Skip to content

[VLLM] Allows for max tokens to be set in model config file #547

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 11 commits into from
Feb 12, 2025
Merged

[VLLM] Allows for max tokens to be set in model config file #547

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
a19e07c
commit
NathanHB Feb 4, 2025
8e21cd5
commit
NathanHB Feb 4, 2025
2802744
Update src/lighteval/main_vllm.py
NathanHB Feb 4, 2025
b4c2d77
commit
NathanHB Feb 5, 2025
147211c
Merge branch 'main' into nathan-fix-vllm-from-file
NathanHB Feb 6, 2025
6ed696d
change doc
NathanHB Feb 6, 2025
e75bb9e
Merge branch 'nathan-fix-vllm-from-file' of github.com:huggingface/li…
NathanHB Feb 6, 2025
c75869b
change doc
NathanHB Feb 6, 2025
952589e
change doc
NathanHB Feb 6, 2025
1a63efe
allow max new token to be set in model config file
NathanHB Feb 10, 2025
87bdf84
Merge remote-tracking branch 'origin/main' into nathan-fix-vllm-from-…
NathanHB Feb 10, 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
2 changes: 1 addition & 1 deletion examples/model_configs/vllm_model_config.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,5 +10,5 @@ model:
top_k: -1
min_p: 0.0
top_p: 0.9
max_new_tokens: 100
max_new_tokens: 256
stop_tokens: ["<EOS>", "<PAD>"]
4 changes: 3 additions & 1 deletion src/lighteval/models/vllm/vllm_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,7 +316,9 @@ def _generate(
sampling_params = self.sampling_params.clone() or SamplingParams()
if generate:
sampling_params.n = num_samples
sampling_params.max_tokens = max_new_tokens
sampling_params.max_tokens = (
max_new_tokens if sampling_params.max_tokens is None else sampling_params.max_tokens
)
Comment on lines +319 to +321
Copy link

Choose a reason for hiding this comment

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

vLLM has max_tokens as Optional[int] but defaulting to 16 here
That means whenever sampling_params is created, it assumes the value 16 and hence this sampling_params.max_tokens ends up being always equal to 16
Then the lighteval benchmark goes on to warn that the output is not in the Gold Format ...

sampling_params.stop = stop_tokens
sampling_params.logprobs = 1 if returns_logits else 0

Expand Down