EMRY-067: emry slurm-wrap helper #111

New issue

Closed

opened 2026-07-02 21:50:22 +00:00 by vxfemboy · 0 comments

vxfemboy commented

2026-07-02 21:50:22 +00:00

(Migrated from github.com)

A CLI that runs the SLURM sidecar pattern (from the runbook) as one command: start the engine, bind a socket, point the training command at it, run it, drain, clean up.

emry slurm-wrap [--project NAME] [--socket PATH] [--log-dir DIR] -- <training command…>

Spawns the engine via the current executable (… engine --project --socket --log-dir), waits for the socket, then runs the command with EMRY_MODE=sidecar + EMRY_SOCKET set (stdio inherited).
After the command exits, gives the engine a grace period to drain on RUN_FINISHED, then terminates it if still blocked on accept(). Propagates the command's exit code.
Socket path defaults to $TMPDIR/emry-{SLURM_JOB_ID or pid}.sock.

A CLI that runs the SLURM sidecar pattern (from the runbook) as one command: start the engine, bind a socket, point the training command at it, run it, drain, clean up. `emry slurm-wrap [--project NAME] [--socket PATH] [--log-dir DIR] -- <training command…>` - Spawns the engine via the current executable (`… engine --project --socket --log-dir`), waits for the socket, then runs the command with `EMRY_MODE=sidecar` + `EMRY_SOCKET` set (stdio inherited). - After the command exits, gives the engine a grace period to drain on `RUN_FINISHED`, then terminates it if still blocked on `accept()`. Propagates the command's exit code. - Socket path defaults to `$TMPDIR/emry-{SLURM_JOB_ID or pid}.sock`.