Issues / #3275

#3275 [Other] calibrate dies with "there is no GPU 0,1" on a dual config whose gpu is the string form

open · @MEsfits · 3 comentários · No GitHub

BenchmarksSetup & installServer & APIMulti-GPUNVIDIA / CUDAModels & quants

Descrição

### Area

[Other]  (anything else, including setup and installation)

### Desktop app, Terminal or source?

Source (setup.sh / setup.py in a source tree at 0.2.0)

### What happened

Setup's calibration of a dual-GPU config whose `"gpu"` is the string form `"0,1"` runs all measurement rounds, then stops with:

```
  [X]  there is no GPU 0,1: 0 = NVIDIA GeForce RTX 3090 Ti, 1 = NVIDIA GeForce RTX 3090

Setup stopped. Fix the item above and run it again - everything already done is kept and skipped.
```

Expected: the calibration completes and is saved, exactly as for the list form `"gpu": [0, 1]`. Nothing is saved, because the hardware key is what the measured rounds are saved under, and computing it is where it dies.

Cause: `hardware_key()` passes `cfg["gpu"]` to `gpu_info()` unchanged; `gpu_info()` matches one card by number (`x["index"] == pick`), so the string `"0,1"` matches nothing. Serve already splits the same string form, and list-form duals work. `recommend_remote_expert_opt()` also silently skips string-form duals (`isinstance(..., list)` check), so they never get the helper-cache recommendation.

### Smallest steps or command that reproduce it

1. Source tree at 0.2.0 (engine 0.1.45.3) with a dual-GPU config whose `"gpu"` is the string form `"0,1"` (serve accepts it).
2. Run setup and let it calibrate that config.
3. After "Measuring the expert tier ..." and the bench round, setup stops with the message above.

### Reproduction status

Reproduced locally

### Strata version

Strata 0.2.0-beta · engine 0.1.45.3 (source tree at main e88ee39d)

### Graphics card, memory and system

2x RTX 3090 (3090 Ti 24 GB + 3090 24 GB) · 59 GB RAM · Ubuntu 26.04

### Model, size and settings

Qwen3.8-Flash-Next IQ3_S dual config · 256K context · pictures on

### Logs

```
  Measuring the expert tier every 1, 160 swaps, decay 0.97 (restarts the engine) ...[strata] starting the engine: reading the model's weights ...
[...]
    146.4 tok/s

  [X]  there is no GPU 0,1: 0 = NVIDIA GeForce RTX 3090 Ti, 1 = NVIDIA GeForce RTX 3090
```

### Notes

This was reported before (#2378 on the 0.1.45 line) and a fix was sent, but it is not in the 0.2.0 tree (`setup.py` still has the pre-fix line). The earlier issue/PR were closed without it landing, so this is resubmitted for 0.2.0.

No site

Links install, modelos, releases.