Local AI & Edge

vLLM v0.32.0 fixes Mamba2 NaN errors on sm90 hardware

Key Takeaway

The vLLM update resolves an issue where the Mamba2 chunk scan kernel produced NaN outputs on sm90 architecture.

Summary

The vLLM project has released version 0.32.0, which specifically addresses a bug in the Mamba2 chunk scan kernel. This bug caused the kernel to produce NaN (Not a Number) outputs when running on sm90 architecture. The fix ensures numerical stability for Mamba2 models on this specific hardware.

Mamba2 kernel fix details

  • Fixes NaN output issues in the Mamba2 chunk scan kernel.
  • Specifically targets hardware using the sm90 architecture.

Original reference: github.com →