whispercpp 1.3.5 → 1.3.7

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

@@ -3,14 +3,14 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "binary-ops.h"
+#include "simd-gemm.h"
 #include "ggml.h"
 #include "unary-ops.h"
 #include "vec.h"
 
-#include <cfloat>
 #include <algorithm>
+#include <cfloat>
 #include <cmath>
-#include <functional>
 
 // ggml_compute_forward_dup
 
@@ -375,7 +375,7 @@ static void ggml_compute_forward_dup_bytes(
         const size_t rs = ne00 * type_size;
 
         if (nb00 == type_size) {
-            // src0 is contigous on first dimension, copy by rows
+            // src0 is contiguous on first dimension, copy by rows
             for (int64_t i03 = 0; i03 < ne03; i03++) {
                 for (int64_t i02 = 0; i02 < ne02; i02++) {
                     id += rs * ir0;
@@ -664,12 +664,14 @@ void ggml_compute_forward_add(
             {
                 ggml_compute_forward_add_non_quantized(params, dst);
             } break;
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1112,6 +1114,7 @@ void ggml_compute_forward_add1(
                     GGML_ABORT("fatal error");
                 }
             } break;
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1119,6 +1122,7 @@ void ggml_compute_forward_add1(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1240,6 +1244,7 @@ void ggml_compute_forward_acc(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -1247,6 +1252,7 @@ void ggml_compute_forward_acc(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1795,7 +1801,7 @@ void ggml_compute_forward_repeat(
             {
                 ggml_compute_forward_repeat_f32(params, dst);
             } break;
-        // TODO: templateify the implemenation and support for I64
+        // TODO: templateify the implementation and support for I64
         //       ref https://github.com/ggml-org/llama.cpp/pull/14274#discussion_r2169492225
         //case GGML_TYPE_I64:
         //    {
@@ -2097,10 +2103,14 @@ static void ggml_compute_forward_gelu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2114,19 +2124,23 @@ static void ggml_compute_forward_gelu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2136,10 +2150,14 @@ static void ggml_compute_forward_gelu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2153,20 +2171,24 @@ static void ggml_compute_forward_gelu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2213,8 +2235,42 @@ 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
@@ -2277,10 +2333,14 @@ static void ggml_compute_forward_gelu_erf_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2294,19 +2354,23 @@ static void ggml_compute_forward_gelu_erf_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2316,10 +2380,14 @@ static void ggml_compute_forward_gelu_erf_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2333,20 +2401,24 @@ static void ggml_compute_forward_gelu_erf_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_erf_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2380,10 +2452,14 @@ static void ggml_compute_forward_gelu_quick_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2397,19 +2473,23 @@ static void ggml_compute_forward_gelu_quick_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2419,10 +2499,14 @@ static void ggml_compute_forward_gelu_quick_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2436,20 +2520,24 @@ static void ggml_compute_forward_gelu_quick_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_gelu_quick_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2483,10 +2571,14 @@ static void ggml_compute_forward_silu_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2500,19 +2592,23 @@ static void ggml_compute_forward_silu_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const float x = ((float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*(dst->nb[1])))[k];
             GGML_UNUSED(x);
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2522,10 +2618,14 @@ static void ggml_compute_forward_silu_f16(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
     assert(ggml_are_same_shape(src0, dst));
 
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -2539,20 +2639,24 @@ static void ggml_compute_forward_silu_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
         ggml_vec_silu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
 
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
-            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2702,7 +2806,7 @@ static void ggml_compute_forward_silu_back_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2738,7 +2842,7 @@ static void ggml_compute_forward_silu_back_f16(
                 (ggml_fp16_t *) ((char *) src1->data + i1*(src1->nb[1])),
                 (ggml_fp16_t *) ((char *) grad->data + i1*(grad->nb[1])));
 
-    #ifndef NDEBUG
+#ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
             const float v = GGML_CPU_FP16_TO_FP32(x);
@@ -2746,7 +2850,7 @@ static void ggml_compute_forward_silu_back_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-    #endif
+#endif // NDEBUG
     }
 }
 
@@ -2829,7 +2933,7 @@ static void ggml_compute_forward_reglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2889,7 +2993,7 @@ static void ggml_compute_forward_reglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -2972,7 +3076,7 @@ static void ggml_compute_forward_geglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3032,7 +3136,7 @@ static void ggml_compute_forward_geglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3115,7 +3219,7 @@ static void ggml_compute_forward_swiglu_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3175,7 +3279,7 @@ static void ggml_compute_forward_swiglu_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3266,7 +3370,7 @@ static void ggml_compute_forward_swiglu_oai_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3345,7 +3449,7 @@ static void ggml_compute_forward_geglu_erf_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3405,7 +3509,7 @@ static void ggml_compute_forward_geglu_erf_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3488,7 +3592,7 @@ static void ggml_compute_forward_geglu_quick_f32(
             assert(!isnan(x));
             assert(!isinf(x));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3548,7 +3652,7 @@ static void ggml_compute_forward_geglu_quick_f16(
             assert(!isnan(v));
             assert(!isinf(v));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -3643,11 +3747,27 @@ 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));
 
@@ -3656,11 +3776,10 @@ 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
@@ -3670,25 +3789,32 @@ 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);
+                }
             }
         }
     }
@@ -3703,7 +3829,31 @@ 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:
             {
@@ -3858,12 +4008,12 @@ 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;
+                }
             }
         }
     }
@@ -4264,12 +4414,14 @@ void ggml_compute_forward_out_prod(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4538,6 +4690,7 @@ void ggml_compute_forward_set(
             } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4545,6 +4698,7 @@ void ggml_compute_forward_set(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4760,6 +4914,7 @@ void ggml_compute_forward_get_rows(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4767,6 +4922,7 @@ void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5239,7 +5395,7 @@ static void ggml_compute_forward_soft_max_f32(
                     //printf("p[%d] = %f\n", i, p[i]);
                     assert(!isnan(wp[i]));
                 }
-#endif
+#endif // NDEBUG
 
                 float max = -INFINITY;
                 ggml_vec_max_f32(ne00, &max, wp);
@@ -5264,7 +5420,7 @@ static void ggml_compute_forward_soft_max_f32(
                     assert(!isnan(dp[i]));
                     assert(!isinf(dp[i]));
                 }
-#endif
+#endif // NDEBUG
             }
         }
     }
@@ -5338,7 +5494,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dy[i]));
             assert(!isnan(y[i]));
         }
-#endif
+#endif // NDEBUG
         // Jii = yi - yi*yi
         // Jij = -yi*yj
         // J = diag(y)-y.T*y
@@ -5371,7 +5527,7 @@ static void ggml_compute_forward_soft_max_ext_back_f32(
             assert(!isnan(dx[i]));
             assert(!isinf(dx[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -5484,6 +5640,7 @@ 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_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -5491,6 +5648,7 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5739,28 +5897,33 @@ static void ggml_compute_forward_rope_flt(
 
     const int32_t * pos = (const int32_t *) src1->data;
 
+    int64_t last_i2 = -1;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) { // batch
         for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
-
-            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            if (!mrope_used) {
-                const int64_t p = pos[i2];
-                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-            else {
-                const int64_t p_t = pos[i2];
-                const int64_t p_h = pos[i2 + ne2];
-                const int64_t p_w = pos[i2 + ne2 * 2];
-                const int64_t p_e = pos[i2 + ne2 * 3];
-                ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
-                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
-            }
-
             for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads
-                if (ir++ < ir0) continue;
+                if (ir++ < ir0) continue; // skip rows mapped to other threads
                 if (ir   > ir1) break;
 
+                float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+                if (last_i2 != i2) {
+                    if (!mrope_used) {
+                        const int64_t p = pos[i2];
+                        ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+                    }
+                    else {
+                        const int64_t p_t = pos[i2];
+                        const int64_t p_h = pos[i2 + ne2];
+                        const int64_t p_w = pos[i2 + ne2 * 2];
+                        const int64_t p_e = pos[i2 + ne2 * 3];
+                        ggml_mrope_cache_init(
+                            p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
+                            freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+                    }
+
+                    last_i2 = i2;
+                }
+
                 T * src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
                 T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1);
 
@@ -6129,7 +6292,7 @@ static void ggml_compute_forward_im2col_f16(
     const ggml_tensor * src1 = dst->src[1];
 
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F16);
 
     GGML_TENSOR_BINARY_OP_LOCALS;
@@ -6160,7 +6323,7 @@ static void ggml_compute_forward_im2col_f16(
     int ofs1 = is_2D ? nb12 : nb11;
 
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
 
     // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
     {
@@ -6173,7 +6336,12 @@ static void ggml_compute_forward_im2col_f16(
 
                         // micro kernel
                         ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
-                        const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW]
+                        const float * const src_data_f32 = src1->type == GGML_TYPE_F32
+                            ? (const float *)((const char *) src1->data + in*ofs0 + iic*ofs1)
+                            : nullptr; // [IH, IW]
+                        const ggml_fp16_t * const src_data_f16 = src1->type == GGML_TYPE_F16
+                            ? (const ggml_fp16_t *)((const char *) src1->data + in*ofs0 + iic*ofs1)
+                            : nullptr; // [IH, IW]
 
                         for (int64_t ikh = 0; ikh < KH; ikh++) {  // 1
                             for (int64_t ikw = 0; ikw < KW; ikw++) {
@@ -6183,7 +6351,11 @@ static void ggml_compute_forward_im2col_f16(
                                 if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
                                     dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
                                 } else {
-                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
+                                    if (src_data_f32 != nullptr) {
+                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data_f32[iih*IW + iiw]);
+                                    } else {
+                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = src_data_f16[iih*IW + iiw];
+                                    }
                                 }
                             }
                         }
@@ -6558,6 +6730,78 @@ 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
 
 
@@ -6838,16 +7082,15 @@ 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);
 
@@ -6858,7 +7101,7 @@ 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) {
@@ -6866,12 +7109,12 @@ 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];
@@ -6883,13 +7126,17 @@ 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];
+                        }
                     }
                 }
             }
@@ -6911,21 +7158,27 @@ 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;
                     }
                 }
@@ -6934,19 +7187,41 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
-// ggml_compute_forward_conv_2d_dw
+void ggml_compute_forward_conv_transpose_2d(
+        const ggml_compute_params * params,
+              ggml_tensor * dst) {
 
-struct ggml_conv_2d_dw_params {
-    int64_t channels;
-    int64_t batch;
-    int64_t src_w;
-    int64_t src_h;
-    int64_t dst_w;
-    int64_t dst_h;
-    int64_t knl_w;
-    int64_t knl_h;
-    int stride_x;
-    int stride_y;
+    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 {
+    int64_t channels;
+    int64_t batch;
+    int64_t src_w;
+    int64_t src_h;
+    int64_t dst_w;
+    int64_t dst_h;
+    int64_t knl_w;
+    int64_t knl_h;
+    int stride_x;
+    int stride_y;
     int pad_x;
     int pad_y;
     int dilation_x;
@@ -7110,12 +7385,13 @@ void ggml_compute_forward_conv_2d_dw(
     }
 }
 
-// ggml_compute_forward_pool_1d_sk_p0
-
-static void ggml_compute_forward_pool_1d_sk_p0(
+// ggml_compute_forward_pool_1d_ksp
+static void ggml_compute_forward_pool_1d_ksp(
         const ggml_compute_params * params,
         const ggml_op_pool op,
         const int k,
+        const int s,
+        const int p,
         ggml_tensor * dst) {
 
     const ggml_tensor * src = dst->src[0];
@@ -7126,39 +7402,56 @@ static void ggml_compute_forward_pool_1d_sk_p0(
         return;
     }
 
-    const char * cdata = (const char *)src->data;
-    const char * const data_end = cdata + ggml_nbytes(src);
-    float * drow = (float *)dst->data;
+    const int64_t IW = src->ne[0];
+    const int64_t OW = dst->ne[0];
 
-    const int64_t rs = dst->ne[0];
+    const int64_t nr = ggml_nrows(src);
 
-    while (cdata < data_end) {
-        const void * srow = (const void *)cdata;
-        int j = 0;
-        for (int64_t i = 0; i < rs; ++i) {
+    for (int64_t ir = 0; ir < nr; ++ir) {
+        const char * srow_bytes =            (const char *) src->data + ir * src->nb[1];
+        float      * drow       = (float *) ((      char *) dst->data + ir * dst->nb[1]);
+
+        for (int64_t ow = 0; ow < OW; ++ow) {
+            float res = 0;
             switch (op) {
-                case GGML_OP_POOL_AVG:   drow[i] = 0;        break;
-                case GGML_OP_POOL_MAX:   drow[i] = -FLT_MAX; break;
+                case GGML_OP_POOL_AVG: res = 0.0f;     break;
+                case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
                 case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
+
+            int count = 0;
+            const int base = (int) ow * s - p;
+
             for (int ki = 0; ki < k; ++ki) {
-                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
+                const int j = base + ki;
+                if (j < 0 || j >= (int) IW) {
+                    continue;
+                }
+
+                float v;
+                if (src->type == GGML_TYPE_F32) {
+                    v = ((const float *) srow_bytes)[j];
+                } else {
+                    v = GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) srow_bytes)[j]);
+                }
+
                 switch (op) {
-                    case GGML_OP_POOL_AVG:                         drow[i] += srow_j; break;
-                    case GGML_OP_POOL_MAX:   if (srow_j > drow[i]) drow[i]  = srow_j; break;
-                    case GGML_OP_POOL_COUNT:                       GGML_ABORT("fatal error");
+                    case GGML_OP_POOL_AVG: res += v;                break;
+                    case GGML_OP_POOL_MAX: res =  std::max(v, res); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
-                ++j;
+
+                ++count;
             }
+
             switch (op) {
-                case GGML_OP_POOL_AVG:         drow[i] /= k; break;
-                case GGML_OP_POOL_MAX:                       break;
+                case GGML_OP_POOL_AVG: res = (count > 0) ? (res / count) : 0.0f; break;
+                case GGML_OP_POOL_MAX:                                           break;
                 case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
-        }
 
-        cdata += src->nb[1];
-        drow  += rs;
+            drow[ow] = res;
+        }
     }
 }
 
@@ -7173,10 +7466,8 @@ void ggml_compute_forward_pool_1d(
     const int k0 = opts[1];
     const int s0 = opts[2];
     const int p0 = opts[3];
-    GGML_ASSERT(p0 == 0); // padding not supported
-    GGML_ASSERT(k0 == s0); // only s = k supported
 
-    ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst);
+    ggml_compute_forward_pool_1d_ksp(params, op, k0, s0, p0, dst);
 }
 
 // ggml_compute_forward_pool_2d
@@ -7194,6 +7485,7 @@ void ggml_compute_forward_pool_2d(
     }
 
     const int32_t * opts = (const int32_t *)dst->op_params;
+
     ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
     const int k0 = opts[1];
     const int k1 = opts[2];
@@ -7217,11 +7509,13 @@ void ggml_compute_forward_pool_2d(
     while (cdata < data_end) {
         for (int oy = 0; oy < py; ++oy) {
             float * const drow = dplane + oy * px;
+            float * const out  = drow;
+
             for (int ox = 0; ox < px; ++ox) {
-                float * const out =  drow + ox;
+                float res = 0;
                 switch (op) {
-                    case GGML_OP_POOL_AVG:     *out = 0;        break;
-                    case GGML_OP_POOL_MAX:     *out = -FLT_MAX; break;
+                    case GGML_OP_POOL_AVG: res = 0;        break;
+                    case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
                     case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
 
@@ -7229,24 +7523,32 @@ void ggml_compute_forward_pool_2d(
                 const int iy = offset1 + oy * s1;
 
                 for (int ky = 0; ky < k1; ++ky) {
-                    if (iy + ky < 0 || iy + ky >= src->ne[1]) continue;
+                    if (iy + ky < 0 || iy + ky >= src->ne[1]) {
+                        continue;
+                    }
+
                     const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky));
                     for (int kx = 0; kx < k0; ++kx) {
                         int j = ix + kx;
-                        if (j < 0 || j >= src->ne[0]) continue;
+                        if (j < 0 || j >= src->ne[0]) {
+                            continue;
+                        }
+
                         const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                         switch (op) {
-                            case GGML_OP_POOL_AVG:                     *out += srow_j; break;
-                            case GGML_OP_POOL_MAX: if (srow_j > *out)  *out  = srow_j; break;
+                            case GGML_OP_POOL_AVG: res += srow_j;                break;
+                            case GGML_OP_POOL_MAX: res =  std::max(srow_j, res); break;
                             case GGML_OP_POOL_COUNT:               GGML_ABORT("fatal error");
                         }
                     }
                 }
                 switch (op) {
-                    case GGML_OP_POOL_AVG:           *out /= ka; break;
-                    case GGML_OP_POOL_MAX:                       break;
+                    case GGML_OP_POOL_AVG:           res /= ka; break;
+                    case GGML_OP_POOL_MAX:                      break;
                     case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
+
+                out[ox] = res;
             }
         }
 
@@ -7603,8 +7905,7 @@ static void ggml_compute_forward_pad_f32(
 
     const ggml_tensor * src0 = dst->src[0];
 
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-    GGML_ASSERT( dst->nb[0] == sizeof(float));
+    assert(dst->nb[0] == sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -8016,12 +8317,14 @@ void ggml_compute_forward_top_k(
     }
 }
 
-// ggml_compute_forward_flash_attn_ext
-
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
         ggml_tensor * dst,
-        int ir0, int ir1) {
+        int ir0, int ir1,
+        int64_t ic_start, int64_t ic_end,
+        float * partials, int64_t partial_stride) {
+
+    const bool write_partials = (partials != nullptr);
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -8098,7 +8401,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
     int ith = params->ith;
 
-    // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
         const int iq3 = ir/(neq2*neq1);
@@ -8138,7 +8440,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         // online softmax / attention
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
-        for (int64_t ic = 0; ic < nek1; ++ic) {
+
+        for (int64_t ic = ic_start; ic < ic_end; ++ic) {
             const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
@@ -8211,8 +8514,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             }
         }
 
-        // sinks
-        if (sinks) {
+        // sinks - apply only on the first kv-chunk
+        if (sinks && ic_start == 0) {
             const float s = ((float *)((char *) sinks->data))[h];
 
             float ms = 1.0f;
@@ -8220,6 +8523,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
 
             if (s > M) {
                 ms = expf(M - s);
+                M = s;
                 ggml_vec_scale_f32(DV, VKQ32, ms);
             } else {
                 vs = expf(s - M);
@@ -8228,20 +8532,386 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
             S = S*ms + vs;
         }
 
-        // V /= S
-        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
-        ggml_vec_scale_f32(DV, VKQ32, S_inv);
+        if (write_partials) {
+            // Write M, S, VKQ to partials for later reduction
+            // partials layout: [M, S, VKQ[DV]] per query head
+            float * partial = partials + ir * partial_stride;
+            partial[0] = M;
+            partial[1] = S;
+            memcpy(partial + 2, VKQ32, DV * sizeof(float));
+        } else {
+            // V /= S
+            const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
+            ggml_vec_scale_f32(DV, VKQ32, S_inv);
 
-        // dst indices
-        const int i1 = iq1;
-        const int i2 = iq2;
-        const int i3 = iq3;
+            // dst indices
+            const int i1 = iq1;
+            const int i2 = iq2;
+            const int i3 = iq3;
+
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+        }
+    }
+}
+
+static void ggml_compute_forward_flash_attn_ext_tiled(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        int ir0, int ir1) {
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(k->type == v->type);
+    const ggml_type kv_type = k->type;
+
+
+    // broadcast factors
+    const int64_t rk2 = neq2/nek2;
+    const int64_t rk3 = neq3/nek3;
+
+    const int64_t rv2 = neq2/nev2;
+    const int64_t rv3 = neq3/nev3;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head      = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    int ith = params->ith;
+
+    static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
+    static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
+
+    int ir = ir0;
+    while (ir < ir1) {
+        // q indices for the start of this tile
+        const int iq3 = ir/(neq2*neq1);
+        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
+        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
+
+        // Number of valid rows in this tile:
+        // - limited by tile size (Q_TILE_SZ)
+        // - limited by chunk boundary (ir1 - ir)
+        // - limited by head boundary (neq1 - iq1) to avoid crossing into next head
+        const int tile_rows = MIN(Q_TILE_SZ, MIN((int)(ir1 - ir), (int)(neq1 - iq1)));
+        GGML_ASSERT(tile_rows > 0);
+
+        const uint32_t h = iq2; // head index
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
+        float S[Q_TILE_SZ];
+        float M[Q_TILE_SZ];
+
+        for (int i = 0 ; i < Q_TILE_SZ; ++i) {
+            S[i] = 0.;
+            M[i] = -INFINITY;
+        }
+
+        // Per-thread scratch layout:
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile — F32 for GEMM, KV type for scalar)
+        // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
+        // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
+        // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile)
+        // K_f32:  KV_TILE_SZ * DK (F32 buffer for K tile — GEMM path)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + KV_TILE_SZ*DK + CACHE_LINE_SIZE_F32);
+
+        void  * Q_q    = base;
+        float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
+        float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
+        float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;
+        float * K_f32  = V32 + KV_TILE_SZ * DV;
+
+        memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
+        memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+
+        // k indices
+        const int ik3 = iq3 / rk3;
+        const int ik2 = iq2 / rk2;
+
+        // v indices
+        const int iv3 = iq3 / rv3;
+        const int iv2 = iq2 / rv2;
+
+        {
+            float * Q_f32 = (float *)Q_q;
+            for (int tq = 0; tq < tile_rows; tq++) {
+                const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+                memcpy(Q_f32 + tq * DK, pq, DK * sizeof(float));
+            }
+            for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+                memset(Q_f32 + tq * DK, 0, DK * sizeof(float));
+            }
+        }
+
+        memset(K_f32, 0, DK * KV_TILE_SZ * sizeof(float));
+        memset(V32,   0, KV_TILE_SZ * DV * sizeof(float));
+
+        for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, nek1 - ic);
+
+            // skip the tile entirely if all the masks are -inf
+            if (mask) {
+                bool can_skip = true;
+                for (int tq = 0; tq < tile_rows; tq++) {
+                    const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
+                    for (int tk = 0; tk < kv_tile; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
+                        if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
+                            can_skip = false;
+                        }
+                    }
+                    // Pad remaining mask entries with -inf
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
+                }
+
+                if (can_skip) {
+                    continue;
+                }
+            }
+
+            // Pack K tile transposed: K_f32[dk][kv] so KV_TILE is contiguous (SIMD dim)
+            // Zero-pad the last tile so the GEMM always operates on KV_TILE_SZ columns
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * k_data = (const char *)k->data + (ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
+                if (kv_type == GGML_TYPE_F16) {
+                    const ggml_fp16_t * k_f16 = (const ggml_fp16_t *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = GGML_CPU_FP16_TO_FP32(k_f16[dk]);
+                    }
+                } else {
+                    const float * k_f32_src = (const float *)k_data;
+                    for (int64_t dk = 0; dk < DK; dk++) {
+                        K_f32[dk * KV_TILE_SZ + tk] = k_f32_src[dk];
+                    }
+                }
+            }
+            memset(KQ, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+            simd_gemm(KQ, (const float *)Q_q, K_f32, Q_TILE_SZ, DK, KV_TILE_SZ);
+            ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, scale);
+
+            // Set padded KQ entries to -inf so softmax gives them zero weight
+            if (kv_tile < KV_TILE_SZ) {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    for (int tk = kv_tile; tk < KV_TILE_SZ; tk++) {
+                        KQ[tq * KV_TILE_SZ + tk] = -INFINITY;
+                    }
+                }
+            }
+
+            if (logit_softcap != 0.0f) {
+                ggml_vec_tanh_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, KQ);
+                ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, logit_softcap);
+            }
+
+            if (mask) {
+                ggml_vec_add_f32(tile_rows * KV_TILE_SZ, KQ, KQ, mask32);
+            }
+
+            bool skip[Q_TILE_SZ] = {};
+
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                float * kq_row = KQ + tq * KV_TILE_SZ;
+
+                float tile_max;
+                ggml_vec_max_f32(KV_TILE_SZ, &tile_max, kq_row);
+
+                if (tile_max == -INFINITY) {
+                    skip[tq] = true;
+                    continue;
+                }
+
+                const float Mold = M[tq];
+                const float Mnew = fmaxf(Mold, tile_max);
+
+                if (Mnew > Mold) {
+                    const float ms = expf(Mold - Mnew);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                    S[tq] *= ms;
+                }
+                M[tq] = Mnew;
+
+
+                S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
+            }
+
+            // V accumulation: VKQ32 += softmax(KQ) * V
+            // Pack V tile to contiguous F32, zero-padded
+            for (int tk = 0; tk < kv_tile; tk++) {
+                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
+                if (kv_type == GGML_TYPE_F16) {
+                    ggml_fp16_to_fp32_row((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
+                } else {
+                    memcpy(V32 + tk * DV, v_data, DV * sizeof(float));
+                }
+            }
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                if (skip[tq]) {
+                    memset(KQ + tq * KV_TILE_SZ, 0, KV_TILE_SZ * sizeof(float));
+                }
+            }
+            simd_gemm(VKQ32, KQ, V32, Q_TILE_SZ, KV_TILE_SZ, DV);
+        }
+
+        // sinks (apply only to valid rows in the tile)
+        if (sinks) {
+            const float s = ((float *)((char *) sinks->data))[h];
+
+            for (int tq = 0; tq < tile_rows; tq++) {
+                float ms = 1.0f;
+                float vs = 1.0f;
+
+                if (s > M[tq]) {
+                    ms = expf(M[tq] - s);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                } else {
+                    vs = expf(s - M[tq]);
+                }
+
+                S[tq] = S[tq] * ms + vs;
+            }
+        }
+
+        for (int tq = 0; tq < tile_rows; tq++) {
+            // V /= S
+            const float S_inv = S[tq] == 0.0f ? 0.0f : 1.0f / S[tq];
+            ggml_vec_scale_f32(DV, VKQ32 + tq * DV, S_inv);
+
+            // dst indices
+            const int i1 = iq1 + tq;
+            const int i2 = iq2;
+            const int i3 = iq3;
+
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32 + tq * DV, nb1);
+        }
+
+        ir += tile_rows;
+    }
+}
+
+// Reduction function: combines partial results across KV chunks
+// Partials layout in wdata: [n_q_heads][n_chunks][2 + DV]
+static void ggml_flash_attn_ext_reduce_partials(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        const int64_t n_chunks,
+        const int64_t chunk_size) {
+
+    const ggml_tensor * q = dst->src[0];
+    const ggml_tensor * k = dst->src[1];
+    const ggml_tensor * v = dst->src[2];
+
+    const int64_t DK        = k->ne[0];
+    const int64_t DV        = v->ne[0];
+    const int64_t nek1      = k->ne[1];
+    const int64_t n_q_heads = q->ne[2];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t wdata_per_thread = DK + 2*DV + CACHE_LINE_SIZE_F32;
+    float *       thread_wdata     = (float *) params->wdata + ith * wdata_per_thread;
+
+    const int64_t partials_offset  = nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
+    const int64_t partial_size     = 2 + DV;
+    const float * partials_base    = (const float *) params->wdata + partials_offset;
+
+    // Output layout
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const size_t  nb1 = dst->nb[1];
 
-        // original
-        //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
+    // Each thread reduces a subset of query heads
+    for (int64_t q_head = ith; q_head < n_q_heads; q_head += nth) {
+        float   M_final   = -INFINITY;
+        float   S_final   = 0.0f;
+        float * VKQ_final = thread_wdata;
+        memset(VKQ_final, 0, DV * sizeof(float));
 
-        // permute(0, 2, 1, 3)
-        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+        // Combine partials from all chunks
+        for (int64_t chunk_idx = 0; chunk_idx < n_chunks; ++chunk_idx) {
+            const int64_t ic_start = chunk_idx * chunk_size;
+            if (ic_start >= nek1) continue;
+
+            const float * partial   = partials_base + (q_head * n_chunks + chunk_idx) * partial_size;
+            const float   M_chunk   = partial[0];
+            const float   S_chunk   = partial[1];
+            const float * VKQ_chunk = partial + 2;
+
+            if (S_chunk == 0.0f) continue;
+
+            const float M_new     = fmaxf(M_final, M_chunk);
+            const float scale_old = expf(M_final - M_new);
+            const float scale_new = expf(M_chunk - M_new);
+
+            for (int64_t d = 0; d < DV; ++d) {
+                VKQ_final[d] = VKQ_final[d] * scale_old + VKQ_chunk[d] * scale_new;
+            }
+            S_final = S_final * scale_old + S_chunk * scale_new;
+            M_final = M_new;
+        }
+
+        // Normalize and write to output
+        if (S_final != 0.0f) {
+            const float S_inv = 1.0f / S_final;
+            ggml_vec_scale_f32(DV, VKQ_final, S_inv);
+        }
+        // iq1=0, iq3=0 for decode
+        memcpy((char *) dst->data + (0*ne2*ne1 + q_head + 0*ne1)*nb1, VKQ_final, nb1);
     }
 }
 
@@ -8266,6 +8936,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const int64_t DV = nev0;
     const int64_t N  = neq1;
 
+
     GGML_ASSERT(ne0 == DV);
     GGML_ASSERT(ne2 == N);
 
@@ -8286,47 +8957,97 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int64_t nr = neq1*neq2*neq3;
-
-    // rows per thread
     const int ith = params->ith;
     const int nth = params->nth;
 
-    // disable for NUMA
-    const bool disable_chunking = ggml_is_numa();
+    // When use_ref is set, force the vec-only reference implementation (no tiling, no KV-chunking)
+    const bool use_ref = params->use_ref;
 
-    // 4x chunks per thread
-    int nth_scaled = nth * 4;
-    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
-    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+    const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
+    const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;
 
-    if (nth == 1 || nchunk < nth || disable_chunking) {
-        nchunk = nth;
-    }
+    if (use_split_kv_path) {
+        const int64_t chunk_size = (nek1 + nth - 1) / nth;
 
-    if (ith == 0) {
-        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-        ggml_threadpool_chunk_set(params->threadpool, nth);
-    }
+        // Partials buffer layout: [q_head][kv_chunk][M, S, VKQ]
+        const int64_t partial_size  = 2 + DV;
+        float *       partials_base = (float *) params->wdata + nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
 
-    ggml_barrier(params->threadpool);
+        const int64_t ic_start = ith * chunk_size;
+        const int64_t ic_end   = std::min(ic_start + chunk_size, nek1);
 
-    // The number of elements in each chunk
-    const int64_t dr = (nr + nchunk - 1) / nchunk;
+        const int64_t partial_stride = nth * partial_size;
+        float *       chunk_partials = partials_base + ith * partial_size;
 
-    // The first chunk comes from our thread_id, the rest will get auto-assigned.
-    int current_chunk = ith;
+        if (ic_start < nek1) {
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(
+                    params, dst, q_head, q_head + 1, ic_start, ic_end,
+                    chunk_partials, partial_stride);
+            }
+        } else {
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                float * q_partials = chunk_partials + q_head * partial_stride;
+                q_partials[0] = -INFINITY;  // M
+                q_partials[1] = 0.0f;       // S
+            }
+        }
 
-    while (current_chunk < nchunk) {
-        const int64_t ir0 = dr * current_chunk;
-        const int64_t ir1 = MIN(ir0 + dr, nr);
+        ggml_barrier(params->threadpool);
+        ggml_flash_attn_ext_reduce_partials(params, dst, nth, chunk_size);
+    } else {
 
-        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+        // total rows in q
+        const int64_t nr = neq1*neq2*neq3;
 
-        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        if (ith == 0) {
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+        static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+        bool use_tiled = !use_ref &&
+                               (q->type == GGML_TYPE_F32 &&
+                                kv_is_f32_or_f16 &&
+                                k->type == v->type &&
+                                neq1 >= Q_TILE_SZ);
+#ifdef GGML_SIMD
+#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;
+
+        while (current_chunk < nchunk) {
+            const int64_t ir0 = dr * current_chunk;
+            const int64_t ir1 = MIN(ir0 + dr, nr);
+
+            if (use_tiled) {
+                ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+            } else {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, nek1, nullptr, 0);
+            }
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
     }
 }
 
@@ -9107,7 +9828,7 @@ void ggml_compute_forward_win_unpart(
     }
 }
 
-//gmml_compute_forward_unary
+//ggml_compute_forward_unary
 
 void ggml_compute_forward_unary(
         const ggml_compute_params * params,
@@ -9396,13 +10117,9 @@ 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;
@@ -9613,13 +10330,9 @@ 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;
@@ -9870,6 +10583,219 @@ void ggml_compute_forward_solve_tri(const struct ggml_compute_params * params, s
     }
 }
 
+// ggml_compute_forward_gated_delta_net
+static void ggml_compute_forward_gated_delta_net_one_chunk(
+    const ggml_compute_params * params,
+    ggml_tensor * dst,
+    int64_t ir0,
+    int64_t ir1) {
+
+    ggml_tensor * src_q     = dst->src[0];
+    ggml_tensor * src_k     = dst->src[1];
+    ggml_tensor * src_v     = dst->src[2];
+    ggml_tensor * src_g     = dst->src[3];
+    ggml_tensor * src_beta  = dst->src[4];
+    ggml_tensor * src_state = dst->src[5];
+
+    const int64_t S_v      = src_v->ne[0];
+    const int64_t H        = src_v->ne[1];
+    const int64_t n_tokens = src_v->ne[2];
+    const int64_t n_seqs   = src_v->ne[3];
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src_q));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_k));
+    GGML_ASSERT(ggml_is_contiguous_rows(src_v));
+    GGML_ASSERT(ggml_is_contiguous(src_g));
+    GGML_ASSERT(ggml_is_contiguous(src_beta));
+    GGML_ASSERT(ggml_is_contiguous(src_state));
+
+    GGML_ASSERT(src_g->ne[0] == 1 || src_g->ne[0] == S_v);
+    GGML_ASSERT(src_beta->ne[0] == 1);
+
+    GGML_TENSOR_LOCALS(int64_t, neq, src_q, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbq, src_q, nb);
+    GGML_TENSOR_LOCALS(int64_t, nek, src_k, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbk, src_k, nb);
+    GGML_TENSOR_LOCALS(int64_t, nev, src_v, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbv, src_v, nb);
+    GGML_TENSOR_LOCALS(int64_t, neg, src_g, ne);
+    GGML_TENSOR_LOCALS(size_t,  nbg, src_g, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbb, src_beta, nb);
+
+    const bool kda = (neg0 == 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 * 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 * 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;
+  //const int64_t rk1 = nev1 / nek1;
+    const int64_t rq3 = nev3 / neq3;
+    const int64_t rk3 = nev3 / nek3;
+
+    const float scale = 1.0f / sqrtf((float) S_v);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t iv1 = ir % H; // head_index
+        const int64_t iv3 = ir / H; // sequence
+
+        const int64_t iq1 = iv1 % neq1;
+        const int64_t ik1 = iv1 % nek1;
+
+        const int64_t iq3 = iv3 / rq3;
+        const int64_t ik3 = iv3 / rk3;
+
+        // 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 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)
+        float * attn_data = attn_out_base + (iv3 * n_tokens * H + iv1) * S_v;
+
+        for (int64_t t = 0; t < n_tokens; t++) {
+            const float * q_d = (const float *)((const char *)src_q->data + iq3 * nbq3 + t * nbq2 + iq1 * nbq1);
+            const float * k_d = (const float *)((const char *)src_k->data + ik3 * nbk3 + t * nbk2 + ik1 * nbk1);
+            const float * v_d = (const float *)((const char *)src_v->data + iv3 * nbv3 + t * nbv2 + iv1 * nbv1);
+
+            const float beta_val = *(const float *)((const char *)src_beta->data + iv3 * nbb3 + t * nbb2 + iv1 * nbb1);
+            const float * g_d    =  (const float *)((const char *)src_g->data    + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
+
+            // state is stored transposed: s_out[j*S_v + i] = S[i][j]
+            // so row j of s_out = column j of S (contiguous access)
+
+            if (kda) {
+                // precompute exp(g) into delta scratch (reused below)
+                for (int64_t i = 0; i < S_v; ++i) {
+                    delta[i] = expf(g_d[i]);
+                }
+                // S[i][:] *= exp(g[i]) => for each row j of M: M[j][i] *= exp(g[i])
+                for (int64_t j = 0; j < S_v; ++j) {
+                    ggml_vec_mul_f32(S_v, &s_out[j * S_v], &s_out[j * S_v], delta);
+                }
+            } else {
+                ggml_vec_scale_f32(S_v * S_v, s_out, expf(g_d[0]));
+            }
+
+            // delta[j] = sum_i S[i][j] * k[i] = dot(row j of M, k)
+            for (int64_t j = 0; j < S_v; ++j) {
+                float sum = 0.0f;
+                ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, k_d, 0, 1);
+                delta[j] = (v_d[j] - sum) * beta_val;
+            }
+
+            // outer product: S[i][j] += k[i] * delta[j] => M[j][i] += delta[j] * k[i]
+            for (int64_t j = 0; j < S_v; ++j) {
+                ggml_vec_mad_f32(S_v, &s_out[j * S_v], k_d, delta[j]);
+            }
+
+            // attn_out[j] = sum_i S[i][j] * q[i] = dot(row j of M, q)
+            for (int64_t j = 0; j < S_v; ++j) {
+                float sum = 0.0f;
+                ggml_vec_dot_f32(S_v, &sum, 0, &s_out[j * S_v], 0, q_d, 0, 1);
+                attn_data[j] = sum * scale;
+            }
+
+            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));
+                }
+            }
+        }
+    }
+}
+
+
+static void ggml_compute_forward_gated_delta_net_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    ggml_tensor * V = dst->src[2];
+    int64_t nr = V->ne[1] * V->ne[3];
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    int nth = params->nth;
+    int ith = params->ith;
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+      nchunk = nth;
+    }
+
+    if (ith == 0) {
+      ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_gated_delta_net_one_chunk(params, dst, ir0, ir1);
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
+void ggml_compute_forward_gated_delta_net(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_gated_delta_net_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_rwkv_wkv7
 
 static void ggml_compute_forward_rwkv_wkv7_f32(
@@ -9887,13 +10813,9 @@ 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;
@@ -10195,7 +11117,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         float max = -INFINITY;
         ggml_vec_max_f32(nc, &max, s0);
@@ -10214,7 +11136,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
             assert(!isnan(st[i]));
             assert(!isinf(st[i]));
         }
-#endif
+#endif // NDEBUG
     }
     sums[ith] = sum_thread;
     ggml_barrier(params->threadpool);
@@ -10287,7 +11209,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
-#endif
+#endif // NDEBUG
 
         // soft_max
         float max = -INFINITY;
@@ -10305,7 +11227,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
             assert(!isnan(ds0[i]));
             assert(!isinf(ds0[i]));
         }
-#endif
+#endif // NDEBUG
     }
 }
 
@@ -10471,3 +11393,95 @@ 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");
+            }
+    }
+}