diff --git a/csrc/gpu/sample_kernels/top_p_sampling_from_probs.cu b/csrc/gpu/sample_kernels/top_p_sampling_reject.cu similarity index 100% rename from csrc/gpu/sample_kernels/top_p_sampling_from_probs.cu rename to csrc/gpu/sample_kernels/top_p_sampling_reject.cu diff --git a/csrc/setup_cuda.py b/csrc/setup_cuda.py index 269a37acded3..a2e75fb465f5 100644 --- a/csrc/setup_cuda.py +++ b/csrc/setup_cuda.py @@ -107,7 +107,7 @@ def get_gencode_flags(): "./gpu/dequant_int8.cu", "./gpu/flash_attn_bwd.cc", "./gpu/tune_cublaslt_gemm.cu", - "./gpu/sample_kernels/top_p_sampling_from_probs.cu", + "./gpu/sample_kernels/top_p_sampling_reject", ] cutlass_dir = "third_party/cutlass"