Skip to content
Problems.

PyTorch users keep filing bugs

MLOps14 posts from 12 people8d active+11 post in the last 7 days, 3 the 7 days before (fading)

Posts per day

Posts per day14 posts, Sep 2 to Sep 25

The posts behind it

14, newest first
PostDate
[XPU] Building Pytorch XPU with older ccache (4.9 - 4.11) failsStarting from 6d2509d4f397087f72d64112d2fc9a65e0f4f433, on Ubuntu 24.04, with oneAPI DLE 2026.1. Pytorch XPU build fails on Ubuntu 24.04 with the system native `ccache` version `4.9.1` with the following kind of errors. Below is error example for the test app included into...pytorch/pytorchdvrogozhSep 232 days ago
[Accelerator] empty_host_cache() segfaults on macOS / PyTorch 2.13.0Calling the public `torch.accelerator.empty_host_cache()` API in a fresh Python process crashes with SIGSEGV on Apple Silicon environment. ## Reproduction Above crashes with SIGSEGV due to the null pointer dereference in...pytorch/pytorchsuntpSep 187 days ago
4D mask created under torch.compile when padding_mask is None### System Info ### Summary When running models with `torch.compile` and unpadded inputs (`attention_mask=None`), 4D masks are still materialized instead of passing `attn_mask=None` to SDPA. This causes noticeable SDPA slowdown due to unnecessary mask-add operations. ###...huggingface/transformersmsnliuSep 187 days ago
B200 dgxb200-03 in bad state> NOTE: Remember to label this issue with "`ci: sev`" > If you want autorevert to be disabled, keep the ci: disable-autorevert label ## Current Status *Status could be: preemptive, ongoing, mitigated, closed. Also tell people if they need to take action to fix it (i.e. rebase)*....pytorch/pytorchnWEIdiaSep 1510 days ago
105045167### 🚀 The feature, motivation and pitch ### Alternatives _No response_ ### Additional context _No response_pytorch/pytorchhha831928-sudoSep 102 weeks ago
[MPS] "Failed to created pipeline state object" due to "unsupported deferred-static-alloca-size" in cooperative_tensor during cholesky_ex on virtualized macOS runners...on virtualised macOS environments (e.g. GitHub Actions `macos-latest` runners). I am unable to reproduce this issue locally on actual Mac hardware. Because I am unable to reproduce this locally, here is the output from one of the virtualised runners where this happened (see...pytorch/pytorchjank324Sep 102 weeks ago
[MPS] Incorrect complex batched underdetermined lstsq solutions in PyTorch 2.14### 🐛 Describe the bug `torch.linalg.lstsq` returns incorrect solutions on MPS for complex batched underdetermined systems: `A.shape=(8, 32, 48)`, `B.shape=(8, 32, 3)`, `complex64`. Both devices use identical CPU-generated inputs, `driver="gelsd"`, and `rcond=1e-4`. All...pytorch/pytorchifsheldonSep 52 weeks ago
DistributedSampler(num_replicas=0) raises a rank error instead of rejecting num_replicasI confirmed this on current pytorch `main`. `num_replicas ### 🐛 Describe the bug > > `DistributedSampler(..., num_replicas=0, rank=0)` raises a `ValueError` about `rank`, not `num_replicas`. > > > > The constructor only checks `rank >= num_replicas or rank > ### Versions > >...pytorch/pytorchYeonwooSungSep 52 weeks ago
DataLoader(prefetch_factor=0) constructs, then raises AssertionError on iterateI confirmed this on current pytorch `main`. The constructor allows 0, but the multiprocessing iterator requires a positive factor. Drafted with AI assistance; I reviewed the repro and the two checks. > ### 🐛 Describe the bug > > `DataLoader(..., num_workers>0,...pytorch/pytorchYeonwooSungSep 52 weeks ago
`lerp_` of two non-negative tensors produces negative output on MPS on 2.14### 🐛 Describe the bug ### Versions cc @ezyang @gchanan @kadeng @msaroufim @kulinseth @malfet @DenisVieriu97 @jhavukainen @aditvenk @Isalia20pytorch/pytorchMikhailRubinchikSep 43 weeks ago
[dynamo][while_loop] constant-False cond_fn drops captured-tensor mutationsI reviewed the AI-assisted analysis below against the reproducer and current source. I am filing it because `cond_fn` must be evaluated once even when it immediately returns `False`, and inference explicitly permits mutation of captured tensors. > [AI-assisted analysis] > > ##...pytorch/pytorchruddyscentSep 43 weeks ago
[inductor][while_loop] carried-input mutation validation is bypassed by auto-functionalizationI re-reviewed the updated PR and reproduced the `cpp_wrapper` failure. The new test makes carried-input mutation depend on an unrelated captured-tensor mutation, contradicting the documented contract. > [AI-assisted analysis] > > ## Bug > > `torch.while_loop` documents that...pytorch/pytorchruddyscentSep 43 weeks ago
[inductor] mix-order reduction: split column sum ignores xmask, wrong results when rows % XBLOCK != 0With `triton.mix_order_reduction` on (OSS default), a row and a column reduction over the same input fuse into one kernel. The column reduction is accumulated per RSPLIT chunk with `tl.sum(<value, 0)` without masking the tile by `xmask`. Padded rows load `other=0.0`, so any...pytorch/pytorcheellisonSep 33 weeks ago
Integrate with ONNX 1.23.0 release branch### New Feature for Release We are releasing ONNX 1.23.0. A release branch is created (https://github.com/onnx/onnx/tree/rel-1.23.0). The planned release date is September 14, 2026. Release candidates are also available from TestPyPI: pip install -i https://pypi.org/simple/...pytorch/pytorchsfatimarSep 23 weeks ago

Companies and products named

Company or productPosts naming it
PyTorch14
GitHub6
Hugging Face1
NVIDIA1
vLLM1
GitHub Actions1
About this problem

Evidence

14 posts from 12 people in 2 places, about 4 a week over 22 days. Mostly on pytorch/pytorch, huggingface/transformers. Tools named alongside: GitHub, Hugging Face, NVIDIA, vLLM.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts about PyTorch that state a pain but match no known issue yet. First post Sep 2, 2026, latest Sep 23, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method

History

  • 2026-09-25 Statement: PyTorch users keep asking for help to PyTorch users keep filing bugs
  • 2026-09-25 Added: PyTorch users keep asking for help (14 posts)

Rising problems by email

Mondays: the problems in data, tech and AI that grew fastest that week.

Double opt-in. Unsubscribe any time.