sgl-project/sglang

[Performance] Remove instances of `torch.nonzero()` where appropriate

Open

#9,889 opened on Sep 2, 2025

View on GitHub
 (9 comments) (1 reaction) (0 assignees)Python (6,216 forks)auto 404
good first issueperformance

Repository metrics

Stars
 (28,442 stars)
PR merge metrics
 (Avg merge 2d 1h) (1,000 merged PRs in 30d)

Description

Right now we have:

4 results - 4 files

python/sglang/srt/configs/janus_pro.py:
  478          image_token_mask: torch.Tensor = (input_ids == self.image_id).to(torch.bool)
  479:         image_indices = image_token_mask.nonzero()
  480          input_ids, num_image_tokens = self.add_image_token(

python/sglang/srt/models/gemma3_mm.py:
  211          # Distinguish sequences by position id 0
  212:         start_indices = (positions == 0).cpu().nonzero()
  213          num_seqs = len(start_indices)

python/sglang/srt/models/phi4mm_utils.py:
  83      seq_range = torch.arange(0, x_len).unsqueeze(-1)  # seq_range size: [x_len, 1]
  84:     idx = ((seq_range < end_pad) & (seq_range >= start_pad)).nonzero()[
  85          :, 1

We should explore whether we can update these to reduce the host-device sync. You could profile it before trying to update to see if it really happens, or analyze the call sites deeply and test changes. However, if it doesn't force a sync, then updating it will do nothing and using some more advanced pytorch operations might be even more overhead. So edit the code judiciously. Why? See below.

[!CAUTION] When input is on CUDA, torch.nonzero() causes host-device synchronization.

Contributor guide

Research direction
Analyze each call site of torch.nonzero in the listed files. Test whether synchronization actually occurs using profiling tools like torch.cuda.synchronize or nsys. If it does, replace with alternatives like torch.where, boolean indexing, or torch.nonzero(..., as tuple=True) to avoid host device sync. For example, use `start indices = (positions == 0).nonzero(as tuple=True)[0]` or `image token mask.nonzero(as tuple=True)[0]` or use `torch.where` and `torch.split`. Ensure the changes maintain correctness and do not increase overhead.
Tech stack
pythonpytorch
Domain
backend
Issue type
Performance
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
2
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyTorchCUDA
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
50