Skip to content

Revert "[rollout] feat: use dummy load_format when init AsyncServer" #3207

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
vermouth1992 merged 1 commit into from
Aug 25, 2025
Merged

Revert "[rollout] feat: use dummy load_format when init AsyncServer" #3207

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion verl/workers/rollout/vllm_rollout/vllm_async_server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -276,7 +276,7 @@ async def init_engine(self):
skip_tokenizer_init=False,
max_model_len=self.max_model_len,
max_num_seqs=config.max_num_seqs,
load_format=config.load_format,
load_format="auto",
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Aug 25, 2025

Choose a reason for hiding this comment

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

high

This change reverts the load_format to a hardcoded "auto". While this fixes the issue from the original commit where an invalid value might be passed, it creates an inconsistency with the synchronous vLLMRollout implementation and disables a potentially important feature for AsyncvLLMServer.

The synchronous counterpart in verl/workers/rollout/vllm_rollout/vllm_rollout_spmd.py handles dummy load formats correctly. By reverting to "auto", the async server will ignore the load_format from the configuration, which can lead to unexpected behavior, especially when external weight loading is required.

To maintain consistency and support for dummy loading in async mode, consider applying the same logic as in the sync implementation.

Suggested change
load_format="auto",
load_format="dummy" if config.load_format.startswith("dummy") else config.load_format,

disable_log_stats=config.disable_log_stats,
max_num_batched_tokens=max_num_batched_tokens,
enable_chunked_prefill=config.enable_chunked_prefill,
Expand Down
Loading