Starstruk Logo MarkStarstruk

v0.18.1

vllm-project/vllmv0.18.1Mar 31, 2026by khluu

AI Summary

Patch release addressing specific issues including reverting SM100 MLA prefill backend, fixing mock.patch resolution failure for Python <= 3.10, and addressing DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8.

Key Highlights

  • Change default SM100 MLA prefill backend back to TRT-LLM
  • Fix mock.patch resolution failure for standalone_compile.FakeTensorMode on Python <= 3.10
  • Disable monolithic TRTLLM MoE for Renormalize routing
  • Pre-download missing FlashInfer headers in Docker build
  • Fix DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8 on Blackwell

New Features

  • SM100 MLA prefill backend configuration adjustment
  • Docker build enhancement for FlashInfer headers

Full Release Notes

This is a patch release on top of v0.18.0 to address a few issues:
- Change default SM100 MLA prefill backend back to TRT-LLM (#38562)
- Fix mock.patch resolution failure for standalone_compile.FakeTensorMode on Python <= 3.10 (#37158)
- Disable monolithic TRTLLM MoE for Renormalize routing #37605
- Pre-download missing FlashInfer headers in Docker build #38391
- Fix DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8 on Blackwell (#38083)