---
title: '49909'
pageLayout: default
slug: thoughts/inference/vllm-prs/49909
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/49909.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#49909](https://github.com/aarnphm/aarnphm.github.io/issues/49909) \[ROCm\] Use backend-default dot precision for ReplaySSM

[vllm-project/vllm#49909](https://github.com/vllm-project/vllm/pull/49909) · AndreasKaratzas · MERGED · created 2026-07-26T22:19:18Z · merged 2026-07-27T10:11:07Z

- Let ROCm Triton select its native dot precision instead of passing the unsupported CUDA-only tf32x3 value.
- Keep tf32x3 on CUDA and move the ROCm Mamba kernel group from MI250 to MI300 so the default ROCm path is exercised.

<https://buildkite.com/vllm/amd-ci/builds/11254/list?sid=019f966b-1e70-4263-a40f-36dcf55d9601&tab=output>

---

## comments

