whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-cuda/set-rows.cu

@@ -53,6 +53,7 @@ static __global__ void k_set_rows_quant(const float * __restrict__ src0,
     const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
+    ggml_cuda_pdl_sync();
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
 
     const float * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
@@ -110,9 +111,9 @@ static void set_rows_cuda_quant(
 }
 
 template <typename src_t, typename idx_t, typename dst_t>
-static __global__ void k_set_rows(const src_t * __restrict__ src0,
-                                  const idx_t * __restrict__ src1,
-                                  dst_t * __restrict__ dst,
+static __global__ void k_set_rows(const src_t * src0_ptr,
+                                  const idx_t * src1_ptr,
+                                  dst_t * dst_ptr,
                                   const int64_t ne_total,
                                   const int64_t ne10,
                                   const int64_t ne11,
@@ -132,6 +133,9 @@ static __global__ void k_set_rows(const src_t * __restrict__ src0,
                                   const uint3   ne02,
                                   const uint3   ne11_fd,
                                   const uint3   ne12_fd) {
+    const src_t * GGML_CUDA_RESTRICT src0 = src0_ptr;
+    const idx_t * GGML_CUDA_RESTRICT src1 = src1_ptr;
+    dst_t       * GGML_CUDA_RESTRICT dst  = dst_ptr;
     const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
 
     if (i >= ne_total) {
@@ -157,7 +161,9 @@ static __global__ void k_set_rows(const src_t * __restrict__ src0,
     const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
+    ggml_cuda_pdl_sync();
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+    ggml_cuda_pdl_lc();
 
     const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
     dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
@@ -203,9 +209,11 @@ static void set_rows_cuda(
         const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
         const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
 
-        k_set_rows<<<grid_size, block_size, 0, stream>>>(src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
-                                                         s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
-                                                         ne11_fd, ne12_fd);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_size, block_size, 0, stream);
+        ggml_cuda_kernel_launch(k_set_rows<src_t, idx_t, dst_t>, launch_params,
+            src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
+            s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
+            ne11_fd, ne12_fd);
     }
 }
 
@@ -314,17 +322,77 @@ static void set_rows_cuda(ggml_backend_cuda_context & ctx, const ggml_tensor * s
     }
 }
 
+template<>
+void set_rows_cuda<half, int32_t>(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const half    * src0_d = (const half *)src0->data;
+    const int32_t * src1_d = (const int32_t *)src1->data;
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    cudaStream_t stream = ctx.stream();
+
+
+    if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
+    }
+}
+
+template<>
+void set_rows_cuda<half, int64_t>(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const half    * src0_d = (const half *)src0->data;
+    const int64_t * src1_d = (const int64_t *)src1->data;
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    cudaStream_t stream = ctx.stream();
+
+
+    if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
+    }
+}
+
 
 void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16));
     GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);
 
-    if (src1->type == GGML_TYPE_I64) {
-        set_rows_cuda<float, int64_t>(ctx, src0, src1, dst);
+    if (src0->type == GGML_TYPE_F32) {
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_cuda<float, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_cuda<float, int32_t>(ctx, src0, src1, dst);
+        }
+    } else if (src0->type == GGML_TYPE_F16) {
+        if (src1->type == GGML_TYPE_I64) {
+            set_rows_cuda<half, int64_t>(ctx, src0, src1, dst);
+        } else {
+            set_rows_cuda<half, int32_t>(ctx, src0, src1, dst);
+        }
     } else {
-        set_rows_cuda<float, int32_t>(ctx, src0, src1, dst);
+        GGML_ABORT("unsupported type %s", ggml_type_name(src0->type));
     }
 }

data/ext/sources/ggml/src/ggml-cuda/snake.cu

@@ -0,0 +1,72 @@
+#include "snake.cuh"
+#include "convert.cuh"
+
+// Fused Snake activation: y = x + sin^2(a * x) * inv_b
+// x: [T, C] (T contiguous), a: [1, C], inv_b: [1, C]
+// Supports F32, F16, BF16 data with F32 compute.
+
+template <typename T>
+static __global__ void snake_kernel(
+        const T     * __restrict__ x,
+        const float * __restrict__ a,
+        const float * __restrict__ inv_b,
+        T           * __restrict__ dst,
+        const int    total,
+        const uint3  T_len_fastdiv) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total) return;
+
+    const int c = (int) fastdiv((uint32_t) idx, T_len_fastdiv);
+
+    const float xi = ggml_cuda_cast<float>(x[idx]);
+    const float s  = sinf(a[c] * xi);
+    dst[idx] = ggml_cuda_cast<T>(xi + s * s * inv_b[c]);
+}
+
+// Internal launcher with explicit x/a/inv_b/dst tensors.
+// Shared by the public op (reads dst->src) and the fusion path (explicit args).
+static void launch_snake(ggml_backend_cuda_context & ctx,
+                         const ggml_tensor * x,
+                         const ggml_tensor * a,
+                         const ggml_tensor * inv_b,
+                         ggml_tensor *       dst) {
+    const float * a_d     = (const float *)a->data;
+    const float * inv_b_d = (const float *)inv_b->data;
+
+    const int   T = (int)x->ne[0];
+    const int   C = (int)x->ne[1];
+    const int   total = T * C;
+    const uint3 T_len_fastdiv = init_fastdiv_values((uint64_t) T);
+
+    const int block_size = 256;
+    const int grid_size  = (total + block_size - 1) / block_size;
+
+    cudaStream_t stream = ctx.stream();
+
+    switch (x->type) {
+        case GGML_TYPE_F32: {
+            snake_kernel<<<grid_size, block_size, 0, stream>>>(
+                (const float *)x->data, a_d, inv_b_d, (float *)dst->data, total, T_len_fastdiv);
+        } break;
+        case GGML_TYPE_F16: {
+            snake_kernel<<<grid_size, block_size, 0, stream>>>(
+                (const half *)x->data, a_d, inv_b_d, (half *)dst->data, total, T_len_fastdiv);
+        } break;
+        case GGML_TYPE_BF16: {
+            snake_kernel<<<grid_size, block_size, 0, stream>>>(
+                (const nv_bfloat16 *)x->data, a_d, inv_b_d, (nv_bfloat16 *)dst->data, total, T_len_fastdiv);
+        } break;
+        default:
+            GGML_ABORT("snake: unsupported type");
+    }
+}
+
+// Fusion entry: caller supplies x/a/inv_b explicitly from the matched
+// mul -> sin -> sqr -> mul -> add pattern. The dst is the trailing add output.
+void ggml_cuda_op_snake_fused(ggml_backend_cuda_context & ctx,
+                              const ggml_tensor * x,
+                              const ggml_tensor * a,
+                              const ggml_tensor * inv_b,
+                              ggml_tensor *       dst) {
+    launch_snake(ctx, x, a, inv_b, dst);
+}

data/ext/sources/ggml/src/ggml-cuda/snake.cuh

@@ -0,0 +1,8 @@
+#include "common.cuh"
+
+// Fusion entry point. Caller supplies x/a/inv_b explicitly.
+void ggml_cuda_op_snake_fused(ggml_backend_cuda_context & ctx,
+                              const ggml_tensor * x,
+                              const ggml_tensor * a,
+                              const ggml_tensor * inv_b,
+                              ggml_tensor *       dst);

data/ext/sources/ggml/src/ggml-cuda/softcap.cu

@@ -1,18 +1,21 @@
 #include "softcap.cuh"
 
 static __global__ void softcap_f32(const float * x, float * dst, const float scale, const float softcap, const int k) {
+    ggml_cuda_pdl_lc();
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
+    ggml_cuda_pdl_sync();
     dst[i] = tanhf(scale * x[i]) * softcap;
 }
 
 static void softcap_f32_cuda(const float * x, float * dst, const float scale, const float softcap, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SOFTCAP_BLOCK_SIZE - 1) / CUDA_SOFTCAP_BLOCK_SIZE;
-    softcap_f32<<<num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream>>>(x, dst, scale, softcap, k);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, CUDA_SOFTCAP_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(softcap_f32, launch_params, x, dst, scale, softcap, k);
 }
 
 // fused GGML_OP_SCALE + GGML_UNARY_OP_TANH + GGML_OP_SCALE

data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu

@@ -1,11 +1,18 @@
+#include "common.cuh"
 #include "ssm-conv.cuh"
 #include "unary.cuh"
 
 template <bool apply_silu, size_t split_d_inner, size_t d_conv>
-static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+static __global__ void ssm_conv_f32(const float * src0_ptr, const float * src1_ptr,
+                                    const float * bias_ptr,
                                     const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
-                                    float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
+                                    float * dst_ptr, const int dst_nb0, const int dst_nb1, const int dst_nb2,
                                     const int64_t n_t) {
+    ggml_cuda_pdl_lc();
+    const float * GGML_CUDA_RESTRICT src0 = src0_ptr;
+    const float * GGML_CUDA_RESTRICT src1 = src1_ptr;
+    const float * GGML_CUDA_RESTRICT bias = bias_ptr;
+    float       * GGML_CUDA_RESTRICT dst  = dst_ptr;
     GGML_UNUSED(src0_nb0);
     const int tid  = threadIdx.x;
     const int bidx = blockIdx.x;
@@ -22,11 +29,14 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
     float x[d_conv] = { 0.0f };
     float w[d_conv] = { 0.0f };
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (size_t j = 0; j < d_conv; j++) {
         w[j] = w_block[tid * stride_w + j];
     }
 
+    float b = bias != nullptr ? bias[bidy * split_d_inner + tid] : 0.0f;
+
     for (int64_t i = 0; i < n_t; i++) {
         float sumf = 0.0f;
 
@@ -42,12 +52,14 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
         for (size_t j = 0; j < d_conv; j++) {
             sumf += x[(i + j) % d_conv] * w[j];
         }
+        sumf += b;
         y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
 template <bool apply_silu, size_t split_d_inner, size_t d_conv, int64_t split_n_t>
 static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+                                               const float * __restrict__ bias,
                                                const int src0_nb0, const int src0_nb1, const int src0_nb2,
                                                const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
                                                const int dst_nb1, const int dst_nb2, const int64_t n_t) {
@@ -97,6 +109,8 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
         w[j] = w_block[tid * stride_w + j];
     }
 
+    float b = bias != nullptr ? bias[bidy * split_d_inner + tid] : 0.0f;
+
     // Compute from shared memory
     for (int64_t i = 0; i < local_n_t; i++) {
         float sumf = 0.0f;
@@ -104,12 +118,13 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
         for (size_t j = 0; j < d_conv; j++) {
             sumf += smem[tid * n_cols + i + j] * w[j];
         }
+        sumf += b;
         y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
 template <bool apply_silu>
-static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
+static void ssm_conv_f32_cuda(const float * src0, const float * src1, const float * bias, const int src0_nb0, const int src0_nb1,
                               const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
                               const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t,
                               const int64_t n_s, cudaStream_t stream) {
@@ -120,30 +135,40 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         constexpr int kNC = decltype(NC)::value;
         if (n_t <= 32) {
             const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
-            ssm_conv_f32<apply_silu, threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
-                                                                       dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_conv_f32<apply_silu, threads, kNC>, launch_params, src0, src1, bias, src0_nb0, src0_nb1,
+                                                                        src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
             const int64_t split_n_t = 32;
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
             const size_t  smem_size = threads * (kNC - 1 + split_n_t) * sizeof(float);
             ssm_conv_long_token_f32<apply_silu, threads, kNC, split_n_t><<<blocks, threads, smem_size, stream>>>(
-                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+                src0, src1, bias, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         }
     };
 
     switch (nc) {
-        case 3: launch_kernel(std::integral_constant<int, 3>{}); break;
-        case 4: launch_kernel(std::integral_constant<int, 4>{}); break;
-        case 9: launch_kernel(std::integral_constant<int, 9>{}); break;
-        default: GGML_ABORT("Only support kernel sizes 3, 4, 9 right now.");
+        case 3:  launch_kernel(std::integral_constant<int, 3 >{}); break;
+        case 4:  launch_kernel(std::integral_constant<int, 4 >{}); break;
+        case 5:  launch_kernel(std::integral_constant<int, 5 >{}); break;
+        case 9:  launch_kernel(std::integral_constant<int, 9 >{}); break;
+        case 15: launch_kernel(std::integral_constant<int, 15>{}); break;
+        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9, 15 right now.");
     }
 }
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst) {
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * bias_add_node, ggml_tensor * silu_dst) {
     const struct ggml_tensor * src0 = dst->src[0];  // conv_x
     const struct ggml_tensor * src1 = dst->src[1];  // conv1d.weight
+    const bool fuse_bias = bias_add_node != nullptr;
     const bool fuse_silu = silu_dst != nullptr;
 
+    // bias always comes with silu.
+    GGML_ASSERT(!fuse_bias || fuse_silu);
+
+    // The bias (when fused) is the non-conv operand of the ADD node.
+    const struct ggml_tensor * bias = fuse_bias ? (bias_add_node->src[0] == dst ? bias_add_node->src[1] : bias_add_node->src[0]) : nullptr;
+
     // When fusing, write to silu_dst (the node downstream references).
     const struct ggml_tensor * out = fuse_silu ? silu_dst : dst;
 
@@ -159,16 +184,23 @@ void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, g
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
+    const float * bias_d = fuse_bias ? (const float *) bias->data : nullptr;
     float *       dst_d  = (float *) out->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(out->type == GGML_TYPE_F32);
+    if (fuse_bias) {
+        GGML_ASSERT(bias->type == GGML_TYPE_F32);
+        GGML_ASSERT(ggml_is_contiguous(bias));
+        GGML_ASSERT(ggml_nelements(bias) == nr);
+    }
+
     if (fuse_silu) {
-        ssm_conv_f32_cuda<true>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+        ssm_conv_f32_cuda<true>(src0_d, src1_d, bias_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
                           out->nb[2], nc, nr, n_t, n_s, stream);
     } else {
-        ssm_conv_f32_cuda<false>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+        ssm_conv_f32_cuda<false>(src0_d, src1_d, bias_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
                           out->nb[2], nc, nr, n_t, n_s, stream);
     }
 }

data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh

@@ -1,3 +1,3 @@
 #include "common.cuh"
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst = nullptr);
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * bias_add_node = nullptr, ggml_tensor * silu_dst = nullptr);

data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu

@@ -17,15 +17,24 @@ using namespace cub;
 #endif // __clang__
 template <size_t splitD, size_t N, size_t L_template>
 __global__ void __launch_bounds__(splitD, 1)
-    ssm_scan_f32(const float *__restrict__ src0, const float *__restrict__ src1, const float *__restrict__ src2,
-                 const float *__restrict__ src3, const float *__restrict__ src4, const float *__restrict__ src5,
-                 const int32_t * __restrict__ src6, float * __restrict__ dst,
+    ssm_scan_f32(const float * src0_ptr, const float * src1_ptr, const float * src2_ptr,
+                 const float * src3_ptr, const float * src4_ptr, const float * src5_ptr,
+                 const int32_t * src6_ptr, float * dst_ptr,
                  const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
                  const int src2_nb1, const int src2_nb2, const int src3_nb1,
                  const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
                  const int64_t s_off, const int64_t d_inner, const int64_t L_param)
 {
+    const float   * GGML_CUDA_RESTRICT src0 = src0_ptr;
+    const float   * GGML_CUDA_RESTRICT src1 = src1_ptr;
+    const float   * GGML_CUDA_RESTRICT src2 = src2_ptr;
+    const float   * GGML_CUDA_RESTRICT src3 = src3_ptr;
+    const float   * GGML_CUDA_RESTRICT src4 = src4_ptr;
+    const float   * GGML_CUDA_RESTRICT src5 = src5_ptr;
+    const int32_t * GGML_CUDA_RESTRICT src6 = src6_ptr;
+    float         * GGML_CUDA_RESTRICT dst  = dst_ptr;
     const size_t L = L_template == 0 ? L_param : L_template;
+    ggml_cuda_pdl_sync();
     const float *s0_block = (const float *)((const char *)src0 + src6[blockIdx.x] * src0_nb3 + blockIdx.y * splitD * src0_nb2);
     const float *x_block = (const float *)((const char *)src1 + (blockIdx.x * src1_nb3) + blockIdx.y * splitD * sizeof(float));
     const float *dt_block = (const float *)((const char *)src2 + (blockIdx.x * src2_nb2) + blockIdx.y * splitD * sizeof(float));
@@ -58,6 +67,7 @@ __global__ void __launch_bounds__(splitD, 1)
     __shared__ CubTempStorage cub_temp_storage;
 
     BlockLoad(cub_temp_storage.load_temp).Load(A_block, regA);
+    __syncthreads();
     BlockLoad(cub_temp_storage.load_temp).Load(s0_block, regs0);
 #else
     const int stride_s0 = src0_nb2 / sizeof(float);
@@ -96,6 +106,7 @@ __global__ void __launch_bounds__(splitD, 1)
             regs0[n] = state;
         }
         y_block[i * stride_y + threadIdx.x] = sumf;
+        __syncthreads();
     }
 
 #ifdef USE_CUB
@@ -117,13 +128,21 @@ __global__ void __launch_bounds__(splitD, 1)
 template <int c_factor, int d_state>
 __global__ void __launch_bounds__(d_state, 1)
     ssm_scan_f32_group(
-        const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
-        const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
-        const int32_t * __restrict__ src6, float * __restrict__ dst,
+        const float * src0_ptr, const float * src1_ptr, const float * src2_ptr,
+        const float * src3_ptr, const float * src4_ptr, const float * src5_ptr,
+        const int32_t * src6_ptr, float * dst_ptr,
         const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
         const int src2_nb1, const int src2_nb2, const int src3_nb1,
         const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
         const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
+    const float   * GGML_CUDA_RESTRICT src0 = src0_ptr;
+    const float   * GGML_CUDA_RESTRICT src1 = src1_ptr;
+    const float   * GGML_CUDA_RESTRICT src2 = src2_ptr;
+    const float   * GGML_CUDA_RESTRICT src3 = src3_ptr;
+    const float   * GGML_CUDA_RESTRICT src4 = src4_ptr;
+    const float   * GGML_CUDA_RESTRICT src5 = src5_ptr;
+    const int32_t * GGML_CUDA_RESTRICT src6 = src6_ptr;
+    float         * GGML_CUDA_RESTRICT dst  = dst_ptr;
 
     const int warp     = threadIdx.x / WARP_SIZE;
     const int lane     = threadIdx.x % WARP_SIZE;
@@ -135,6 +154,7 @@ __global__ void __launch_bounds__(d_state, 1)
 
     const int group_off = (head_idx / (n_head / n_group)) * d_state * sizeof(float);
 
+    ggml_cuda_pdl_sync();
     // TODO: refactor strides to be in elements/floats instead of bytes to be cleaner and consistent with the rest of the codebase
     const float * s0_warp = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
     const float * x_warp  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + (warp_idx * sizeof(float)));
@@ -206,7 +226,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
             constexpr int num_warps = threads/WARP_SIZE;
 
             const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
-            ssm_scan_f32_group<128/WARP_SIZE, 128><<<blocks, threads, 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_scan_f32_group<128/WARP_SIZE, 128>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                     src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
                     src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
@@ -215,7 +236,8 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
             constexpr int num_warps = threads/WARP_SIZE;
 
             const dim3 blocks((n_head * head_dim + (num_warps - 1)) / num_warps, n_seq, 1);
-            ssm_scan_f32_group<256/WARP_SIZE, 256><<<blocks, threads, 0, stream>>>(
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
+            ggml_cuda_kernel_launch(ssm_scan_f32_group<256/WARP_SIZE, 256>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                     src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
                     src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
@@ -229,60 +251,60 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
         GGML_ASSERT(head_dim == 1);
         GGML_ASSERT(n_group == 1);
         const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
-        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
         if (d_state == 16) {
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks, threads, 0, stream);
             switch (n_tok)
             {
             case 1:
-                ssm_scan_f32<threads, 16, 1><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 1>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 2:
-                ssm_scan_f32<threads, 16, 2><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 2>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 3:
-                ssm_scan_f32<threads, 16, 3><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 3>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 4:
-                ssm_scan_f32<threads, 16, 4><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 4>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 5:
-                ssm_scan_f32<threads, 16, 5><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 5>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 6:
-                ssm_scan_f32<threads, 16, 6><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 6>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 7:
-                ssm_scan_f32<threads, 16, 7><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 7>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             case 8:
-                ssm_scan_f32<threads, 16, 8><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 8>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
                 break;
             default:
-                ssm_scan_f32<threads, 16, 0><<<blocks, threads, smem_size, stream>>>(
+                ggml_cuda_kernel_launch(ssm_scan_f32<threads, 16, 0>, launch_params,
                     src0, src1, src2, src3, src4, src5, src6, dst,
                 src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
                 src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);

data/ext/sources/ggml/src/ggml-cuda/sumrows.cu

@@ -7,10 +7,12 @@ void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int
     const dim3 block_nums(nrows, 1, 1);
     if ((nrows / nsm) < 2) {
         const dim3 block_dims(512, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, x, dst, ncols);
     } else {
         const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, x, dst, ncols);
     }
 }
 
@@ -34,10 +36,12 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     if ((nrows / nsm) < 2) {
         // Increase num threads to 512 for small nrows to better hide the latency
         const dim3 block_dims(512, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, src0_d, dst_d, ncols);
     } else {
         // Enough active SMs to hide latency, use smaller blocks to allow better scheduling
         const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
-        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+        ggml_cuda_kernel_launch(reduce_rows_f32</*norm=*/false>, launch_params, src0_d, dst_d, ncols);
     }
 }

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(192, 128, 1, 16);
 DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(320, 256, 1, 32);
 DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu

@@ -7,4 +7,6 @@ DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8);
 DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8);
+DECL_FATTN_MMA_F16_CASE(192, 128, 1, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 1, 8);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2);
+DECL_FATTN_MMA_F16_CASE(512, 512, 16, 2);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 16, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(192, 128, 2, 16);
 DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(320, 256, 2, 32);
 DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32);