vLLM#40677: FlashInfer head_size not supported on Blackwell SM120

Fields

id

sources/src5-gotcha-vllm-40677-flashinfer-head-size

URL

https://github.com/vllm-project/vllm/issues/40677

Publisher

vLLM

Kind

forum

Published

2026-04-23

Accessed

2026-10-04

notes

Gemma-4 fails when forcing FLASHINFER attention backend on Blackwell SM120 (RTX PRO 6000). Error: head_size not supported. Gemma-4 config: head_dim=256, global_head_dim=512. Works with TRITON_ATTN. Affects base model, FP8, and NVFP4 checkpoints. Demonstrates hardware-specific kernel compatibility issues.

Notes

Gemma-4 fails when forcing FLASHINFER attention backend on Blackwell SM120 (RTX PRO 6000). Error: head_size not supported. Gemma-4 config: head_dim=256, global_head_dim=512. Works with TRITON_ATTN. Affects base model, FP8, and NVFP4 checkpoints. Demonstrates hardware-specific kernel compatibility issues.

Cross-references

This record references no other records.