Problemsby GlyphPyTorch GPU access and drivers are painful
14 posts from 14 people11d activeRising, +5 this week
Posts per day
The posts behind it
[DTensor] threshold() keeps Partial(max) / Partial(min) when value > threshold
...mirrored) | | input `Shard` / `Replicate` | correct (no partial rule involved) | ## Workaround Call `x.redistribute(placements=[Replicate()])` (or `.full_tensor()`) before `threshold` when the input may be `Partial(max/min)` and `value > threshold`. ## Fix direction Make the...
DISABLED test_short_json_timing_enabled_True_include_collectives_True (__main__.NCCLTraceTest)
Platforms: rocm This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_short_json_timing_enabled_True_include_collectives_True&suite=NCCLTraceTest&limit=100) and the most recent trunk [workflow...
[RISC-V] matrix_exp effectively hangs on a small batch containing NaN
### 🐛 Describe the bug On two RISC-V CPU systems, `torch.linalg.matrix_exp` did not return within 12 seconds for a batch of two 2×2 matrices containing one NaN. A full `test_linalg.py` process on SG2044 remained CPU-bound for roughly 28 hours, with sampled native stacks in...
TaskThreadPoolBase::defaultNumThreads() does not cache sched_getaffinity()
### 🐛 Describe the bug Since #194605, c10::TaskThreadPoolBase::defaultNumThreads() calls sched_getaffinity(0, ...) (instead of querying the # of processors/cores, which remains stable) without caching the result. In an environment that changes cpuset.cpus during the job's...
[Bug]: get_top_k_tokens applies scaling and soft capping in the wrong order
### Your current environment Collecting environment information... ============================== System Info ============================== OS : Ubuntu 22.04.5 LTS (x86_64) GCC version : (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0 Clang version : Could not collect CMake version :...
[LocalTensor] fully_shard under LocalTensorMode uses ProcessGroup.rank()==0, so every simulated rank keeps chunks[0]
### 🐛 Describe the bug fully_shard on a Linear under FakePG + LocalTensorMode returns the wrong values (full-shaped, no error). Real multi-process FSDP is unaffected. ### Observed eager [28.0, 92.0, 156.0, 220.0, 284.0, 348.0, 412.0, 476.0] fsdp [28.0, 92.0, 28.0, 92.0, 28.0,...
[c10d][nccl] test_init_with_idx flakes: NCCL watchdog + "driver shutting down" + SIGABRT at teardown (missing destroy_process_group)
...29, SHA 2e9b4aff, failed both attempts ~14.4s). - Test is byte-identical between the failing SHA and latest main; no upstream DISABLED/flaky-bot record exists for it. - Legacy backend has no watchdog, hence only the new path is exposed. Fix: add the same explicit...
[ROCm][Windows] Device-side assert prints no message: C10_USE_ROCM_KERNEL_ASSERT is off in the Windows ROCm wheel
...device-side assert triggered`. Only the missing message is PyTorch's. Expected: the failing assertion's message is printed, as it is on CUDA. ### Versions Windows 11 Pro 10.0.26100.9457; PyTorch 2.11.0+rocm7.13.0a20260416 (HIP 7.2.0, `amdhip64_7.dll` 10.0.3581.0); Python 3.14.6;...
[Inductor][PDL] triton.enable_pdl: a kernel loads a buffer written two kernels back before gdc_wait() when the kernel in between is store-only, giving wrong results
### 🐛 Describe the bug The analysis was done with assistance of AI and was reviewed by me. With `torch.compile(..., options={"triton.enable_pdl": True})`, a kernel's first load of a buffer written *two kernels back* is emitted before `tl.extra.cuda.gdc_wait()` when the kernel...
[Inductor] Compiled softshrink turns NaN into zero
### 🐛 Describe the bug F.softshrink(NaN) returns NaN in eager and aot_eager, but zero with the Inductor backend. This reproduces on CPU and CUDA. ### Error logs No exception. Eager and aot_eager return tensor([nan]); Inductor returns tensor([0.]). ### Versions PyTorch version:...
Tensor.requires_grad_ typing is incorrectly defined in __init__.pyi
### 🐛 Describe the bug `TensorBase.requires_grad_` is typed as `def requires_grad_(self, mode: _bool = True) -> Tensor:` however it actually accepts argument `requires_grad` rather than `mode`: To produce the error: ### Versions PyTorch version: 2.14.0 Is debug build: False...
[PT2] Device-TMA reduction stores lose tma_min_block_sizes
...the bug ## Describe the bug `store_reduction()` bypasses `codegen_block_ptr()` and fails to record an in-kernel device-TMA descriptor. The resulting kernel can select XBLOCK=1 for an fp32 descriptor store, which Triton rejects because 1 * 4 < 16 bytes. ## Minimal reproducer...
[FSDP2][NCCL] Large slowdown for non-16-byte-aligned per-rank collective payloads
### 🐛 Describe the bug ## Summary Using PyTorch's NCCL tensor collectives on a flat buffer with the same per-rank layout as FSDP2's default non-symmetric-memory communication path, a per-rank payload whose byte count is not a multiple of 16 causes a substantial latency...
[MPS] bmm/baddbmm/addbmm silently wrong on M1/M2 when K >= 32767 (2-D paths already routed by #183535)
### 🐛 Describe the bug > This report was prepared with the help of an AI assistant; everything in it was verified by me on an M2 Max running macOS 26.6.2. On Apple7/8 GPUs (M1/M2), MPSGraph matmuls with a reduction dimension K >= 32767 and both output dims >= 16 can sum memory...
Companies and products named
About this problem
Evidence
14 posts from 14 people in 2 places, about 1 a week over 72 days. Mostly on pytorch/pytorch, vllm-project/vllm. Tools named alongside: NVIDIA, GitHub, Inc., Red Hat.
Frustration Frustration 0 of 3· Seen on GitHub issues
How it was grouped
Posts that state a pain and match the "gpu-access" rule about PyTorch. First post Aug 28, 2026, latest Nov 7, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
NVIDIA GPU access and drivers are painful
12 posts from 12 people on GitHub issues, Mastodon, Reddit.
GPU access, drivers and CUDA versions are painful
7 posts from 7 people on GitHub issues, Reddit.
DeepSeek GPU access and drivers are painful
7 posts from 6 people on GitHub issues.
vLLM GPU access and drivers are painful
5 posts from 3 people on GitHub issues.
Claude Code GPU access and drivers are painful
3 posts from 3 people on GitHub issues.
Ollama GPU access and drivers are painful
3 posts from 3 people on Hacker News, Reddit, GitHub issues.
Chroma GPU access and drivers are painful
2 posts from 1 person on GitHub issues. Unverified.
Mistral GPU access and drivers are painful
2 posts from 2 people on GitHub issues. Unverified.
Other PyTorch problems
PyTorch upgrades break working setups
26 posts from 21 people on GitHub issues, Reddit.
PyTorch users keep filing bugs
Top post: "Meta-issue: torch.compile gaps found via the CPython test suite", one of 46 posts from 37 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