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

vLLM issue #43559: accuracy drops ~20% with --enable-prefix-caching together with MTP speculative decoding

forum

Fields

id
sources/gotcha-vllm-issue-43559-mtp-prefix-caching-accuracy
URL
https://github.com/vllm-project/vllm/issues/43559
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-05-25
Accessed
2026-10-03
notes
Four-configuration ablation table isolating the interaction: prefix caching alone and MTP alone are both fine.

Notes

Four-configuration ablation table isolating the interaction: prefix caching alone and MTP alone are both fine.

Cross-references

This record references no other records.