[Bugfix] Fix trtllm fused allreduce+rms_norm for transformers backend (#45307)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2026-06-16 08:34:27 +00:00
committed by GitHub
parent c4fd9794e9
commit ebf3a6d705
@@ -155,6 +155,13 @@ if flashinfer_comm is not None:
scale_factor: torch.Tensor | None = None,
weight_bias: float = 0.0,
) -> None:
# handle transformers backend passing outer batch dim.
if allreduce_in.dim() != 2:
hidden = allreduce_in.shape[-1]
allreduce_in = allreduce_in.view(-1, hidden)
residual = residual.view(-1, hidden)
if norm_out is not None:
norm_out = norm_out.view(-1, hidden)
num_tokens, hidden_size = allreduce_in.shape
element_size = allreduce_in.element_size()
current_tensor_size = num_tokens * hidden_size * element_size