A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Remove contiguous output req for context parallel MLA (#25414)
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Michael Goin committed
Sep 22, 2025 at 23:26 UTC
78237e43bf3ac70f31b27684dc27d7eca0fa3677
1 file changed
-1
vllm/attention/ops/common.py
-1
@@ -134,6 +134,5 @@ def cp_lse_ag_out_rs(cp_attn_out: torch.Tensor,
134
cp_attn_lse = cp_attn_lse.contiguous()
135
lses = cp_group.all_gather(cp_attn_lse, dim=0).view_as(lses)
136
out, _ = correct_attn_out(cp_attn_out, lses, cp_group.rank_in_group, ctx)
137
- assert out.is_contiguous()
137
out = cp_group.reduce_scatter(out, dim=1)
138
return out