Merge branch 'develop_v2.5' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine

transformer_engine/pytorch/cpp_extensions/gemm.py

@@ -570,7 +570,7 @@ def general_grouped_gemm(
                 dw.view(-1, dw.size(-1)),
                 num_gemms,
                 None,
-                TE_DType[out_dtype],
+                out_dtype,
                 None,
                 bias_dtype,
                 gelu,

transformer_engine/pytorch/module/base.py

@@ -1183,7 +1183,7 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
             ):
                 grad_bias = grad_output.dequantize().view(-1, grad_output.shape[-1]).sum(dim=0)
             else:
-                if isinstance(quantizer, Float8BlockQuantizer):
+                if isinstance(quantizer, Float8BlockQuantizer) or (isinstance(quantizer, Float8CurrentScalingQuantizer) and IS_HIP_EXTENSION):
                     # unfuse bgrad for now until cast_transpose + dgrad calculation is ready for Float8BlockQuantizer.
                     grad_bias = grad_output.view(-1, grad_output.shape[-1]).sum(dim=0)
                 else: