Symptom
After restarting forgather_server (or just opening the SubmitModal afresh on a config that previously used DiLoCo), the DiLoCo configuration section starts folded. The operator can't see at a glance whether the previous DiLoCo settings were restored — and when the prior settings could not be restored (e.g. the previously-selected base URL no longer corresponds to a running server, or the registry entry was deleted), the form silently falls back to base = None (= no DiLoCo) with no warning.
The result: an operator who submits expecting their previous DiLoCo config to be in effect gets a job spawned without DiLoCo, then hits a later error like:
ValueError: shard_dataset.method='work_units' requires DILOCO_SERVER to be set …
which is correct but only diagnoses the symptom several layers downstream of the root cause.
Acceptance
Pick one or both:
- Default the DiLoCo tab to open when the project's config indicates DiLoCo is in use (e.g.
shard_dataset.method='work_units' is set in the rendered config), OR whenever the persisted form snapshot had a non-empty base.
- Surface a warning in the modal when the previously-saved DiLoCo
base cannot be resolved against the current server list. The warning should name the missing URL and link to the DiLoCo view so the operator can restart that server (or register it). The form should keep the prior base displayed (even though it's stale) so the operator sees what's broken, rather than silently resetting to None.
Related
Symptom
After restarting forgather_server (or just opening the SubmitModal afresh on a config that previously used DiLoCo), the DiLoCo configuration section starts folded. The operator can't see at a glance whether the previous DiLoCo settings were restored — and when the prior settings could not be restored (e.g. the previously-selected base URL no longer corresponds to a running server, or the registry entry was deleted), the form silently falls back to base = None (= no DiLoCo) with no warning.
The result: an operator who submits expecting their previous DiLoCo config to be in effect gets a job spawned without DiLoCo, then hits a later error like:
which is correct but only diagnoses the symptom several layers downstream of the root cause.
Acceptance
Pick one or both:
shard_dataset.method='work_units'is set in the rendered config), OR whenever the persisted form snapshot had a non-emptybase.basecannot be resolved against the current server list. The warning should name the missing URL and link to the DiLoCo view so the operator can restart that server (or register it). The form should keep the priorbasedisplayed (even though it's stale) so the operator sees what's broken, rather than silently resetting to None.Related
EOF
)