[1/2] Refactor DeepGeem requant for FP8 Linear on Blackwell (#13601)

Co-authored-by: fy1214
2025-11-23 16:07:56 -08:00
parent 9054e844ea
commit 4683e244fe
4 changed files with 39 additions and 79 deletions
@@ -222,7 +222,7 @@ class TestDeepGemmBlackwell(CustomTestCase):

        with torch.inference_mode():
            ref_out = native_w8a8_block_fp8_matmul(
-                A_q, B_q, A_s, B_s, block_size, out_dtype
+                A_qu[0], B_qu[0], A_qu[1], B_qu[1], block_size, out_dtype
            )
            out = torch.empty_like(ref_out)
            fp8_gemm_nt(A_qu, B_qu, out)