RepoJournal
PyTorch

@pytorch

PyTorch and the broader machine-learning ecosystem

Pick a date

The Wire · Showcase

DYNAMO BACKENDS UNBROKEN, PYTORCH COMPLETES OSDC MIGRATION

By RepoJournal · Filed · About PyTorch

The ts and aot_ts backends were silently failing since lazy graph modules became default, but dynamo tests masked it until today's fix ships.

Dynamo's critical backends are now working correctly after a fix that forces LazyGraphModule into plain GraphModule before torch.jit.script [1]. The tests masking this default config are also dropped, so the fix covers production use. In parallel, PyTorch completed its infrastructure migration off EC2 to OSDC across three workflows. The docs build [2], stable-fa3 tests [3], and linux build/test workflows [4] all dropped legacy EC2 paths, making OSDC/ARC the sole job. This cleanup removes roughly 150 lines of conditional logic that no longer has a purpose. On the stability front, CUDAGuardImpl now restores current device via RAII on exception paths, fixing a realistic crash case where cudaErrorNotReady would leak device state [5]. TorchTitan's integration CI is now green after fixes to numa API bindings, H100 cache setup, and per-microbatch D2H syncs from isfinite checks [6] [7].

Action items

References

  1. [1] [Dynamo] Fix ts and aot_ts backends with LazyGraphModule (#188875) pytorch/pytorch
  2. [2] Delete EC2 docs build path from _docs, rename build-docs-osdc to build-docs (#189169) pytorch/pytorch
  3. [3] Delete EC2 test path from _linux-test-stable-fa3, rename test-osdc to test (#189170) pytorch/pytorch
  4. [4] Remove the use-arc parameter from linux build/test/docs workflows (#189171) pytorch/pytorch
  5. [5] [c10][cuda] Restore current device via RAII on the throwing path in CUDAGuardImpl (#188167) pytorch/pytorch
  6. [6] [graph_trainer] Fix red CI: numa + H100 HF-cache + OpaqueBase startup breaks, chunked-loss remat scoping, quarantines ↗ pytorch/torchtitan
  7. [7] Avoid per-microbatch D2H caused by isfinite ↗ pytorch/torchtitan

FAQ

What changed in PyTorch on July 8, 2026?
The ts and aot_ts backends were silently failing since lazy graph modules became default, but dynamo tests masked it until today's fix ships.
What should PyTorch teams do about it?
Pull dynamo backend fix [ref:1] to unblock ts/aot_ts users in your next deploy • Review OSDC migration for any custom workflows still targeting EC2 paths [ref:2] [ref:3] [ref:4] • Monitor TorchTitan integration suites for further regressions after CI fixes [ref:6]
Which PyTorch repositories shipped on July 8, 2026?
pytorch/pytorch, pytorch/torchtitan

Related across the cluster

For your repos

The showcase is a teaser.
Your wire is the product.

Same engine. Different stack. Below: what changes when the wire is yours.

Showcase wire

  • 14 famous open source orgs
  • One wire per day
  • Public, generic
  • Read on the web, when you remember

Your wire

  • Up to 1,500 of your repos - orgs, deps, vendors
  • Morning and evening briefs
  • Action items routed to your team
  • Slack delivery, email, breaking-news CVE alerts

Want a hands-on demo first? Ask a current user for an invite link.