pub fn vram_threshold_warning(caps: &WorkerCapabilities) -> Option<String>Expand description
Operator-facing warning when the configured VRAM threshold exceeds the GPU VRAM the worker actually detected.
The studio matches jobs to a worker purely by its advertised
vram_threshold_gb, so a threshold set above the card’s real
capacity — e.g. the default 12 GB on an 8 GB consumer GPU — makes the
worker accept jobs its GPU can’t fit: they load, exhaust VRAM, and
fail with an OOM the operator then has to trace back to a config
value. Surfacing it on the handshake (one line next to the
capability summary) turns a silent OOM-on-claim into an actionable
“lower your threshold” breadcrumb.
Only fires when the VRAM probe returned a real positive total: a
detected 0 GB means the probe failed (no nvidia-smi / sysfs tree,
or a non-NVIDIA GPU we can’t size), where the threshold is the only
capacity signal we have and second-guessing it would be wrong. The
boundary is strict (threshold > total), so a threshold that exactly
matches the card stays silent. Pure so the wording + boundary are
unit-tested without a live GPU.