Merge branch 'develop_v2.4' into 'main'

[DCU] fix

See merge request dcutoolkit/deeplearing/TransformerEngine!24

tests/pytorch/references/blockwise_fp8_gemm_reference.py

@@ -7,6 +7,7 @@ from typing import Tuple
 import torch
 import triton
 import triton.language as tl
+from torch.utils.cpp_extension import IS_HIP_EXTENSION
 
 
 @triton.jit
@@ -195,14 +196,17 @@ class CuBLASRefBlockwiseGemm:
                     # Perform qgemm with scaling factors fused in the GEMM
                     # Accumulate should be in float32 format, which aligns with the split_accumulate in FP8 GEMM
                     one = torch.tensor(1.0, dtype=torch.float32, device=qx.device)
-                    y_partial = torch._scaled_mm(
-                        qx_block,
-                        qw_block.t(),
-                        scale_a=one,
-                        scale_b=one,
-                        out_dtype=torch.float32,
-                        use_fast_accum=not use_split_accumulator,
-                    )
+                    if IS_HIP_EXTENSION:
+                        y_partial = torch.mm(qx_block.to(torch.float), qw_block.t().to(torch.float))
+                    else:
+                        y_partial = torch._scaled_mm(
+                            qx_block,
+                            qw_block.t(),
+                            scale_a=one,
+                            scale_b=one,
+                            out_dtype=torch.float32,
+                            use_fast_accum=not use_split_accumulator,
+                        )
 
                     # Accumulate the partial result
                     if is_a_1d_scaled and is_b_1d_scaled: