Skip to content
Problems.

Trouble with PyTorch inductor

MLOps6 posts from 5 people4d active+44 posts in the last 7 days, 0 the 7 days before (rising)

Posts per day

Posts per day6 posts, Sep 3 to Oct 5

The posts behind it

6, newest first
PostDate
[inductor] AOTI arrayref (stack-allocation) wrapper tests consistently red on trunk since mid-September...unrelated PRs of mine, and it reproduces on main itself, so here is a tracker. **What is failing** `test/inductor/test_aot_inductor_arrayref.py::AOTInductorTestABICompatibleCpuWithStackAllocation`: 45 distinct tests across the two shards on the latest completed main run I...pytorch/pytorchhe-yufengSep 305 days ago
[Inductor] soft_margin_loss returns wrong dtype for mixed-dtype input and target### 🐛 Describe the bug `torch.compile` with the Inductor backend produces the wrong output dtype for `F.soft_margin_loss` when the input and target have different floating-point dtypes. Eager and `aot_eager` preserve the input dtype (`float32`), while Inductor promotes the...pytorch/pytorchxyt66665000Sep 296 days ago
[Inductor] Tensor.exponential_() can return +inf### 🐛 Describe the bug `torch.compile` with the Inductor backend can occasionally make `Tensor.exponential_()` return `+inf` on both CPU and CUDA. PyTorch documents `Tensor.exponential_()` as sampling from an exponential distribution with actual support `(0, inf)`, so `+inf` is...pytorch/pytorchxyt66665000Sep 296 days ago
[inductor][cpu] Compiled "flex_attention" returns wrong values with strided view with storage offset### 🐛 Describe the bug (On CPU) `torch.compile` of `flex_attention` produces wrong results when q, k and v are slices of projection output and q/k are computed inside the graph. Eager `flex_attention` matches SDPA. Calling `.contiguous()` on v makes the compiled result match...pytorch/pytorchIvMisticosSep 296 days 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 44 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 34 weeks ago

Companies and products named

Company or productPosts naming it
PyTorch6
GitHub2
NVIDIA2
About this problem

Evidence

6 posts from 5 people in 1 place, about 2 a week over 28 days. Mostly on pytorch/pytorch. Tools named alongside: GitHub, NVIDIA.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts about PyTorch with no known issue, sharing the word "inductor". First post Sep 3, 2026, latest Sep 30, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method

History

  • 2026-10-05 Added: Trouble with PyTorch inductor (6 posts)

Rising problems by email

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

Double opt-in. Unsubscribe any time.