karylab_agents/vllm
1
0
Fork 0
forked from Karylab-cklius/vllm
Code Pull Requests 2 Actions 1 Packages Activity

[Bugfix] Fix sequence_parallel_chunk_impl custom op aliasing its input (#44130)

Browse Source 
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Vadim Gimpelson
2026-06-05 23:56:36 +00:00
committed by GitHub
co-authored by Cursor
parent a50e675b0d
commit 4765f0f189
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/models/utils.py
+4 -1
View File
@@ -838,7 +838,10 @@ def sequence_parallel_chunk_impl(x: torch.Tensor) -> torch.Tensor:
chunk = y.shape[0] // tp_size
start = tp_rank * chunk
return torch.narrow(y, 0, start, chunk)
out = torch.narrow(y, 0, start, chunk)
# narrow() returns a view; clone when it aliases the input (no-pad case),
# since a functional custom op must not return a view of an input.
return out.clone() if y is x else out
def sequence_parallel_chunk_impl_fake(x: torch.Tensor) -> torch.Tensor: