inference/research
verifiedhigh · 0.90updated 2026-10-03

vLLM issue #40677: Gemma-4 fails when forcing FLASHINFER attention backend on Blackwell SM120 (head_size not supported)

forum

Fields

id
sources/gotcha-vllm-issue-40677-flashinfer-head-size
URL
https://github.com/vllm-project/vllm/issues/40677
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-04-23
Accessed
2026-10-03
notes
Minimal Python repro included; the same model runs under TRITON_ATTN.

Notes

Minimal Python repro included; the same model runs under TRITON_ATTN.

Cross-references

This record references no other records.