Dennis van der Staay d01a7b0241 Back out "MatMal - fix folding logic" (#167884) ...

Summary:
For sepcific hardware (A100), Autocast will generate a relatively large error on Transformer (torch.nn.TransformerEncoder) when using no_grad decorator on dim=256 (and larger presuably).

H100 seems fine, as does A100 with mig (so less than full SMs).

For now backing out, and revisting next week.

Test Plan:
failed jobs:
https://fburl.com/scuba/remote_execution_action/jzcmujgk

 {F1983543613}

Reviewed By: t-ivan-gr

Differential Revision: D87111518

Pull Request resolved: https://github.com/pytorch/pytorch/pull/167884
Approved by: https://github.com/malfet

2025-11-15 08:29:08 +00:00

..

__init__.py

Revert "Back out "Do not decompose in functionalization/proxy tensor if autograd wouldn't have decomposed (#164939)" (#165910)" (#166812)

2025-11-03 15:06:11 +00:00

decompositions_for_jvp.py

[3/N] Use key in dict for existence checks (#167214)

2025-11-07 02:49:15 +00:00

decompositions_for_rng.py

[5/N] Apply ruff UP035 rule (#164423)

2025-10-02 07:31:11 +00:00

decompositions.py

Back out "MatMal - fix folding logic" (#167884)

2025-11-15 08:29:08 +00:00