CUDA: faster tile FA (Pascal/AMD), headsize 256 (#15769)

2025-09-07 00:26:28 +02:00
parent c4df49a42d
commit 79bc429262
7 changed files with 604 additions and 769 deletions
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst);