forked from mooncake-track/Mooncake
77ae1034a6
During Elastic EP rank recovery, the recovered rank must re-run expensive initialization (especially CUDA graph capture). Previously, this required the recovered and healthy ranks to enter the same communication phase, which paused healthy-rank inference and increased recovery disruption. This change enables **asynchronous recovered-rank initialization** in Mooncake PG: recovered ranks can initialize in isolation first, then join the live process group after local recovery work is finished. |
||
|---|---|---|
| .. | ||
| test_hicache_storage_mooncake_backend.py | ||
| test_moe_mooncake.py | ||
| toy_proxy_server.py | ||