$ the-wire · showcase
PyTorch 2.15.0 lands with a rewritten NCCL backend
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
15 people shipped this
peterdsharpe
@peterdsharpe
2 cited
soulitzer
@soulitzer
1 cited
mottopanikeiku
@mottopanikeiku
1 cited
Anthony Shoumikhin
@shoumikhin
1 cited
itrowbri
@itrowbri
Zitong Zeng
@zztbit
dsashidh
@dsashidh
quantumxiaol
@quantumxiaol
Andrey Talman
@atalman
George Hong
@georgehong
Ivan Zaitsev
@izaitsevfb
Driss Guessous
@drisspg
tianyu-l
@tianyu-l
ryan-li-ryanllli-aws
@ryan-li-ryanllli-aws
ruisizhang123
@ruisizhang123
PyTorch 2.15.0 makes its rewritten NCCL backend the default, and the surrounding repos spent the day hardening the edges that release will run into.
PyTorch 2.15.0 Release pytorch/pytorch
The rewritten NCCL backend is now the default and ships a flight recorder that dumps state on hang, and the release notes carry their own Backwards Incompatible Changes and Deprecations sections. Read those two sections before you pin, since both outline behavior your code may currently rely on.
[Deprecation] set_ warns when a nested key has a missing entry, and raises KeyError where it lost the value or failed pytorch/tensordict
set_ documents a KeyError for a missing key but behaved differently per container: TensorDict, lazy stack and tensorclass created the entry, a locked tensordict raised a lock RuntimeError, and TensorDictParams dropped the value silently; it now warns when a nested key has a missing entry and raises KeyError where it lost the value.
[BugFix] In-place reshape and repeat_interleave on a slice of a PersistentTensorDict raise NotImplementedError before they change anything pytorch/tensordict
reshape(..., inplace=True) and repeat_interleave(..., inplace=True) on a slice of a PersistentTensorDict failed mid-write with an error about set(..., inplace=False); they now raise the same NotImplementedError as flatten, unflatten, repeat and gather, before anything changes. Calls that used to leave a half-mutated dict now fail cleanly up front.
[BugFix] Keep MCAdvantage and TopKRewardSelector working with tensordict 0.17 pytorch/rl
MCAdvantage and TopKRewardSelector only pass as_nested_tensor=True to get when trajectories form a lazy stack, which silences the FutureWarning from tensordict 0.15 and keeps rewards identical on every tensordict version from 0.14 up.
Build Python 3.15 wheels with the final 3.15 release from conda-forge (#9265) pytorch/test-infra
Windows Python 3.15 wheel builds had been failing since Oct 27 in the conda setup step, which installs 3.15.0rc2 and its _python_rc marker package; the recipe now pulls the final 3.15 release from conda-forge, clearing the CondaHTTPError on the new conda 25.3.1 runner image. Elsewhere, torchtitan made its local compile config extensible for out-of-tree backends by replacing model.local_compile_...
Action items
- → Review the PyTorch 2.15.0 Backwards Incompatible Changes and Deprecations sections before pinning it pytorch/pytorch [plan]
