Skip to content

Reintroduce tiny fix sampler error when prob is not contiguous #7354

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 1 commit into from
Jun 19, 2025
Merged

Reintroduce tiny fix sampler error when prob is not contiguous #7354

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 python/sglang/srt/layers/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,7 +91,7 @@ def forward(
)
else:
batch_next_token_ids = top_k_top_p_sampling_from_probs(
probs,
probs.contiguous(),
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jun 19, 2025

Choose a reason for hiding this comment

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

medium

Calling .contiguous() here is a good safeguard to ensure the probs tensor meets the memory layout requirements of the top_k_top_p_sampling_from_probs kernel, which is likely a custom C++/CUDA kernel from sgl_kernel that expects contiguous inputs. This prevents potential runtime errors or incorrect behavior if probs becomes non-contiguous due to upstream operations.

To improve code clarity and maintainability, consider adding a brief inline comment explaining why this call is necessary. This helps future developers understand the requirement, especially since .contiguous() can incur a performance overhead if a copy is made.

For example:

# Ensure probs is contiguous, as required by the underlying sgl_kernel
# for the 'flashinfer' backend to prevent potential errors.
probs.contiguous(),

sampling_info.top_ks,
sampling_info.top_ps,
filter_apply_order="joint",
Expand Down
Loading