Qinghua Zhou 63afb25ab3 tests/ep: LL bench combine uses recv_tokens×hidden for payload bytes ...

Each local expert sends one copy per dispatched token back to its owner,
so the bytes actually on the wire during combine match dispatch. The
previous num_tokens×hidden under-counted by ~num_topk×, making combine
BW look artificially low next to dispatch.

2026-04-23 21:53:34 +00:00

..

ext/ep

tests/ep: LL bench combine uses recv_tokens×hidden for payload bytes

2026-04-23 21:53:34 +00:00