Skip to content

LLM inference in C/C++

C++ ggml

cuda : disables launch_fattn PDL enrollment due to compiler bug (#23825)

Andreas Kieslinger241cbd41d21f
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index debcb6e544..d650b5fbd0 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -1153,8 +1153,8 @@ void launch_fattn(
GGML_ASSERT(block_dim.x % warp_size == 0);
- const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num, block_dim, nbytes_shared, main_stream);
- ggml_cuda_kernel_launch(fattn_kernel, launch_params,
+ // disabled PDL enrollment for now due to a compiler bug.
+ fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
(const char *) Q->data,
K_data,
V_data,