'tsr generate -m <model> ...' without an explicit -r/--remote flag was
running _validate_model_available() against the local [comfyui] url
even when config.toml had default_remote pointing at a different host.
On a typical operator setup (default_remote = 'runpod', local comfyui
url stale or empty) this fails with 'Model X not available on ComfyUI
host (looked in checkpoints/ — 2 entries)' even though the model
exists on the remote and would be reachable for the actual dispatch
a few hundred lines later.
Root cause: two validation gates used the raw 'remote' parameter
(CLI flag only) instead of the resolved remote URL:
- generate(): pre-fanout validation for --parallel-queue path
- _run_generation(): per-call validation
Both now use do_resolve_remote(remote), which returns the resolved
URL when default_remote is set even if -r was omitted. The server-
side validator on the remote tensors API still catches missing
models on the remote host — matches the intent already documented
in the comment block.
Co-authored-by: marauder-actual <marauder@saiden.dev>
Adam Ladachowski
7144b7ac6a