Distributed SWE bench by sharding SWE tasks across cpus for faster runtime - #458

Closed

leopck wants to merge 9 commits into

mlcommons:mainfrom

leopck:swe-dist-6-fleet-scorer

Commits on Aug 20, 2026

feat(swe-bench): content-addressed unit plan + durable mkdir-atomic work queue
leopck
committed
e1dbdb5
feat(swe-bench): all-or-nothing merge gate scoped to exactly one run id
leopck
committed
d53af86
feat(swe-bench): conservative claim reaper + PID-only memory guard
leopck
committed
76319c3
feat(swe-bench): eval-phase infra-vs-genuine error classifier
leopck
committed
b56729a
feat(swe-bench): pre-dispatch gates that must prove their own scale
leopck
committed
c57b089
fix(swe-bench): fingerprint endpoint identity, not the time of asking
leopck
committed
79b866c
feat(swe-bench): SWEBenchFleetScorer - fan out units across a service fleet
leopck
committed
e5003ec
fix(swe-bench): validate model_params before reading generation settings
leopck
committed
153f9ba
feat(swe-bench): bind each unit to an endpoint by shard index
leopck
committed
854f854