vLLM#57838: RowWise FP8 kernel regression on RDNA4 (gfx1201)

Fields

id

sources/src5-gotcha-vllm-57838-rdna4-fp8-regression

URL

https://github.com/vllm-project/vllm/issues/57838

Publisher

vLLM

Kind

forum

Published

2026-09-20

Accessed

2026-10-04

notes

From v0.28.0, RowWiseTorchFP8ScaledMMLinearKernel is selected on RDNA4 (gfx1201) instead of ChannelWise, costing 5-24% decode throughput. v0.27.1 and earlier correctly reject RDNA4 for RowWise. Fix: restore RDNA4 exclusion. Demonstrates ROCm kernel selection regression.

Notes

From v0.28.0, RowWiseTorchFP8ScaledMMLinearKernel is selected on RDNA4 (gfx1201) instead of ChannelWise, costing 5-24% decode throughput. v0.27.1 and earlier correctly reject RDNA4 for RowWise. Fix: restore RDNA4 exclusion. Demonstrates ROCm kernel selection regression.

Cross-references

This record references no other records.