Fix ORTTrainer failure on DeBERTa(base/v2/sew_d) fp16 training - #18529
Closed
JingyaHuang wants to merge 82 commits into
Closed
Fix ORTTrainer failure on DeBERTa(base/v2/sew_d) fp16 training#18529JingyaHuang wants to merge 82 commits into
JingyaHuang wants to merge 82 commits into
Conversation
* fix typos * fix sequence_length docs of LayoutLMv3Model * delete trailing white spaces * fix layoutlmv3 docs more * apply make fixup & quality * change to two versions of input docstring * apply make fixup & quality
…upport Opacus training (huggingface#18486) * changing BartLearnedPositionalEmbedding forward signature and references to it * removing debugging dead code (thanks style checker) * blackened modeling_bart file * removing copy inconsistencies via make fix-copies * changing references to copied signatures in Bart variants * make fix-copies once more * using expand over repeat (thanks @michaelbenayoun) * expand instead of repeat for all model copies Co-authored-by: Daniel Jones <jonesdaniel@microsoft.com>
* Create _config.py * Create _toctree.yml * Create index.mdx not sure about "du / ihr" oder "sie" * Create quicktour.mdx * Update _toctree.yml * Update build_documentation.yml * Update build_pr_documentation.yml * fix build * Update index.mdx * Update quicktour.mdx * Create installation.mdx * Update _toctree.yml
…face#18272) * Fix critical trace warnings to allow ONNX export * Force input to `sqrt` to be float type * Cleanup code * Remove unused import statement * Update model sew * Small refactor Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com> * Use broadcasting instead of repeat * Implement suggestion Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com> * Match deberta v2 changes in sew_d * Improve code quality * Update code quality * Consistency of small refactor * Match changes in sew_d Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>
Contributor
Author
|
close as it turned to be too messy even after rebasing. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
19 participants
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?
Context
It was reported in optimum huggingface/optimum#305 that the training on DeBERTa with optimum.onnxruntime.ORTTrainer is broken.
After investigation, the break comes from two causes:
XDropOutdidn't have a symbolic function. And it has been implemented by @garymm in support ONNX export of XDropout in deberta{,_v2} and sew_d #17502 and has been merged to the main of transformers.However with those two fixes, the fp32 training will work, but the mixed-precision training will fail due to mismatched inputs dtype for some
Matmulnodes. In #18272, somesqrtresults are cast tofp32, and they need to be re-casted to fp16 beforeMatmulops, and this PR is supposed to add the re-cast part.Fixes #huggingface/optimum#305
Who can review?
@LysandreJik @patrickvonplaten @lewtun