0.8.0 rc3 cherry picks #1475
Merged
0.8.0 rc3 cherry picks #1475
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
### Description This PR adds the missing pattern to identify the final norm layer in LoRA models. It also cleans up some of the classes in the model builder. ### Motivation and Context The missing final norm layer in LoRA models caused the generated LoRA models to be incorrect.
### Description This PR adds [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) support in the model builder. ### Motivation and Context Most SLMs and LLMs are trained in bfloat16 precision. Casting from bfloat16 to [float16](https://en.wikipedia.org/wiki/Half-precision_floating-point_format) can cause accuracy loss in models (e.g. Google's Gemma model family). The NumPy dependency when converting a [torch.Tensor](https://pytorch.org/docs/stable/tensors.html) object to an [ONNX TensorProto](https://onnx.ai/onnx/api/helper.html#onnx.helper.make_tensor) object has been removed. This will allow torch.Tensor objects in [other precisions that are not supported in NumPy](https://numpy.org/doc/stable/user/basics.types.html#relationship-between-numpy-data-types-and-c-data-types) to be converted to ONNX TensorProto objects. This PR also fixes [this issue](#691).
### Description This PR fixes accuracy issues with Google's Gemma models by using bfloat16 precision, [always using float32 precision to compute any LayerNorms](https://github.com/huggingface/transformers/blob/fee1190601b5d04ec6d3f7f58fd22788d7f3236d/src/transformers/models/gemma3/modeling_gemma3.py#L141-L146), and casting the output logits to float32 always. ### Motivation and Context This PR has been tested with Gemma-2 and Gemma-3. It is using the bfloat16 changes from [this PR](#1447) as well as the missing final norm changes from [this PR](#1420). --------- Co-authored-by: Nenad Banfic <46795300+nenad1002@users.noreply.github.com> Co-authored-by: Nenad Banfic <nebanfic@microsoft.com>
Most CPUs do not support BF16, hence removing it as an option since we miss some underlying kernel implementation
Integrate Constrained decoding using LLGuidance library. Based on Ying's Constrained Decoding branch (yingxiong/constrained_decoding) --------- Co-authored-by: Ying Xiong <yingxiong@microsoft.com> Co-authored-by: Michał Moskal <michal@moskal.me> Co-authored-by: Kunal Vaishnavi <kvaishnavi@microsoft.com> Co-authored-by: Ryan Hill <38674843+RyanUnderhill@users.noreply.github.com> Co-authored-by: Baiju Meswani <bmeswani@microsoft.com>
Update windows packaging pipelines to use build.py by aciddelgado · Pull Request #1468 · microsoft/onnxruntime-genai
baijumeswani
approved these changes
May 14, 2025
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.
Address previous PR review comments from #1470 (#1473)
Address QNN specific regressions (#1470)
Fix array eos_token_id handling (#1463)
Constrained decoding integration (#1381)
Remove BF16 CPU from valid GQA configuration (#1469)
Avoid adding providers if not requested (#1464)
Persist provider options across ClearProviders, AppendProvider where possible (#1454)
Fix accuracy issues with Gemma models (#1448)
Add bfloat16 support in model builder (#1447)
Add final norm for LoRA models (#1446)
Update version to 0.8.0-rc3