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

vLLM issue #53670: EAGLE/MTP prefix-cache last-block drop causes a 1,648-token recompute per hit on hybrid GDN layouts

forum

Fields

id
sources/gotcha-vllm-issue-53670-eagle-last-block-drop
URL
https://github.com/vllm-project/vllm/issues/53670
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-08-25
Accessed
2026-10-03
notes
Prefix hit rate 97.5% -> 86.9-91.7% with any EAGLE speculative_config set; names the conservative all-groups-flagged fallback in kv_cache_coordinator.py.

Notes

Prefix hit rate 97.5% -> 86.9-91.7% with any EAGLE speculative_config set; names the conservative all-groups-flagged fallback in kv_cache_coordinator.py.

Cross-references

This record references no other records.