File ".../transformers/utils/import_utils.py", line 339, in get_device_type
device = torch.accelerator.current_accelerator() or torch.device("cpu")
AttributeError: module 'torch' has no attribute 'accelerator'
ModuleNotFoundError: Could not import module 'AutoModelForImageTextToText'
Why:SpikerSoft.EventHandlers.ImageDescription.Python/requirements.txt pins torch==2.5.1 / torchvision==0.20.1 but floats transformers>=4.48.0. A recent transformers release calls torch.accelerator unconditionally at import, which needs torch ≥ 2.6. Nothing in the repo changed here: #1299 (the S3 rename) only touched a comment in that requirements file, and the workflow fired because its paths were in the PR. The last time this image was built it resolved an older transformers. Unpinned ML deps = the build is red whenever upstream moves (same class as the TripoSR torchmcubes pin).
Not caused by the RustFS cutover / s3 rename - the failure is before any storage code runs.
Fix options: pin transformers<4.57 (or whatever the last torch-2.5-compatible release is) next to the torch pin, or bump torch/torchvision to 2.6.x and re-test Qwen-VL on the Jetson/4090 image. Either way the Python pins for this image should all be exact.
Impact: image-description workers keep running the previous image (no deploy happened - deploy was skipped). New code to this handler cannot ship until the build is green.
**What:** `spikersoft-image-description.yml` on master failed at `build_and_publish` (run 44288, 2026-10-07 17:51Z, https://git.spikersoft.com/spikerj/spikersoft-backend/actions/runs/44288). The image builds, then the import check dies:
```
File ".../transformers/utils/import_utils.py", line 339, in get_device_type
device = torch.accelerator.current_accelerator() or torch.device("cpu")
AttributeError: module 'torch' has no attribute 'accelerator'
ModuleNotFoundError: Could not import module 'AutoModelForImageTextToText'
```
**Why:** `SpikerSoft.EventHandlers.ImageDescription.Python/requirements.txt` pins `torch==2.5.1` / `torchvision==0.20.1` but floats `transformers>=4.48.0`. A recent transformers release calls `torch.accelerator` unconditionally at import, which needs torch ≥ 2.6. Nothing in the repo changed here: #1299 (the S3 rename) only touched a comment in that requirements file, and the workflow fired because its paths were in the PR. The last time this image was built it resolved an older transformers. Unpinned ML deps = the build is red whenever upstream moves (same class as the TripoSR torchmcubes pin).
**Not caused by the RustFS cutover / s3 rename** - the failure is before any storage code runs.
**Fix options:** pin `transformers<4.57` (or whatever the last torch-2.5-compatible release is) next to the torch pin, or bump torch/torchvision to 2.6.x and re-test Qwen-VL on the Jetson/4090 image. Either way the Python pins for this image should all be exact.
**Impact:** image-description workers keep running the previous image (no deploy happened - `deploy` was skipped). New code to this handler cannot ship until the build is green.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
What:
spikersoft-image-description.ymlon master failed atbuild_and_publish(run 44288, 2026-10-07 17:51Z, https://git.spikersoft.com/spikerj/spikersoft-backend/actions/runs/44288). The image builds, then the import check dies:Why:
SpikerSoft.EventHandlers.ImageDescription.Python/requirements.txtpinstorch==2.5.1/torchvision==0.20.1but floatstransformers>=4.48.0. A recent transformers release callstorch.acceleratorunconditionally at import, which needs torch ≥ 2.6. Nothing in the repo changed here: #1299 (the S3 rename) only touched a comment in that requirements file, and the workflow fired because its paths were in the PR. The last time this image was built it resolved an older transformers. Unpinned ML deps = the build is red whenever upstream moves (same class as the TripoSR torchmcubes pin).Not caused by the RustFS cutover / s3 rename - the failure is before any storage code runs.
Fix options: pin
transformers<4.57(or whatever the last torch-2.5-compatible release is) next to the torch pin, or bump torch/torchvision to 2.6.x and re-test Qwen-VL on the Jetson/4090 image. Either way the Python pins for this image should all be exact.Impact: image-description workers keep running the previous image (no deploy happened -
deploywas skipped). New code to this handler cannot ship until the build is green.