accelerator_type shows "G" for GeForce RTX 3090 (Ray auto-detect passthrough) + latent lstrip prefix-strip bug

[nilsmechtel]

Summary

The worker's per-node accelerator_type reports "G" for an NVIDIA GeForce RTX 3090 (seen on the single-machine Europa worker, bioengine 0.15.0, Ray 2.55.1). Datacenter cards are unaffected — KTH reports A40 and deNBI reports T4 correctly on the same code version.

While investigating I also found a separate latent string-handling bug in the same function that isn't the cause of "G" but should be fixed alongside it.

Evidence

get_status() on the Europa worker (RTX 3090):

{
  "node_ip": "172.17.0.3",
  "total_gpu": 1.0,
  "accelerator_type": "G",
  ...
}

KTH (A40) and deNBI (T4) report correctly, so this is specific to the GeForce card, not a regression in the field plumbing.

Root cause of "G" — upstream in Ray, passed through verbatim

BioEngine never sets the accelerator resource; it only reads Ray's accelerator_type:<TYPE> node resource in bioengine/cluster/proxy_actor.py::_get_accelerator_type. The one string transform there (see below) cannot turn GeForce-RTX-3090 into G. Therefore the raw Ray resource on that node is literally accelerator_type:G — Ray's own GPU auto-detection labels the RTX 3090 as G, and BioEngine faithfully passes it through.

Datacenter GPUs map cleanly to a model code (A40 → A40, T4 → T4); the consumer GeForce product name does not, and Ray collapses it to G.

Suggested fix for the label

Give BioEngine a clean, explicit accelerator label instead of trusting Ray's auto-detection for unrecognized GPUs — e.g. set accelerator_type:GeForce-RTX-3090 as a custom resource when the worker starts Ray on such a node (this is what our Chiron worker on the same physical box does, and it displays the full name correctly).

Separate latent bug in the same function

bioengine/cluster/proxy_actor.py:334:

return resource_name.lstrip("accelerator_type:")

str.lstrip(chars) strips any leading character in the set {a, c, e, l, r, t, o, _, y, p, :} — it does not strip the literal prefix "accelerator_type:". It happens to be correct for A40, T4, and G (each starts with an out-of-set char), but it would silently corrupt any accelerator whose value begins with one of those letters. Examples:

"accelerator_type:A40".lstrip("accelerator_type:")            # -> "A40"   (ok, by luck)
"accelerator_type:GeForce-RTX-3090".lstrip("accelerator_type:") # -> "GeForce-RTX-3090" (ok, by luck)
# a hypothetical value starting with an in-set char would be truncated, e.g.
"accelerator_type:tesla".lstrip("accelerator_type:")          # -> "sla"   (BUG)

Fix:

return resource_name.removeprefix("accelerator_type:")

Note: fixing lstrip alone does not change the "G" display — the G comes from Ray's resource value, so the label fix above is also needed.

Impact

Cosmetic/observability: the dashboard and bioengine cluster status show G instead of the GPU model, and the lstrip misuse is a latent correctness bug for future accelerator names.

Repro

Run get_status() (or bioengine cluster status) against a worker on a GeForce RTX 3090 node and observe accelerator_type == "G".