Problemsby GlyphPyTorch users keep filing bugs
46 posts from 37 people25d activeRising, +5 this week
Posts per day
The posts behind it
[distributions] FisherSnedecor and GeneralizedPareto return nan for moments that are infinite
### 🐛 Describe the bug Most distributions in `torch.distributions` return `inf` when a moment diverges to infinity and `nan` only when it is undefined. `FisherSnedecor` and `GeneralizedPareto` return `nan` for moments that are infinite, so the same situation gives a different...
DISABLED test_multiheadattention_fastpath_fp16_head_dim_alignment_xpu (__main__.TestTransformersAcceleratorXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/42342271695/job/128317745150#step:14:1572)). cc @mruberry @gujinghui @EikanWang @fengyuan14 @guangyey
Cached `torch.func.vjp` fails inside `torch.while_loop`
AI Disclosure: GPT-6.1-sol helped in generating the minimal working example and in drafting the issue. ### 🐛 Describe the bug I am trying to use `torch.while_loop` for the iteration `u ← g + Jᵀu` in an implicit backward solve. The Jacobian stays fixed, so I want to create the...
cross compile gcc 14 error: unrecognizable insn
pytorch 2.15.0 gcc 14.3.0 cc @malfet @zklaus @snadampal @milpuz01 @aditew01 @nikhil-arm @fadara01 @robert-hardwick @nWEIdia
[Bug][CPU] torch.binomial hangs indefinitely when count is NaN
### 🐛 Describe the bug ## Description `torch.binomial` may enter a non-terminating loop when `count` is NaN and `0 count`. This comparison is always false when `count` is NaN. **Suggested fix:** Validate non-finite counts before sampling and ensure the inversion loop cannot run...
Make varlen_attn visible to torch.compile
### 🚀 The feature, motivation and pitch Currently, `nn.attention.varlen.varlen_attn` is compiled as an opaque Python custom operator (`torch_attn::_varlen_attn`). Its implementation remains outside the compiled graph, so each invocation still executes Python backend selection...
Bug: NameError: name 'nn' is not defined in transformers/integrations/accelerate.py
### System Info - OS: Windows - Python version: 3.11 - PyTorch version: 2.5.1+cu121 - Transformers version: 4.38.2 ### Who can help? _No response_ ### Reproduction A `NameError` occurs because `nn` is not imported in `transformers/integrations/accelerate.py`, causing a crash...
Random weight init bug when using FSDP2 + CPU Offload
...On 2.13 training never gets past `init_weights`. Until the RNG tracker is fixed, a workaround is to initialize the local tensor directly, for example `nn.init.trunc_normal_(param.to_local(), ...)` for CPU-offloaded DTensors. That makes the CPU generator advance, though every...
[DCP] FileSystemWriter / FsspecWriter with thread_count > 1 can write incorrectly
### 🐛 Describe the bug If core team thinks this is a bug, I would actually love to make a patch for the team! With `thread_count > 1`, DCP starts extra writer threads in `FileSystemWriter._write_data` that target `_write_files_from_queue`. If one of those threads raises (e.g....
A Fix for Softmax and Cross-Entropy
### 🚀 The feature, motivation and pitch A request to implement the numerical stability approaches described by the paper: Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen. A Fix for Softmax and Cross-Entropy. *ArXiv pre-print*, 2026....
DISABLED test_grad_nn_functional_grid_sample_xpu_float32 (__main__.TestOperatorsDeviceXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203203#step:14:2097)). cc @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_operator_nn_functional_grid_sample_xpu_float32 (__main__.TestCompositeComplianceXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203183#step:14:7245)). cc @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_jvp_nn_functional_grid_sample_xpu_float32 (__main__.TestOperatorsDeviceXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203200#step:14:2663)). cc @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_noncontiguous_samples_arange_xpu_complex64 (__main__.TestCommonXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203199#step:14:5256)). cc @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_fn_grad_nn_functional_grid_sample_xpu_float64 (__main__.TestBwdGradientsXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203172#step:14:2820)). cc @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_redispatch_arange_xpu_complex128 (__main__.TestTorchFunctionRedispatchOpsDeviceXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203217#step:14:3823)). cc @mruberry @gujinghui @EikanWang @fengyuan14 @guangyey
DISABLED test_vmap_autograd_grad_nn_functional_grid_sample_xpu_float64 (__main__.TestOperatorsDeviceXPU)
Platforms: xpu This test was disabled because it is failing on main branch ([recent examples](https://github.com/pytorch/pytorch/actions/runs/40102459449/job/120904203197#step:14:3770)). cc @mruberry @gujinghui @EikanWang @fengyuan14 @guangyey
cholesky_inverse gives a nonzero derivative to the triangle of the factor it does not read
### 🐛 Describe the bug `torch.cholesky_inverse(L)` reads only the lower triangle of `L` (the upper one when `upper=True`), but its backward and forward AD formulas treat `L` as dense, so the unread triangle gets a nonzero gradient and tangent. Same bug as #199069 for...
Forked child processes exit with code 1: QueueService at-exit drain asserts after reset_for_fork
### Bug summary A process forked after a Prefect `QueueService` has started (for example the events or log workers inside a flow run) always exits with code 1, even when its target returns normally. `QueueService.start()` registers `self._at_exit` with...
torchrun exits 0 when a worker group in D-state runs out of restarts
### 🐛 Describe the bug When the uninterruptible-sleep (D-state) detection marks the worker group UNHEALTHY and no restarts are left, the agent stops the workers and then reports success: `elastic_launch` returns an empty dict instead of raising `ChildFailedError`, and torchrun...
Elastic c10d rendezvous reuses stale MASTER_ADDR/MASTER_PORT across rounds
### 🐛 Describe the bug With the c10d rendezvous backend and the shared rendezvous TCPStore (the default, `TORCH_DISABLE_SHARE_RDZV_TCP_STORE` unset), `DynamicRendezvousHandler` caches the bootstrap store info (`MASTER_ADDR`/`MASTER_PORT`) from its first round and reuses it in...
TORCHELASTIC_ENABLE_FILE_TIMER: one expired worker timer SIGKILLs torchrun
### 🐛 Describe the bug With `TORCHELASTIC_ENABLE_FILE_TIMER=1`, when a worker's `expires()` timer runs out, the `FileTimerServer` kills the worker and then sends the same signal (SIGKILL by default) to its own process. That server runs as a thread inside the torchrun agent, so...
Quick survey: your experience as a PyTorch Ecosystem maintainer
Hi, I'm reaching out from the PyTorch Foundation's Ecosystem team. We're running a short survey to understand how participation in the Ecosystem is going for maintainers like you, and where we can improve. It should take about 5 minutes:...
LocalResponseNorm accepts size <= 0 (late RuntimeError / silent no-op in F.local_response_norm)
### 🐛 Describe the bug `torch.nn.LocalResponseNorm` accepts invalid `size` values and fails late with a confusing kernel error, instead of rejecting at construction/validation time. Also, when calling `torch.nn.functional.local_response_norm` directly with a 3D input and `size...
UNSTABLE pull / linux-noble-xpu-n-py3.11 / build
✴️ iz2: The build fails at link time on every `main` commit since the xpu-n CI image was rebuilt and re-pushed under its existing tag on 2026-10-16 at 15:36Z: (also `urKernelSetArgValue/Local/Pointer/Sampler`). `.ci/docker/common/install_xpu.sh` runs `apt-get install -y...
Add documentation for SDPBackend Members
### 📚 The doc issue There's insufficient documentation on the [SDPBackend](https://docs.pytorch.org/docs/2.14/generated/torch.nn.attention.SDPBackend.html#torch.nn.attention.SDPBackend) class and no material about its members (eg: MATH, FLASH_ATTENTION, EFFICIENT_ATTENTION) and...
Introducting pytest-torch!
...in alpha, and I’d appreciate feedback-especially examples of model bugs that are frustrating to test or reproduce.. GitHub: https://github.com/AKJ7/pytest-torch . Thanks and please drop a star if possible..
AutoVideoProcessor.register treats positional exist_ok as a backend dictionary
### System Info - Transformers: 5.19.0.dev0, source revision 90ef040d400e94c771edb8b806c9c294e1b6a13f - Platform: Linux-5.15.0-88-generic-x86_64-with-glibc2.39 - Python: 3.12.13 - PyTorch: 2.11.0+cu129 - huggingface_hub: 1.31.0 - pytest: 8.4.2 - Device used for reproduction: CPU...
Trainer checkpoint rotation can delete valid checkpoints when backup directories match the checkpoint prefix
### System Info - Transformers: `5.19.0.dev0` - Upstream commit: `4cc2aa8430` - Python: `3.11` (Conda) - PyTorch: `2.14.1+cpu` - Platform: Windows - Device: CPU - Installation: Editable source checkout (`pip install -e ".[torch,testing]"`) The issue is reproducible without model...
JIT trace debug timing only works through torch.cuda
### 🚀 The feature, motivation and pitch The debug helper _time in torch/jit/_trace.py records elapsed time with CUDA events. It is gated by PYTORCH_JIT_TIME or an explicit time=True, and nothing in-tree calls it. On a machine whose accelerator is not CUDA, that debug timing...
https://github.com/pytorch/pytorch/issues/198563
### New Feature for Release ### ### Point(s) of contact @Przemas ### Release Mode (pytorch/pytorch features only) In-tree ### Out-Of-Tree Repo _No response_ ### Description and value to the user _No response_ ### Link to design doc, GitHub issues, past submissions, etc _No...
need help in finding the right resources
hey guys im an undergrad student (currently in 3rd year) and want to start learning ML and explore fields beyond that in the future. So I have seen a lot of people suggesting others to learn from Andrew Ng on coursera. I have the pdf of Hands-On Machine Learning with...
Caching allocator: cross-stream memory when running two streams on green contexts
Previously, I opened #195378 related to SM green context memory mapping. This is another question related to memory in the same setup. Currently, I'm working on LLM prefill and decode running in parallel using green contexts. Each of them runs in a separate stream. As you know,...
Unverified potential RCE / insecure shell usage in script helpers or build workflows
### 🐛 Describe the bug File: .ci/magma/Makefile Issue: Shell execution blocks use dynamic variables like $(shell git rev-parse --show-toplevel) and environment substitutions directly in command strings (DOCKER_RUN) without strict sanitization, which could lead to command...
CSM codebook embedding tying is ignored when `tie_word_embeddings=False`
### System Info - `transformers` version: 5.17.0 - Platform: Linux-7.0.0-30-generic-x86_64-with-glibc2.43 - Python version: 3.10.21 - Huggingface_hub version: 1.33.0 - Safetensors version: 0.8.0 - Accelerate version: 1.15.0 - Accelerate config: not found - DeepSpeed version: not...
DISABLED test_non_contig_byte_cpu_float16 (__main__.TestUnaryUfuncsCPU)
Platforms: macos This test was disabled because it is failing on main branch ([recent examples](https://torch-ci.com/failure?failureCaptures=%5B%22test%2Ftest_unary_ufuncs.py%3A%3ATestUnaryUfuncsCPU%3A%3Atest_non_contig_byte_cpu_float16%22%5D)). This test is flaky on macOS...
[XPU] Building Pytorch XPU with older ccache (4.9 - 4.11) fails
Starting 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...
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. ###...
Silent broadcasting is still a big problem and could be derailing your work right now
...silently broadcast the output of your model when the target shape mismatches, causing hard to diagnose issues. It still exists a ton in the wild, so I thought I'd actually highlight the symptoms of a silent broadcasting bug to keep people in the know.....
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)*....
105045167
### 🚀 The feature, motivation and pitch ### Alternatives _No response_ ### Additional context _No response_
[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...
DistributedSampler(num_replicas=0) raises a rank error instead of rejecting num_replicas
I 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 > >...
`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 @Isalia20
[dynamo] Meta-issue: torch.compile gaps found via the CPython test suite
...is the usual example). ## How to run PyTorch must be built against Python 3.13. A known-failing test reports as `skipped` with the graph-break reason attached, because it has a file under `test/dynamo_expected_failures/CPython313- - . `. **Deleting that file is part of your...
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/...
Companies and products named
About this problem
Evidence
46 posts from 37 people in 6 places, about 5 a week over 67 days. Mostly on pytorch/pytorch, huggingface/transformers, r/learnmachinelearning. Tools named alongside: GitHub, Hugging Face, NVIDIA, PyPI.
Frustration Frustration 0 of 3· Seen on GitHub issues, Reddit
Summary, model-written
PyTorch developers encounter continuous integration issues, null pointer SIGSEGV crashes on Apple Silicon, and test failures on virtualized macOS environments.
How it was grouped
Posts about PyTorch that state a pain but match no known issue yet. First post Sep 2, 2026, latest Nov 7, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
Other PyTorch problems
PyTorch GPU access and drivers are painful
14 posts from 14 people on GitHub issues.
PyTorch upgrades break working setups
26 posts from 21 people on GitHub issues, Reddit.
Trouble with PyTorch values
Top post: "How can I sum elements of a tensor based on a condition of one out of two values in its index pair into a new", one of 3 posts from 3 people on GitHub issues, Stack Overflow.
Trouble with PyTorch causing
Top post: "PyVista plotting running slowly, resetting box size and causing collisions", one of 3 posts from 3 people on GitHub issues, Stack Overflow.
PyTorch deprecations force unplanned migrations
8 posts from 5 people on GitHub issues.
Trouble with PyTorch true
Top post: "PaddleOCR-VL: position embedding interpolation changed from align_corners=False to True", one of 9 posts from 6 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 12 posts from 9 people on GitHub issues.
PyTorch builds are slow or flaky
9 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.
Running PyTorch locally hits memory and speed limits
5 posts from 5 people on GitHub issues, Hacker News.
PyTorch lock-in makes switching costly
4 posts from 4 people on GitHub issues.
Trouble with PyTorch windows
Top post: "Bundled cuDNN fails on first use since #202206 when no cuDNN is on PATH (Cannot load symbol cudnnGetVersion)", one of 4 posts from 4 people on GitHub issues.
History