Skip to content

Fix sampling #674

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

Closed
merrymercy wants to merge 1 commit into from
Closed

Fix sampling #674

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
35 changes: 16 additions & 19 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 @@ -665,27 +665,24 @@ def sample(self, logits: torch.Tensor):

# TODO(lmzheng): apply penalty
probs = torch.softmax(logits, dim=-1)
try:
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(
probs, uniform_samples, self.top_ks, self.top_ps
)
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(
probs, uniform_samples, self.top_ks, self.top_ps
)

# FIXME: This is a temporary fix for the illegal token ids in sampling.
illegal_mask = (
batch_next_token_ids < 0 or batch_next_token_ids >= probs.shape[-1]
# FIXME: This is a temporary fix for the illegal token ids in sampling.
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 token ids in sampling.")
batch_next_token_ids = torch.where(
illegal_mask, torch.argmax(probs, dim=-1), batch_next_token_ids
)
if torch.any(illegal_mask):
warnings.warn("Illegal token ids in sampling.")
batch_next_token_ids = torch.where(
illegal_mask, torch.argmax(probs, dim=-1), batch_next_token_ids
)
except RuntimeError as e:
warnings.warn(f"Ignore errors in sampling: {e}")
batch_next_token_ids = torch.argmax(probs, dim=-1)

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