CollectiveX: graph-replay every capturable EP backend with rank alignment and the chained period - #3437

Open

Oseltamivir wants to merge 18 commits into

mainfrom

cx-graphs-rank-major

Commits on Sep 25, 2026

CollectiveX: graph-replay every capturable EP backend, keeping rank alignment and the chained period
Oseltamivir
committed
e8fc87a
CollectiveX: selectable nccl-ep LL layout (expert-major alongside rank-major); keep MoRI AsyncLL eager
Oseltamivir
committed
a86227f
CollectiveX: keep MoRI eager -- ROCm torch rejects external events inside a graph capture
Oseltamivir
committed
828c7fb
fix: route MI325X CollectiveX sweeps through the live Slurm pool
cquil11
committed
b25ca14
CollectiveX: keep nccl-ep HT eager -- graphed zero-copy HT corrupts intermittently across nodes on x86
Oseltamivir
committed
15796e5
Merge remote-tracking branch 'origin/main' into cx-graphs-rank-major
Oseltamivir
committed
85de08d
Merge remote-tracking branch 'origin/cx-graphs-rank-major' into cx-graphs-rank-major
Oseltamivir
committed
23b0f3f
CollectiveX: address the per-library review of graph replay
Oseltamivir
committed
b0d7acf
CollectiveX: render withheld graph-replay tails as '-' in the step summaries
Oseltamivir
committed
901fbcd
CollectiveX: publish per-pass oracle verdicts so a failure can be placed before or after the measured regimes
Oseltamivir
committed
3f4a146
CollectiveX: barrier the graph replay value check after poisoning; report a non-finite output mismatch as inf
Oseltamivir
committed
858dc2d

Commits on Sep 26, 2026