[AMD][V4.1] Fix wo_b input flatten for packed fp8 activations on gfx950 (#42796)

This commit is contained in:
Xinyi Song
2026-10-06 21:35:51 -07:00
committed by GitHub
parent 9e26600f3a
commit 168458fd90
+1 -1
View File
@@ -2621,7 +2621,7 @@ class MQALayer(MqaAttentionBase):
and not should_skip_mlp_all_reduce()
)
o, _ = self.wo_b(
o if isinstance(o, Mxfp8SwizzledInput) else o.flatten(1),
o.flatten(1) if isinstance(o, torch.Tensor) else o,
skip_all_reduce=defer_all_reduce,
)
if defer_all_reduce: