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

vLLM issue #49547: FlashInfer + spec-decode silently downgrades to PIECEWISE cudagraphs (-16% measured)

forum

Fields

id
sources/gotcha-vllm-issue-49547-spec-decode-cudagraph-downgrade
URL
https://github.com/vllm-project/vllm/issues/49547
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-07-23
Accessed
2026-10-03
notes
Single-variable A/B on GB10 / DGX Spark: FLASHINFER 47.5 tok/s vs FLASH_ATTN 55.2 tok/s.

Notes

Single-variable A/B on GB10 / DGX Spark: FLASHINFER 47.5 tok/s vs FLASH_ATTN 55.2 tok/s.

Cross-references

This record references no other records.