vLLM-v0 Upgrade #122
Merged
vLLM-v0 Upgrade #122
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Merged
rafapi
reviewed
Jan 21, 2026
rafapi
reviewed
Jan 21, 2026
rafapi
reviewed
Jan 21, 2026
rafapi
reviewed
Jan 21, 2026
|
|
||
| # Run HTTP server | ||
| sock_addr = (args.host or "", args.port) | ||
| sock = create_server_socket(sock_addr) |
Collaborator
There was a problem hiding this comment.
Not sure when this happened, but we are running the http server twice without dropping the previous one. see line 159. we need to remove this line and the one above
Collaborator
Author
There was a problem hiding this comment.
Looks like it's intentional, see the comment in the first run:
# workaround to make sure that we bind the port before the engine is set up.
# This avoids race conditions with ray.
# see https://github.com/vllm-project/vllm/issues/8204
rafapi
reviewed
Jan 21, 2026
|
|
||
| if args.load_as_bf16: | ||
| loading_args["torch_dtype"] = torch.bfloat16 | ||
| loading_args["dtype"] = torch.bfloat16 |
Collaborator
There was a problem hiding this comment.
why this change? has the transformers API changed here?
Collaborator
Author
There was a problem hiding this comment.
torch_dtype became deprecated and prints a warning
rafapi
reviewed
Jan 21, 2026
|
|
||
| logger.info(f"Merge lora checkpoint {lora_model_path}") | ||
| model = lora_load_and_merge(lora_model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True) | ||
| model = lora_load_and_merge(lora_model_path, dtype=torch.bfloat16, low_cpu_mem_usage=True) |
Collaborator
Author
There was a problem hiding this comment.
torch_dtype renamed to dtype
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR upgrades vLLM to a recent version where the V0 engine is not removed, which is
v0.10.0(an intermediate step to fully migrate to V1 in #121)Notable upgrades:
python:
3.11to3.12vllm:
0.8.5.post1to0.10.0torch:
2.6.0to2.7.1transformers:
4.51.1to4.57.6flash-attention:
2.7.4.post1to2.8.3GSPO (blue=
v0.8.5post1, pink/purple=v0.10.0)GRPO (orange=
v0.8.5post1, purple=v0.10.0)Potentially, the latest version to upgrade to is
v0.10.2, but this error occurs from the bundled flash attention in vLLM:Also some v0 features that we use in the code were removed in
>v0.10.0such as multi-step scheduler (it's ok in this case as we don't normally use it)