Files
mscclpp/test/python/ext/ep
Qinghua Zhou e752dbaf97 tests/ep: add NCCL-EP six-metric BW breakdown (send/recv x total/nvl/rdma)
For HT intra/internode benches, compute per-rank avg total_send/rdma_send
and total_recv/rdma_recv token counts (matching NCCL-EP ep_bench
accounting) and print send-side and recv-side BW split into total / nvl
/ rdma columns. Combine reverses send<->recv. Byte-count line mirrors
NCCL-EP's '(per rank avg)' summary so numbers are directly comparable.
2026-04-29 20:44:10 +00:00
..