w + src_x) * c + c_i];
                     }
                 }
@@ -7129,9 +7400,11 @@ static void ggml_compute_forward_conv_2d_dw_whcn(
     const int64_t per_thread = (n + params->nth - 1) / params->nth;
     const int64_t start = params->ith * per_thread;
     const int64_t end = MIN(start + per_thread, n);
+    const char * knl_base = (const char *)kernel->data;
+    const ggml_type knl_type = kernel->type;
 
     for (int64_t i = start; i < end; ++i) {
-        const float * knl_data = (const float *)kernel->data + (i % p.channels) * p.knl_w * p.knl_h;
+        const int64_t knl_offset = (i % p.channels) * p.knl_w * p.knl_h;
         const float * src_data = (const float *)src->data + i * p.src_w * p.src_h;
         float * dst_data = (float *)dst->data + i * p.dst_w * p.dst_h;
 
@@ -7149,7 +7422,7 @@ static void ggml_compute_forward_conv_2d_dw_whcn(
                         if (src_x < 0 || src_x >= p.src_w) {
                             continue;
                         }
-                        sum += knl_data[knl_y * p.knl_w + knl_x]
+                        sum += ggml_conv_2d_dw_knl_f32(knl_base, knl_offset + knl_y * p.knl_w + knl_x, knl_type)
                              * src_data[src_y * p.src_w + src_x];
                     }
                 }
@@ -7181,13 +7454,13 @@ void ggml_compute_forward_conv_2d_dw(
     p.dilation_x = dst->op_params[4];
     p.dilation_y = dst->op_params[5];
 
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 || kernel->type == GGML_TYPE_F16);
     GGML_ASSERT(kernel->ne[3] == p.channels);
     GGML_ASSERT(dst->ne[3] == p.batch);
 
     if (ggml_is_contiguous(src)) {
         ggml_compute_forward_conv_2d_dw_whcn(params, src, kernel, dst, p);
     } else if (ggml_is_contiguous_channels(src)) {
-        // kernel should also have channels most contiguous in memory
         GGML_ASSERT(kernel->nb[0] >= kernel->nb[2] && kernel->nb[1] >= kernel->nb[0]);
         ggml_compute_forward_conv_2d_dw_cwhn(params, src, kernel, dst, p);
     } else {
@@ -8837,7 +9110,12 @@ static void ggml_compute_forward_flash_attn_ext_f16(
                                 k->type == v->type &&
                                 neq1 >= Q_TILE_SZ);
 #ifdef GGML_SIMD
-        use_tiled &= (DV % GGML_F32_EPR == 0);
+#if defined(__ARM_FEATURE_SVE)
+        const int64_t f32_epr = svcntw();
+#else
+        const int64_t f32_epr = GGML_F32_EPR;
+#endif
+        use_tiled &= (DV % f32_epr == 0);
 #endif
         int current_chunk = ith;
 
@@ -9922,13 +10200,9 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * k =          (float *) dst->src[0]->data;
     float * v =          (float *) dst->src[1]->data;
@@ -10139,13 +10413,9 @@ static void ggml_compute_forward_gla_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * k = (float *) dst->src[0]->data;
     float * v = (float *) dst->src[1]->data;
@@ -10437,19 +10707,29 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
 
     const bool kda = (neg0 == S_v);
 
-    // scratch layout per thread: [delta(S_v)]
-    const int64_t scratch_per_thread = S_v;
+    // K (snapshot slot count) is an op param; state holds s0 only [S_v, S_v, H, n_seqs].
+    const int64_t K = ggml_get_op_params_i32(dst, 0);
+    GGML_ASSERT(K >= 1);
+    // per-seq stride in floats (seq s starts at state + s * seq_stride)
+    const int64_t state_seq_stride = src_state->nb[3] / sizeof(float);
+
+    const int64_t per_thread = S_v + (K > 1 ? S_v * S_v : 0);
     const int ith = params->ith;
 
-    float * delta = (float *)params->wdata + ith * scratch_per_thread + CACHE_LINE_SIZE_F32;
+    float * delta       = (float *)params->wdata + ith * per_thread + CACHE_LINE_SIZE_F32;
+    float * state_work  = K > 1 ? (delta + S_v) : nullptr;
 
     // output layout: [attn_scores | new_states]
-    // attn_scores: S_v * H * n_tokens * n_seqs floats
-    // new_states:  S_v * S_v * H * n_seqs floats
-    const int64_t attn_score_elems = S_v * H * n_tokens * n_seqs;
+    // attn_scores: S_v * H * n_tokens * n_seqs    floats
+    // new_states:  S_v * S_v * H * n_seqs * K     floats  (K snapshot slots; last min(n_tokens, K))
+    const int64_t attn_score_elems    = S_v * H * n_tokens * n_seqs;
+    const int64_t state_size_per_snap = S_v * S_v * H * n_seqs;
     float * attn_out_base  = (float *)dst->data;
     float * state_out_base = (float *)dst->data + attn_score_elems;
 
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
     const float * state_in_base = (const float *)src_state->data;
 
   //const int64_t rq1 = nev1 / neq1;
@@ -10469,10 +10749,15 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
         const int64_t iq3 = iv3 / rq3;
         const int64_t ik3 = iv3 / rk3;
 
-        float * s_out = state_out_base + (iv3 * H + iv1) * S_v * S_v;
+        // For K=1, write directly to the single output slot to avoid an extra memcpy at the end.
+        // For K>1, work in scratch and copy out per-token when the slot is in range.
+        float * s_out = (K > 1)
+            ? state_work
+            : state_out_base + (iv3 * H + iv1) * S_v * S_v;
 
-        // copy input state into output buffer and operate in-place
-        const float * s_in = state_in_base + (iv3 * H + iv1) * S_v * S_v;
+        // copy input state into the working buffer and operate in-place
+        // state layout [S_v, S_v, H, n_seqs]: seq iv3 starts at iv3 * state_seq_stride.
+        const float * s_in = state_in_base + iv3 * state_seq_stride + iv1 * S_v * S_v;
         memcpy(s_out, s_in, S_v * S_v * sizeof(float));
 
         // attn output pointer for first token of this (head, seq)
@@ -10522,6 +10807,15 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
             }
 
             attn_data += S_v * H; // advance to next token
+
+            if (K > 1) {
+                const int64_t target_slot = n_tokens - 1 - t;
+                if (target_slot >= 0 && target_slot < K) {
+                    float * curr_state_o = state_out_base + target_slot * state_size_per_snap +
+                                     (iv3 * H + iv1) * S_v * S_v;
+                    memcpy(curr_state_o, s_out, S_v * S_v * sizeof(float));
+                }
+            }
         }
     }
 }
@@ -10602,13 +10896,9 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * r = (float *) dst->src[0]->data;
     float * w = (float *) dst->src[1]->data;
@@ -11186,3 +11476,95 @@ void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_
             }
     }
 }
+
+static void ggml_compute_forward_fwht_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t n = ne10;
+    GGML_ASSERT((n & (n - 1)) == 0); // must be power of 2
+
+    const int64_t nr = ne11 * ne12 * ne13;
+    const int64_t rows_per_thread = (nr + nth - 1) / nth;
+    const int64_t start_row = ith * rows_per_thread;
+    const int64_t end_row = MIN(start_row + rows_per_thread, nr);
+
+    const float scale = 1.0f / sqrtf((float)n);
+
+#if defined(GGML_SIMD)
+    const GGML_F32_VEC v_minus_one = GGML_F32_VEC_SET1(-1.0f);
+#endif
+
+    for (int64_t r = start_row; r < end_row; r++) {
+        const int64_t i13 = r / (ne11 * ne12);
+        const int64_t i12 = (r - i13 * ne11 * ne12) / ne11;
+        const int64_t i11 = r - i13 * ne11 * ne12 - i12 * ne11;
+
+        const float * src_row = (const float *) ((const char *) src1->data + i11 * nb11 + i12 * nb12 + i13 * nb13);
+        float * dst_row = (float *) ((char *) dst->data + i11 * nb1 + i12 * nb2 + i13 * nb3);
+
+        for (int64_t j = 0; j < n; j++) {
+            dst_row[j] = src_row[j] * scale;
+        }
+
+        // Scalar passes
+#if defined(GGML_SIMD)
+#if defined(__ARM_FEATURE_SVE)
+        const int step = svcntw();
+#else
+        const int step = GGML_F32_EPR;
+#endif
+#else
+        const int step = n;
+#endif
+        for (int64_t len = 1; len < step && len < n; len <<= 1) {
+            for (int64_t i = 0; i < n; i += 2 * len) {
+                for (int64_t j = 0; j < len; j++) {
+                    float u = dst_row[i + j];
+                    float v = dst_row[i + len + j];
+                    dst_row[i + j] = u + v;
+                    dst_row[i + len + j] = u - v;
+                }
+            }
+        }
+
+        // SIMD passes using GGML_F32_VEC_* macros for multi-architecture support
+#if defined(GGML_SIMD)
+        for (int64_t len = step; len < n; len <<= 1) {
+            for (int64_t i = 0; i < n; i += 2 * len) {
+                for (int64_t j = 0; j < len; j += step) {
+                    GGML_F32_VEC u = GGML_F32_VEC_LOAD(dst_row + i + j);
+                    GGML_F32_VEC v = GGML_F32_VEC_LOAD(dst_row + i + len + j);
+
+                    GGML_F32_VEC_STORE(dst_row + i + j,       GGML_F32_VEC_ADD(u, v));
+                    GGML_F32_VEC_STORE(dst_row + i + len + j, GGML_F32_VEC_FMA(u, v, v_minus_one));
+                }
+            }
+        }
+#endif
+    }
+}
+
+void ggml_compute_forward_fwht(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src1 = dst->src[1];
+
+    switch (src1->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_fwht_f32(params, dst);
+            }
+            break;
+        default:
+            {
+                GGML_ABORT("fatal error - fwht is F32 only");
+            }
+    }
+}