Starstruk Logo MarkStarstruk

v0.20.2

niklasvh/html2canvasv0.20.2May 10, 2026by khluu

AI Summary

A minor patch release focusing on bug fixes for DeepSeek V4, gpt-oss, and Qwen3-VL models.

Key Highlights

  • Fixes DeepSeek V4 sparse attention (MTP=1 hang)
  • Fixes DeepSeek V4 KV cache allocation error
  • Fixes gpt-oss MXFP4 compatibility with torch.compile
  • Fixes Qwen3-VL deepstack boundary check

New Features

  • Bug fixes for DeepSeek V4 sparse attention
  • Bug fixes for KV cache manager
  • Bug fixes for gpt-oss and Qwen3-VL

Full Release Notes

# vLLM v0.20.2

## Highlights
This release features 6 commits from 6 contributors (0 new)!

This is a small patch release with bug fixes for DeepSeek V4, gpt-oss, and Qwen3-VL

### Bug Fixes
* **DeepSeek V4 sparse attention**: Re-enable the persistent topk path on Hopper and ensure the memset kernel runs at CUDA graph capture time regardless of `max_seq_len`, fixing the MTP=1 hang on DeepSeek V4 (#41665, revert of #41605).
* **DeepSeek V4 KV cache**: Fixed a "failure to allocate KV blocks" error in the V1 engine KV cache manager (#41282).
* **gpt-oss MXFP4 + torch.compile**: Plumbed `hidden_dim_unpadded` through the `moe_forward` fake op so MXFP4 works under `torch.compile` on v0.20.x (#42002, backport of #41646).
* **Qwen3-VL**: Removed an invalid deepstack boundary check that could fail under heavy load (#40932).

## Contributors
@ywang96, @zyongye, @stecasta, @wzhao18, @Isotr0py, @khluu