RepoJournal
PyTorch @pytorch
90 wires and counting

$ follow PyTorch

Keep up with PyTorch in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-09-13
stories 74
GitHub RSS Archive Reviews Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

PyTorch trims CuTeDSL known-good versions, fixes lerp on MPS

By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology

tianyu-l Irakli Salia YeonwooSung Driss Guessous qedbot Nikita Shulga vmoens
7 people shipped this
tianyu-l @tianyu-l 3 cited Irakli Salia @Isalia20 2 cited YeonwooSung @YeonwooSung 2 cited Driss Guessous @drisspg 1 cited qedbot @qedbot 1 cited Nikita Shulga @malfet 1 cited vmoens @vmoens 1 cited

PyTorch dropped nvidia-cutlass-dsl 4.4.1 through 4.5.2 from the torch/_native known-good list, because the re-vendored QuACK 0.6.4 imports cutlass APIs that only exist in nvidia-cutlass-dsl >= 4.6.

The native CuTeDSL story is the day's consequential one. A forward fix for the QuACK 0.6.4 re-vendoring removed 4.4.1, 4.4.2, and 4.5.2 from the torch/_native known-good versions [1]. Those builds would have imported cutlass.utils.block_copy, cutlass.cute.experimental.iket, and cutlass.pipeline.alloc_reserved_mbarrier, which only exist in nvidia-cutlass-dsl >= 4.6 (upstream QuACK pins ==4.6.2), so the known-good list was advertising combinations the code cannot load. The internal default is 4.4.2, which is now correctly outside the supported set.

On the correctness side, PyTorch made sparse tensor validation state thread-local by moving the deferred sparse-validation collector onto the existing _SerializationLocal state [2]. Concurrent torch.load calls used to be able to validate or clear one another's tensors; now they cannot. The change addresses the issue described in #194750, and it ships with a deterministic integration test that pauses a malformed weights-only load at validation. Separately on MPS, lerp with a mixed-dtype scalar weight now mirrors CUDA's lerp_tensor_kernel: read the CPU scalar weight on the host, drop it from the iterator with remove_operand, and let the scalar-weight path cast it to the compute dtype [3]. Metal LERP still does not support mixed dtypes in general, so an error check was added for other incompatible combinations.

The test suite also shrank where coverage was already duplicated. Five forward-only tests were removed from test/test_mps.py: test_log, test_log_ten, test_log_two, test_logsumexp, and test_gelu_tanh, all already covered by the MPS OpInfo consistency tests, which additionally check backward [4]. A second batch removed test_eq, test_eq_int64, test_logical_not, test_logical_and, test_logical_or, and test_logical_xor, plus the _wrap_tensor helper that only the logical tests used [5].

In the RL and TensorDict stacks, YeonwooSung stopped Minari mission stacks from being shared between observation and ("next", "observation") [6] and fixed RandomTruncationTransform to resolve a nested step_count key rather than only reading root "step_count" and writing root "truncated"/"done" [7]. vmoens added shared buffers for ParallelEnv workers that report their own metadata under metadata_from_workers=True [8]. torchtitan moved quantization and lora into dedicated homes, with lora becoming a ModelConfigTransform [9], relocated logging.py, profiler.py, and metrics.py into torchtitan/observability with each module logging under its own namespace [10], and deprecated minimal_async_ep as unmaintained, noting that the HybridEP backend will be deprecated once DistMoE is available [11].

Action items

References

  1. [1] [native] Drop CuTeDSL <= 4.5.x from known-good versions (#196834) ↗ pytorch/pytorch
  2. [2] Make sparse tensor validation state thread-local (#194751) ↗ pytorch/pytorch
  3. [3] [MPS] Fix lerp with a mixed-dtype scalar weight (#196662) ↗ pytorch/pytorch
  4. [4] [BE] Remove tests from test_mps for log, log10, log2, logsumexp and gelu (tanh) (#196877) ↗ pytorch/pytorch
  5. [5] [BE] Remove tests from test_mps for eq, logical_not, logical_and, logical_or and logical_xor (#196878) ↗ pytorch/pytorch
  6. [6] [BugFix] Stop sharing Minari mission stacks between observation and next ↗ pytorch/rl
  7. [7] [BugFix] Honor nested step_count in RandomTruncationTransform ↗ pytorch/rl
  8. [8] [Feature] Shared buffers for ParallelEnv workers that report their own metadata ↗ pytorch/rl
  9. [9] move quantization and lora to new home ↗ pytorch/torchtitan
  10. [10] improve logging folder and module organization ↗ pytorch/torchtitan
  11. [11] deprecate minimal_async_ep ↗ pytorch/torchtitan

$ ls pytorch/reviews/ # the step-back reads

the week in review Helion autotunes Metal, ending default 16x16x16 tiles the month in review ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave
Topics: Python AI / ML Full archive → Weekly reviews →

Quick answers

What shipped in PyTorch on September 13, 2026?
PyTorch dropped nvidia-cutlass-dsl 4.4.1 through 4.5.2 from the torch/_native known-good list, because the re-vendored QuACK 0.6.4 imports cutlass APIs that only exist in nvidia-cutlass-dsl >= 4.6. In total, 50 commits and 24 pull requests landed.
Who contributed to PyTorch on September 13, 2026?
7 developers shipped this update, including Driss Guessous, qedbot, Nikita Shulga, Irakli Salia, YeonwooSung, vmoens, and tianyu-l.
What were the notable PyTorch updates?
[native] Drop CuTeDSL <= 4.5.x from known-good versions (#196834), Make sparse tensor validation state thread-local (#194751), and [MPS] Fix lerp with a mixed-dtype scalar weight (#196662).
♥

$ log @pytorch --recent

torchtitan fixes multi-node HF export, torch._native opens to non-aten namespaces 57 shipped Fake-tensor hardening, ROCm graph fixes, and an XNNPACK use-after-free 180 shipped CPU caching allocator teardown race fixed, torchtitan shares one training engine 67 shipped ROCm test timeouts traced to duplicated runs, and distributed/test_store drops 15 minutes 100 shipped

$ log --all # elsewhere on the wire

@openclaw Cold-start Gateway diagnostics stop crying wolf, doctor repairs plugin drift Agentic Coding Cline desktop bundles shrink, goose unblocks its own clicks

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.