Problemsby GlyphTrouble with PyTorch torch
3 posts from 3 people3d activeSteady, +1 this week
Posts per day
The posts behind it
torch 2.14.1 macOS arm64 cp315 wheel was built against pre-b4 CPython headers: segfault in gil_scoped_acquire::dec_ref on Python 3.15.0
### 🐛 Describe the bug On Python 3.15.0 (final), the `torch==2.14.1` macOS arm64 cp315 wheel segfaults whenever libtorch has to re-acquire the GIL from a non-main thread. For example, C++ dropping the last reference to a tensor whose Python object is already gone. The same code...
torch 2.7: every `from_pretrained` fails in 5.20.0 (`hf_storage` import behind a `torch>=2.7` guard)
...script?: No ### Who can help? @michaelbenayoun @ArthurZucker ### Reproduction Any model fails to load on torch 2.7.x with transformers 5.20.0: Importing the module directly shows the real error: Same two commands with `transformers==5.19.0` and the same torch 2.7.1 work fine, so...
5.19.0: models cannot be imported on torch 2.5, or on torch 2.6 without an accelerator (get_device_type calls torch.accelerator.current_accelerator())
...below 2.5, with "Disabling PyTorch because PyTorch >= 2.5 is required") instead of failing at import with a message that points at the model. A guard like this fixes the import on 2.5.1, 2.6.0 and 2.7.1 (CPU-only, checked by patching 5.19.0 in place): We hit this in a CI job...
Companies and products named
About this problem
Evidence
3 posts from 3 people in 2 places, about 1 a week over 18 days. Mostly on huggingface/transformers, pytorch/pytorch. Tools named alongside: Hugging Face, NVIDIA, GitHub Actions.
Frustration Frustration 0 of 3· Seen on GitHub issues
How it was grouped
Posts about PyTorch with no known issue, sharing the word "torch". First post Oct 7, 2026, latest Oct 24, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
Other PyTorch problems
PyTorch users keep filing bugs
Top post: "Meta-issue: torch.compile gaps found via the CPython test suite", one of 43 posts from 34 people on GitHub issues, Reddit.
Trouble with PyTorch docs
Top post: "Clarify UTF-8 / Vietnamese text handling in DataLoader and tokenization pipeline", one of 4 posts from 3 people on GitHub issues.
Trouble with PyTorch inductor
Top post: "`[Inductor][CPU] torch.var returns NaN for large-magnitude constant float64 inputs while eager returns 0`", one of 11 posts from 8 people on GitHub issues.
PyTorch GPU access and drivers are painful
9 posts from 9 people on GitHub issues.
PyTorch builds are slow or flaky
8 posts from 1 person on GitHub issues. Unverified.
Trouble with PyTorch input
Top post: "Generation crashes with encoder_repetition_penalty when encoder_outputs are provided without input_ids", one of 6 posts from 6 people on GitHub issues.
PyTorch deprecations force unplanned migrations
5 posts from 4 people on GitHub issues.
Running PyTorch locally hits memory and speed limits
4 posts from 4 people on GitHub issues, Hacker News.
Trouble with PyTorch true
Top post: "torch.cuda.memory_snapshot() drops a reference to True on every call and aborts Python < 3.12 (fix #199044", one of 4 posts from 4 people on GitHub issues.
PyTorch lock-in makes switching costly
3 posts from 3 people on GitHub issues.
Trouble with PyTorch drops
Top post: "constant-False cond_fn drops captured-tensor mutations", one of 3 posts from 3 people on GitHub issues.
PyTorch vulnerabilities are hard to triage and patch
3 posts from 3 people on Glyph Ibis.
History