Skip to content

Fallback when sampling failed #678

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
Ying1123 merged 2 commits into from
Jul 20, 2024
Merged

Fallback when sampling failed #678

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
11d3369
fallback when sampling failed
ispobock Jul 20, 2024
87bdf79
fill nan
ispobock Jul 20, 2024
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
16 changes: 7 additions & 9 deletions python/sglang/srt/managers/controller/infer_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -668,18 +668,16 @@ def sample(self, logits: torch.Tensor):

max_top_k_round, batch_size = 32, probs.shape[0]
uniform_samples = torch.rand((max_top_k_round, batch_size), device=probs.device)
batch_next_token_ids, _ = top_k_top_p_sampling_from_probs(
batch_next_token_ids, success = top_k_top_p_sampling_from_probs(
probs, uniform_samples, self.top_ks, self.top_ps
)

# FIXME: this is a temporary fix for the illegal token ids
illegal_mask = torch.logical_or(
batch_next_token_ids < 0, batch_next_token_ids >= probs.shape[-1]
)
if torch.any(illegal_mask):
warnings.warn("Illegal sampled token ids")
probs = probs.masked_fill(torch.isnan(probs), 0.0)
batch_next_token_ids = torch.argmax(probs, dim=-1)
if torch.any(~success):
warnings.warn("Sampling failed, fallback to top_k=1 strategy")
argmax_ids = torch.argmax(probs, dim=-1)

@merrymercy merrymercy Jul 20, 2024

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Do we need to make sure probs is not NaN before calling argmax?

@ispobock ispobock Jul 20, 2024

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

Yes, we should make sure it.

batch_next_token_ids = torch.where(
success, batch_next_token_ids, argmax_ids
)

if has_regex:
batch_next_token_ids_cpu = batch_next_token_ids.cpu().numpy()
Expand Down