Fix missing initializations for models created in 2023 #39239
+272
−163
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
cc @Cyrilvallez as we are also working on putting the default in the PreTrainedModel |
|
Super happy to see you're following again on a new batch of models @bvantuan! 🚀🤗 Let me know when this is ready! |
|
Yes, of course! Really excited to keep contributing whenever I have time. |
|
[For maintainers] Suggested jobs to run (before merge) run-slow: align, autoformer, bridgetower, bros, clap, clvp, efficientnet, fastspeech2_conformer, informer, kosmos2, mgp_str, mobilevit, mobilevitv2, mra, nllb_moe, owlv2 |
bvantuan
commented
Jul 9, 2025
Comment on lines
+605
to
+606
| elif isinstance(module, BrosRelationExtractor): | ||
| nn.init.normal_(module.dummy_node, std=std) |
There was a problem hiding this comment.
bvantuan
commented
Jul 9, 2025
| elif isinstance(module, nn.Embedding): | ||
| module.weight.data.normal_(mean=0.0, std=factor * 0.02) | ||
|
|
||
| elif isinstance(module, nn.LayerNorm): | ||
| elif isinstance(module, (nn.LayerNorm, nn.BatchNorm2d)): |
There was a problem hiding this comment.
bvantuan
commented
Jul 9, 2025
Comment on lines
+445
to
+461
| elif isinstance(module, nn.Linear): | ||
| shape = module.weight.data.shape | ||
| gain = 1.0 | ||
| scale = 1.0 # extra scale for gain | ||
| if module.bias is not None: | ||
| module.bias.data.zero_() | ||
| if shape[0] > shape[1]: | ||
| gain = math.sqrt(shape[0] / shape[1]) | ||
| if shape[0] == self.config.vocab_size and shape[1] == self.config.hidden_size: # final projection? | ||
| scale = 0.5 | ||
|
|
||
| gain *= scale | ||
| nn.init.orthogonal_(module.weight, gain=gain) | ||
| elif isinstance(module, nn.Embedding): | ||
| shape = module.weight.data.shape | ||
| gain = 1e-4 * math.sqrt(max(shape[0], shape[1])) | ||
| nn.init.orthogonal_(module.weight, gain=gain) |
There was a problem hiding this comment.
|
Cc @Cyrilvallez ! The PR is now ready and awaiting your review😊. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Fixes missing weight initializations for models created in 2023.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@Cyrilvallez