#6649 Docker: inkling-2-mini setup crashes with AttributeError ('int' object has no attribute 'get') - install_inkling2 expects GPU dicts but the docker preset passes int indices
open · @ArkadiyBL · 1 コメント · GitHub で見る
本文
## Environment
- Strata: 0.2.19, docker image `ghcr.io/niko1221/strata:0.2.19-cuda13` (engine 0.1.65)
- Host: Ubuntu 22.04.5, Docker 29.9.3, NVIDIA driver 580.178.04
- GPUs: 5x RTX 3060 Ti (8 GB). One card is dead at the PCI level; hidden from the container with `CUDA_VISIBLE_DEVICES`
## Steps to reproduce
```sh
docker run -d --name strata --gpus all \
-e CUDA_DEVICE_ORDER=PCI_BUS_ID -e CUDA_VISIBLE_DEVICES=0,2,3,4,5 \
-e FAMILY=inkling-2-mini -e MODEL=SQ-10 -e API_KEY=... \
-p 8080:8080 --ulimit memlock=-1:-1 --shm-size 2g --stop-timeout 120 \
-v strata-data:/data \
ghcr.io/niko1221/strata:0.2.19-cuda13
```
Also reproduces with a single visible GPU (`CUDA_VISIBLE_DEVICES=5`) and with all cards visible - the crash is in setup, before any GPU-count logic matters.
## Result
```
strata: The model could not be installed. Setup stopped with an error (AttributeError: 'int' object has no attribute 'get').
```
Full traceback from `/data/state/strata/launch/docker.setup.log`:
```
Traceback (most recent call last):
File "/opt/strata/terminal/setup.py", line 11913, in <module>
code = main()
File "/opt/strata/terminal/setup.py", line 11376, in main
return install_inkling2(a, family, gpu, hip, amd, cuda_tk, multi, port, ram, adopted)
File "/opt/strata/terminal/setup.py", line 10197, in install_inkling2
dual_gpus = [g for g in (multi or []) if g.get("vendor", "nvidia") == "nvidia"] if (not hip and not amd and not unified) else []
^^^^^
AttributeError: 'int' object has no attribute 'get'
```
## Cause
The docker preset path builds `multi` as a list of **int** GPU indices (setup.py:11212/11236: `multi = sel if len(sel) > 1 else []`). The qwen installers (`install_own_path`, `install_hy4_flash`) consume it as ints, but `install_inkling2()` (setup.py:10197) is the only consumer that expects a list of **dicts** with `vendor` / `vram_gb` / `index` keys - so every docker inkling launch crashes at setup.
## Workaround (verified working)
Normalize int entries to real GPU dicts at the top of `install_inkling2` (there is already a `gpus()` helper that returns dicts with `index` / `name` / `vram_gb`):
```python
# setup.py, install_inkling2(), replacing the current dual_gpus line:
_ms = [g if isinstance(g, dict) else next((f for f in gpus() if f["index"] == g),
{"index": g, "vendor": "nvidia", "vram_gb": 0.0}) for g in (multi or [])]
dual_gpus = [g for g in _ms if g.get("vendor", "nvidia") == "nvidia"] if (not hip and not amd and not unified) else []
```
With this patch, Dual-GPU EP works end to end: SQ-10 was downloaded, placed on 2 GPUs (`ep_split=0.50`, `experts: 3072 VRAM, 0 host`) and served the OpenAI/Anthropic API normally.
## Notes
- The qwen36 family is not affected by this crash (its installer handles int `multi`), but on low-RAM hosts (16 GB) it fails later by design: `no host memory for the experts (19195 MiB)` - the expert arena needs more host RAM than the machine has.
- Everything else in the 0.2.19 docker image worked as documented (qwen36 download, error messages, engine start attempts).関連リンク
インストール・モデル・リリースへのサイト内リンク。