[MiniMax-M3][LMCache] Drive ATOM's byte-level KV offload from the vLLM plugin - #2146

Open

XiaobingSuper wants to merge 14 commits into

mainfrom

perzhang/m3-vllm-lmcache-offload

Commits on Sep 6, 2026

feat(vllm-plugin): map vLLM KV registration onto ATOM's KVCacheTensor
zgplvyou
committed
f837220
feat(vllm-plugin): drive ATOM's byte-level LMCache offload from vLLM
zgplvyou
committed
85c700d
fix(vllm-plugin): project every config field the offload path reads
zgplvyou
committed
aca5bfa
fix(vllm-plugin): resolve nested HF configs for the offload namespace
zgplvyou
committed
3972adf
fix(vllm-plugin): map ATOM's four completion sets onto vLLM's two
zgplvyou
committed
17c8d2f
fix(vllm-plugin): only report a save as finished_sending once the request is finished
zgplvyou
committed
c6d5a61
fix(vllm-plugin): carry M3's fp8 KV scales through the offload transfer
zgplvyou
committed
7be1766
fix(vllm-plugin): return the worker's completions to the scheduler half
zgplvyou
committed
ef5d384
fix(vllm-plugin): only promise a load ATOM has agreed to issue
zgplvyou
committed
a1f8787
feat(vllm-plugin): name a promised load that never gets dispatched
zgplvyou
committed
d9eef3e

Commits on Sep 7, 2026