From ae93d8ae2edec52b5d7f58eebacabd5b1093caf2 Mon Sep 17 00:00:00 2001 From: Adam Mainz Date: Wed, 5 Feb 2025 12:16:40 -0800 Subject: [PATCH] fixing fp8 gemm on amd Summary: adding specific options for fp8 rowwise to tritonbench + fixing an error caused by logic changes in triton 3.2 Reviewed By: nmacchioni Differential Revision: D69197406 --- fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py b/fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py index bf8d52cb1..41471067d 100644 --- a/fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py +++ b/fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py @@ -2991,7 +2991,7 @@ def get_full_non_persistent_tuning_space(): # For now we see better perf with num_stages=0 for all gemm configs we care # But keep this explicit so that we do not forget we may need to set it to # other values in the future - num_stage_range = [0] + num_stage_range = [2] waves_per_eu_range = [0] matrix_instr_nonkdim_range = [16, 32] kpack_range = [1, 2]