whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 2, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu

@@ -7,4 +7,6 @@ DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8);
 DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8);
+DECL_FATTN_MMA_F16_CASE(192, 128, 2, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 2, 8);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2);
 DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2);
 DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2);
 DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2);
+DECL_FATTN_MMA_F16_CASE(512, 512, 32, 2);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu

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

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(192, 128);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(320, 256);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(512, 512);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_F16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_F16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q4_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q4_0);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q4_1);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q4_1);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q5_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q5_0);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q5_1);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q5_1);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_BF16, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_BF16, GGML_TYPE_Q8_0);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_BF16, GGML_TYPE_Q8_0);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu

@@ -0,0 +1,7 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec.cuh"
+
+DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_BF16);
+DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_BF16);

data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_NVFP4);

data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q1_0);

data/ext/sources/ggml/src/ggml-cuda/top-k.cu

@@ -5,6 +5,7 @@
 #    include <cub/cub.cuh>
 #    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2)
 #        define CUB_TOP_K_AVAILABLE
+#        include <cuda/iterator>
 using namespace cub;
 #    endif  // CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2
 #endif      // GGML_CUDA_USE_CUB
@@ -25,14 +26,14 @@ static void top_k_cub(ggml_cuda_pool & pool,
     auto indexes_in = cuda::make_counting_iterator(0);
 
     size_t temp_storage_bytes = 0;
-    DeviceTopK::MaxPairs(nullptr, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst, ncols, k,
-                         env);
+    CUDA_CHECK(DeviceTopK::MaxPairs(nullptr, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst, ncols, k,
+                         env));
 
     ggml_cuda_pool_alloc<uint8_t> temp_storage_alloc(pool, temp_storage_bytes);
     void *                        d_temp_storage = temp_storage_alloc.get();
 
-    DeviceTopK::MaxPairs(d_temp_storage, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst,
-                         ncols, k, env);
+    CUDA_CHECK(DeviceTopK::MaxPairs(d_temp_storage, temp_storage_bytes, src, cuda::discard_iterator(), indexes_in, dst,
+                         ncols, k, env));
 }
 
 #elif defined(GGML_CUDA_USE_CUB)  // CUB_TOP_K_AVAILABLE

data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu

@@ -105,6 +105,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         wt[i] = -INFINITY;
     }
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int i = 0; i < n_experts; i += WARP_SIZE) {
         const int expert  = i + threadIdx.x;
@@ -133,7 +134,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
 
     // selection_wt is only needed when bias is present (selection uses wt + bias)
     // when no bias, we use wt directly for both selection and weight values
-    float selection_wt[has_bias ? experts_per_thread : 1];
+    [[maybe_unused]] float selection_wt[has_bias ? experts_per_thread : 1];
 
     if constexpr (has_bias) {
 #pragma unroll
@@ -161,6 +162,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         output_weights[i] = 0.f;
     }
 
+    ggml_cuda_pdl_lc();
     for (int k = 0; k < n_expert_used; k++) {
         float max_val    = wt[0];
         int   max_expert = threadIdx.x;
@@ -271,51 +273,56 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
     dim3         grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
     dim3         block_dims(WARP_SIZE, rows_per_block, 1);
     cudaStream_t stream = ctx.stream();
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, stream);
 
     switch (n_expert) {
         case 1:
-            topk_moe_cuda<1, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<1, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 2:
-            topk_moe_cuda<2, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<2, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 4:
-            topk_moe_cuda<4, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<4, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 8:
-            topk_moe_cuda<8, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                   clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<8, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 16:
-            topk_moe_cuda<16, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<16, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 32:
-            topk_moe_cuda<32, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<32, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 64:
-            topk_moe_cuda<64, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                    clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<64, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 128:
-            topk_moe_cuda<128, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<128, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 256:
-            topk_moe_cuda<256, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<256, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
+            break;
+        case 288: // StepFun 3.7
+            ggml_cuda_kernel_launch(topk_moe_cuda<288, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 512:
-            topk_moe_cuda<512, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<512, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         case 576:
-            topk_moe_cuda<576, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
-                                                                     clamp_val, scale_val, config);
+            ggml_cuda_kernel_launch(topk_moe_cuda<576, has_bias>, launch_params,
+                logits, weights, ids, bias, n_rows, n_expert_used, clamp_val, scale_val, config);
             break;
         default:
             GGML_ASSERT(false && "fatal error");
@@ -374,8 +381,10 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
                                    const ggml_tensor * weights,
                                    const ggml_tensor * logits,
                                    const ggml_tensor * ids) {
+    // must match an instantiation of launch_topk_moe_cuda: a power of 2 up to 512,
+    // or one of the non-power-of-2 expert counts of supported models
     const int n_expert = ids->nb[1] / ids->nb[0];
-    if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 576) {
+    if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 288 && n_expert != 576) {
         return false;
     }
 

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

@@ -65,6 +65,11 @@ static __device__ __forceinline__ float op_sqr(float x) {
     return x * x;
 }
 
+static __device__ __forceinline__ float op_relu_sqr(float x) {
+    const float r = fmaxf(x, 0.0f);
+    return r * r;
+}
+
 static __device__ __forceinline__ float op_sqrt(float x) {
     return sqrtf(x);
 }
@@ -111,19 +116,22 @@ static __device__ __forceinline__ float op_trunc(float x) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, 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] = (T)op((float)x[i]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
-    unary_op_kernel<op><<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_op_kernel<op, T>, launch_params, x, dst, k);
 }
 
 template <float (*op)(float)>
@@ -253,6 +261,7 @@ void ggml_cuda_op_softplus(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
 template <float (*op)(float), typename T>
 static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
+    ggml_cuda_pdl_lc();
     const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
@@ -263,13 +272,15 @@ static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst,
     const int64_t j0 = (i / n) * o0 + (i % n);
     const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
 
+    ggml_cuda_pdl_sync();
     dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
 }
 
 template <float (*op)(float), typename T>
 static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
     const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
-    unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(unary_gated_op_kernel<op, T>, launch_params, x, g, dst, k, n, o0, o1);
 }
 
 template <float (*op)(float)>
@@ -615,3 +626,21 @@ void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary
             GGML_ABORT("Unsupported unary op for fused unary+mul");
     }
 }
+
+/* fused relu + sqr */
+
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node) {
+    const ggml_tensor * src = relu_node->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src));
+    GGML_ASSERT(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
+    GGML_ASSERT(src->type == sqr_node->type);
+
+    const int k = ggml_nelements(src);
+    if (src->type == GGML_TYPE_F16) {
+        unary_cuda<op_relu_sqr>((const half *)src->data, (half *)sqr_node->data, k, stream);
+    } else {
+        unary_cuda<op_relu_sqr>((const float *)src->data, (float *)sqr_node->data, k, stream);
+    }
+}

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

@@ -91,6 +91,8 @@ void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);
 
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node);
+
 __device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
     return x / (1.0f + expf(-x));
 }

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

@@ -106,6 +106,9 @@ static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
+#define VDR_Q1_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
+#define VDR_Q1_0_Q8_1_MMQ  4  // Q1_0 has 128 bits (4 ints) per block
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -322,6 +325,38 @@ static __device__ __forceinline__ float vec_dot_mxfp4_q8_1(
     return d * sumi;
 }
 
+#define VDR_NVFP4_Q8_1_MMVQ 4
+#define VDR_NVFP4_Q8_1_MMQ  8
+
+static __device__ __forceinline__ float vec_dot_nvfp4_q8_1(
+                                        const void * __restrict__ vbq,
+                                        const block_q8_1 * __restrict__ bq8_1,
+                                        const int32_t & kbx,
+                                        const int32_t & iqs) {
+
+    const block_nvfp4 * bq4 = (const block_nvfp4 *) vbq + kbx;
+    float sum = 0.0f;
+#pragma unroll
+    for (int i = 0; i < VDR_NVFP4_Q8_1_MMVQ/2; i++) {
+        const int32_t iqs0 = iqs + 2*i;
+        const int32_t iqs1 = iqs0 + 1;
+        const int32_t is = iqs0 >> 1;
+        const int2 v0 = get_int_from_table_16(get_int_b4(bq4->qs, iqs0), kvalues_mxfp4);
+        const int2 v1 = get_int_from_table_16(get_int_b4(bq4->qs, iqs1), kvalues_mxfp4);
+        const block_q8_1 * bq8 = bq8_1 + (is >> 1);
+        const int32_t i8 = ((is & 1) << 2);
+
+        int sumi = ggml_cuda_dp4a(v0.x, get_int_b4(bq8->qs, i8 + 0), 0);
+        sumi = ggml_cuda_dp4a(v0.y, get_int_b4(bq8->qs, i8 + 2), sumi);
+        sumi = ggml_cuda_dp4a(v1.x, get_int_b4(bq8->qs, i8 + 1), sumi);
+        sumi = ggml_cuda_dp4a(v1.y, get_int_b4(bq8->qs, i8 + 3), sumi);
+
+        const float d = ggml_cuda_ue4m3_to_fp32(bq4->d[is]) * __low2float(bq8->ds);
+        sum += d * float(sumi);
+    }
+
+    return sum;
+}
 #define VDR_Q2_K_Q8_1_MMVQ 1
 #define VDR_Q2_K_Q8_1_MMQ  4
 
@@ -637,6 +672,51 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 }
 
+static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q1_0 * bq1_0 = (const block_q1_0 *) vbq + kbx;
+
+    // Q1_0: 128 elements with ONE scale
+    // Q8_1: 32 elements per block with individual scales
+    // iqs selects which of the 4 chunks of 32 elements to process (0-3)
+
+    const float d1 = bq1_0->d;
+
+    // Process only the chunk specified by iqs
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+
+    // Load 32 bits (4 bytes) for this chunk from Q1_0
+    const int offset = iqs * 4;
+    const int v = bq1_0->qs[offset + 0] | (bq1_0->qs[offset + 1] << 8) |
+                  (bq1_0->qs[offset + 2] << 16) | (bq1_0->qs[offset + 3] << 24);
+
+    // Unpack 32 bits into 32 signed values (-1 or +1)
+    int vi_bytes[8];
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int shift = j * 4;
+        const int bits4 = (v >> shift) & 0x0F;
+        const int b0 = (bits4 & 0x01) ? 1 : -1;
+        const int b1 = (bits4 & 0x02) ? 1 : -1;
+        const int b2 = (bits4 & 0x04) ? 1 : -1;
+        const int b3 = (bits4 & 0x08) ? 1 : -1;
+        vi_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+    }
+
+    // Compute dot product for this 32-element chunk
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int u = get_int_b4(bq8_1_chunk->qs, j);
+        sumi = ggml_cuda_dp4a(vi_bytes[j], u, sumi);
+    }
+
+    // Apply Q1_0's single scale and this chunk's Q8_1 scale
+    const float d8 = __low2float(bq8_1_chunk->ds);
+    return d1 * d8 * sumi;
+}
+
 static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 

data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h

@@ -6,9 +6,14 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 
-#if CUDART_VERSION >= 12050
+#ifdef GGML_USE_NCCL
+#include <nccl.h>
+#endif // GGML_USE_NCCL
+
+#if CUDART_VERSION >= 11080
 #include <cuda_fp8.h>
-#endif // CUDART_VERSION >= 12050
+#define FP8_AVAILABLE
+#endif // CUDART_VERSION >= 11080
 
 #if CUDART_VERSION >= 12080
 #include <cuda_fp4.h>