Skip to content

Make microbatch optimization (DBO) work with general models #37926

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
njhill merged 3 commits into from
Mar 24, 2026
+9 −5
Merged

Make microbatch optimization (DBO) work with general models #37926

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
74abae0
fix dbo
Mar 23, 2026
c16f866
address lint and gemini comments
0xjunhao Mar 23, 2026
1cbe667
Merge branch 'main' into fix-dbo
0xjunhao Mar 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 9 additions & 5 deletions vllm/v1/worker/gpu_ubatch_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -389,16 +389,20 @@ def _slice_model_inputs(
inputs_embeds,
intermediate_tensors,
):
sliced_input_ids = input_ids[tokens_slice]
sliced_input_ids = input_ids[tokens_slice] if input_ids is not None else None
# if we are using mrope. Mrope adds an additional dimension to the
# positions tensor
if positions.ndim == 2:
sliced_positions = positions[:, tokens_slice]
else:
sliced_positions = positions[tokens_slice]
sliced_inputs_embeds = inputs_embeds[tokens_slice] if inputs_embeds else None
sliced_inputs_embeds = (
inputs_embeds[tokens_slice] if inputs_embeds is not None else None
)
sliced_intermediate_tensors = (
intermediate_tensors[tokens_slice] if intermediate_tensors else None
intermediate_tensors[tokens_slice]
if intermediate_tensors is not None
else None
)

return (
Expand Down Expand Up @@ -478,7 +482,7 @@ def __call__(self, *args, **kwargs):
cudagraph_runtime_mode=CUDAGraphMode.NONE,
)
with self.sm_control:
return self._capture_ubatches(ubatch_metadata, self.model)
return self._capture_ubatches(ubatch_metadata, self.runnable)
elif (
num_tokens in self.cudagraphs
and cudagraph_runtime_mode is CUDAGraphMode.FULL
Expand All @@ -504,4 +508,4 @@ def __call__(self, *args, **kwargs):
cudagraph_runtime_mode=CUDAGraphMode.NONE,
)
with self.sm_control:
return self._run_ubatches(ubatch_metadata, self.model)
return self._run_ubatches(ubatch_metadata, self.runnable)
Loading