RepoJournal
PyTorch @pytorch
117 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 2026-07-26
stories 19
GitHub RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

DYNAMO FIXES GRAPH-BREAKING SDPA PARAMS, PAT SVD LEAK BREAKS DOWNSTREAM TESTS

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

Guan-Ming (Wesley) Chiu Denis Boyda Jim Meyering drisspg Lisa Jin
5 people shipped this
Guan-Ming (Wesley) Chiu @guan404ming 2 cited Denis Boyda @dboyda 1 cited Jim Meyering @meyering 1 cited drisspg @drisspg 1 cited Lisa Jin @lisjin 1 cited

PyTorch shipped fixes for two critical compilation and module patching bugs overnight that were silently breaking downstream workflows.

Dynamo can now construct SDPAParams inside compiled regions without graph-breaking under fullgraph=True [1], routing the pybind11 class through TorchInGraphFunctionVariable following the torch.nn.Parameter pattern. The same Dynamo reland [3] restores getattr_static for internal attribute probing after it was reverted for inflating frame counts, but the fix is back in and the frame-count assertions pass. On the CUDA front, _scaled_mm with float8 output was silently dropping the scale_result parameter [5], causing identical bitwise outputs for different scales. Over in pytorch/ao, insert_svd_modules_ was mutating the shared nn.Linear class directly, leaving AttributeError traps in weight access across unrelated modules [6], breaking torch.compile tests in int4/int8/float8 workflows after any pat SVD test ran. The fix patches a private per-instance subclass instead. Meanwhile, c10d's TorchComms integration no longer requires users to mutate c10d internals before calling init_process_group [2], and out-of-range float conversions in tensor creation tests are now restricted to valid ranges [4].

Action items

References

  1. [1] [Dynamo] Support SDPAParams construction in fullgraph (#190839) pytorch/pytorch ↗
  2. [2] [c10d][torchcomms] Add MCCL backend wrapper example (#190292) (#191034) pytorch/pytorch ↗
  3. [3] [Dynamo] Reland: Use getattr_static when probing internal attributes (#190970) pytorch/pytorch ↗
  4. [4] [caffe2] Avoid out-of-range float->integral conversions (UB) in test_tensor_creation_ops (#191025) (#191025) pytorch/pytorch ↗
  5. [5] [CUDA] Restore scale_result for float8 _scaled_mm output (#190979) pytorch/pytorch ↗
  6. [6] [pat] Fix SVD class leakage into `nn.Linear` from #4587 (#4613) ↗ pytorch/ao

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

the week in review torch.compile dtype fix, two Arm CVEs patched the month in review PyTorch 2.14.0 ships, TorchRL 0.14 lands
Full archive → Weekly reviews →

Quick answers

What shipped in PyTorch on July 26, 2026?
PyTorch shipped fixes for two critical compilation and module patching bugs overnight that were silently breaking downstream workflows. In total, 17 commits and 2 pull requests landed.
Who contributed to PyTorch on July 26, 2026?
5 developers shipped this update, including Guan-Ming (Wesley) Chiu, Denis Boyda, Jim Meyering, drisspg, and Lisa Jin.
What were the notable PyTorch updates?
[Dynamo] Support SDPAParams construction in fullgraph (#190839), [c10d][torchcomms] Add MCCL backend wrapper example (#190292) (#191034), and [Dynamo] Reland: Use getattr_static when probing internal attributes (#190970).
♥

$ log @pytorch --recent

Tensordict compiles masked indexing under fullgraph; ExecuTorch links WebGPU into the Wasm build 138 shipped TensorDict 0.15.0 goes pure Python, 1.77x faster 183 shipped Inductor hands Triton kernels to the worker pool, RL rollout KV cache persists across turns 219 shipped ExecuTorch drops Python 3.10, TensorRT trims compile startup, tensordict fixes silent h5 loss 164 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training verl rejects conflicting vLLM engine_kwargs, open-instruct adds MILES OPD behind a Megatron learner Interpretability nnsight 0.8.0 cuts import time from 7.6s to 1.6s

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Interpretability Alignment & Post-training

$ 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.