Skip to content

[Bugfix] Fix wrong CLI defaults for dynamic SchedulerConfig fields #28872

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
merged 2 commits into from
Nov 18, 2025
Merged

[Bugfix] Fix wrong CLI defaults for dynamic SchedulerConfig fields #28872

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
41dcf4d
[Bugfix] Fix wrong CLI defaults for dynamic `SchedulerConfig` fields
DarkLight1337 Nov 17, 2025
5bcc686
Fix
DarkLight1337 Nov 18, 2025
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
18 changes: 15 additions & 3 deletions vllm/engine/arg_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1046,10 +1046,18 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
description=SchedulerConfig.__doc__,
)
scheduler_group.add_argument(
"--max-num-batched-tokens", **scheduler_kwargs["max_num_batched_tokens"]
"--max-num-batched-tokens",
**{
**scheduler_kwargs["max_num_batched_tokens"],
"default": None,
},
)
scheduler_group.add_argument(
"--max-num-seqs", **scheduler_kwargs["max_num_seqs"]
"--max-num-seqs",
**{
**scheduler_kwargs["max_num_seqs"],
"default": None,
},
Comment on lines +1049 to +1060
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

While you're fixing the CLI defaults for dynamic SchedulerConfig fields, it seems like async_scheduling might be another case that needs a similar change.

Currently, EngineArgs.async_scheduling defaults to False (from SchedulerConfig.async_scheduling), and the CLI argument also defaults to False. This means the logic in VllmConfig.__post_init__ to dynamically enable async_scheduling (where self.scheduler_config.async_scheduling is None) will never be triggered.

To enable the dynamic default behavior for async_scheduling, you might need to:

  1. Change the default value of async_scheduling in EngineArgs to None.
  2. Update its add_argument call in add_cli_args to set default=None, similar to the other fields in this PR.

This would make its behavior consistent with the other dynamically configured scheduler arguments.

)
scheduler_group.add_argument(
"--max-num-partial-prefills", **scheduler_kwargs["max_num_partial_prefills"]
Expand All @@ -1071,7 +1079,11 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
"--scheduling-policy", **scheduler_kwargs["policy"]
)
scheduler_group.add_argument(
"--enable-chunked-prefill", **scheduler_kwargs["enable_chunked_prefill"]
"--enable-chunked-prefill",
**{
**scheduler_kwargs["enable_chunked_prefill"],
"default": None,
},
)
scheduler_group.add_argument(
"--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"]
Expand Down
Loading