job-manager / docs /HOST_NOTES.md
codemaivanngu's picture
Mirror of GitHub branch eval-core at f20458a38e06550639cf2011634cb9c48551bf66
36bacba
|
Raw History Blame Contribute Delete
571 Bytes

Multi-machine rules (v0.2)

  • HOST = hostname -s; khai bao GPU cua may: ec host --gpus 0,1,3
  • Lock theo (may, gpu): locks/-gpu-.lock => gpu0 cua may A khong chan gpu0 cua may B.
  • 1 job = 1 CELL: (run, ckpt_sha, step, benchmark, eval_seed, gen_seed).
  • Claim: mv pending -> claimed; ghi claimed_by=":"; touch heartbeats/ moi 60s.
  • ec gc thu hoi theo heartbeat (mac dinh 20m), khong theo mtime file job => dung ca khi lech dong ho.
  • Bank key = sha256(ckpt_sha | benchmark | eval_seed | contract_sha | env_fp) => dedup muc ckpt, xuyen may.