mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
527ee63b7d
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/39205 Context: * https://github.com/pytorch/pytorch/pull/38478 modified convbn folding logic * https://github.com/pytorch/pytorch/pull/38820 fixed the ^ to be backwards compatible and be able to load v1 state dicts This PR is an additional fix on backwards compatibility - it allows older dicts to be loaded with `strict == False`. This is important because there are teams who use this flow to load floating point checkpoints into fused models, with `strict == False` Test Plan: 1. save a floating point and corresponding fused model: https://gist.github.com/vkuzo/177eba811a7a2ac359054fe9d4e3f099 2. load both of them, it works with strict==False and the floating point one fails with a good error message with strict==True: https://gist.github.com/vkuzo/447c9e797f208cb98447ffb24359d73e Imported from OSS Differential Revision: D21774353 fbshipit-source-id: f85f0c7fa956561824c9addb9198fea7a76a91aa |
||
|---|---|---|
| .. | ||
| modules | ||
| qat | ||
| quantized | ||
| __init__.py | ||