sf admin

Administrator commands: users, quotas, and maintenance mode

Requires the admin role. See the Admin page for the corresponding console features.

User Management

sf admin users                          # All users
sf admin user-add <username>            # Create a local account (password entered interactively)
sf admin set-role <username> <role>     # admin / operator
sf admin disable <username> --on        # Disable
sf admin disable <username> --off       # Enable

Quotas

sf admin set-quota <username> \
  --gpus 16 \                  # Concurrent GPU limit (default 8)
  --jobs 6 \                   # Concurrent job limit (default 4)
  --daily-gpu-hours 200 \      # Daily GPU-hours (0 = unlimited)
  --max-storage-gb 2000 \      # Run-directory storage cap in GiB (0 = unlimited)
  --profiles h200,h100 \       # Allowed profiles (comma-separated, empty = all)
  --priority 10                # Queue priority (default 0)

Storage quotas do not work like GPU quotas. GPUs are reserved: submitting a job sets its cards aside up front. Storage can only be measured after the fact — a training process writes to disk without telling the platform, and a job cannot say how much it will write, so there is nothing to reserve. The rule is “already over means no new jobs are dequeued”; running jobs are unaffected.

Usage comes from a periodic background scan (FORGE_STORAGE_SCAN_INTERVAL_S, 600s by default), so it is eventually consistent and a user may briefly exceed the cap between scans. If the scan wedges, the gate admits rather than refusing — better to show “stale” on the dashboard than to stop everyone from submitting.

Maintenance Mode

Drain the cluster → upgrade → resume; jobs resume training from checkpoints without losing progress:

sf admin maintenance status     # Current status and whether the cluster can be safely restarted
sf admin maintenance drain      # Enter maintenance mode and drain (idempotent; run again if not fully drained)
# …… upgrade the cluster / platform ……
sf admin maintenance resume     # Exit maintenance; paused jobs resume training automatically

Maintenance mode is the master switch: after drain, new jobs are not dequeued and running jobs are paused (checkpoints kept). After resume everything recovers automatically, with no per-job operations needed.

sf admin

Manage users and quotas

sf admin <subcommand>

Command	Description
sf admin disable	Disable or enable a user
sf admin maintenance	Drain, upgrade and resume the cluster
sf admin set-quota	Set a user’s compute and storage quota
sf admin set-role	Change a user’s role
sf admin user-add	Create a local account
sf admin users	List every user

sf admin disable

Disable or enable a user

sf admin disable <USERNAME> [options]

Argument	Description
USERNAME	required

Option	Description
--on	—on disables, —off enables

sf admin maintenance

Drain, upgrade and resume the cluster

sf admin maintenance <subcommand>

Command	Description
sf admin maintenance drain	Enter maintenance mode and drain
sf admin maintenance resume	Leave maintenance mode
sf admin maintenance status	Show maintenance state and safety

sf admin maintenance drain

Enter maintenance mode and drain

sf admin maintenance drain [options]

Option	Description
--note	Maintenance note, echoed back to anyone whose submission is blocked

sf admin maintenance resume

Leave maintenance mode

sf admin maintenance resume

No options beyond --help.

sf admin maintenance status

Show maintenance state and safety

sf admin maintenance status

No options beyond --help.

sf admin set-quota

Set a user’s compute and storage quota

sf admin set-quota <USERNAME> [options]

Argument	Description
USERNAME	required

Option	Description
--gpus <int>	Concurrent GPU cap (default 8)
--jobs <int>	Concurrent job cap (default 4)
--daily-gpu-hours <int>	GPU-hours per day (0 = unlimited) (default 0)
--max-storage-gb <int>	Cap on the user’s run directories in GiB (0 = unlimited). Enforced against a periodic scan: going over blocks new jobs, running ones continue (default 0)
--profiles	Allowed profiles (comma separated, empty = all)
--priority <int>	Queue priority (default 0)

sf admin set-role

Change a user’s role

sf admin set-role <USERNAME> <ROLE>

Argument	Description
USERNAME	required
ROLE	required

No options beyond --help.

sf admin user-add

Create a local account

sf admin user-add <USERNAME> [options]

Argument	Description
USERNAME	required

Option	Description
--password, -p	required
--role	admin | operator | viewer (default operator)
--email	—

sf admin users

List every user

sf admin users

No options beyond --help.