Live Feed/vLLM/Fact Record
vLLM logo
vLLM
feature 96% Confidence Gate October 17, 2026

v0.32.0: [Bugfix] Fix NaN output of the Mamba2 chunk scan kernel on sm90 (#61718)

This update resolves a numerical stability issue causing NaN outputs in the Mamba2 chunk scan kernel. The fix specifically targets hardware utilizing the sm90 architecture, such as NVIDIA H100 GPUs.

Verified State Diff

Comparison Mode:
- Previous State
Mamba2 chunk scan kernels produced NaN outputs when executed on sm90 hardware architectures.
+ Verified New State
Mamba2 chunk scan kernels produce numerically stable outputs on sm90 hardware architectures.

Impact & Verification Analysis

WHO IS AFFECTED

Developers and researchers deploying Mamba2 models on NVIDIA Hopper (sm90) GPUs.

WHY IT MATTERS

Ensures the reliability and correctness of state space model inference on modern enterprise-grade hardware, preventing silent failures in production pipelines.

Full Fact Overview

The Mamba2 architecture relies on a specialized chunk scan kernel to handle state space model operations efficiently. On sm90 (Hopper) architectures, the previous implementation of this kernel encountered precision errors leading to NaN (Not a Number) values during computation. This patch corrects the underlying logic in the kernel execution path to ensure numerical consistency on high-performance NVIDIA hardware.

Multi-Source Evidence Chain (1)

v0.32.0: [Bugfix] Fix NaN output of the Mamba2 chunk scan kernel on sm90 (#61718)vLLM
TRACKED ENTITY
Explore all historical vLLM changes
View vLLM Hub ➔