Introduce ORTSessionMixin and enable general io binding (works for diffusers as well)#2234
Merged
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
IlyasMoutawwakil
changed the title
…ConditionalGeneration with positional arguments.
IlyasMoutawwakil
force-pushed
the
ort-session-mixin
branch
from
May 9, 2025 11:09
420f59c to
3d7b976
Compare
There was a problem hiding this comment.
Pull Request Overview
This PR extends the IO binding functionality for multi‐part pipelines by introducing new mixin classes (ORTSessionMixin and ORTSessionsWrapper) and refactoring existing tests and utility functions for both diffusion and non‐diffusion workflows.
- Updates tests for onnxruntime modeling and diffusion pipelines to use updated function names and parameters.
- Refactors utility and exporter modules to adopt new patterns for provider and session handling, and removes deprecated or redundant code.
Reviewed Changes
Copilot reviewed 21 out of 21 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| tests/onnxruntime/test_modeling.py | Removed deprecated parameters and updated assertions to reflect changes in the IO binding interface and model naming. |
| tests/onnxruntime/test_diffusion.py | Renamed helper functions and added tests for IO binding functionality with diffusion pipelines. |
| optimum/utils/constant.py | Added and reorganized constants used across the codebase. |
| optimum/onnxruntime/utils.py | Refactored provider handling and added a new helper for inferring a session’s dtype. |
| optimum/onnxruntime/modeling_decoder.py | Updated deprecated code patterns to use instance configuration attributes consistently and improved IO binding calls. |
| Other files (e.g., optimization.py, exporters tasks, workflows) | Similar refactoring to align with the new multi-part pipeline design and clean up unused code. |
Comments suppressed due to low confidence (3)
tests/onnxruntime/test_modeling.py:140
- The 'use_io_binding' parameter has been removed in favor of a default behavior. Please ensure that this API change is clearly documented in the migration guide to avoid confusion among users.
model = model_cls.from_pretrained(self.LOCAL_MODEL_PATH, use_cache=False, use_io_binding=False)
optimum/onnxruntime/modeling_decoder.py:249
- Updating the condition to use 'self.config' instead of relying on a local 'config' variable improves maintainability. Confirm that similar updates are applied consistently throughout the file.
if self.config.model_type != "gpt_bigcode":
optimum/onnxruntime/utils.py:446
- [nitpick] Consider adding tests and a detailed docstring example for 'get_dtype_from_session' to verify that the correct torch dtype is returned based on the ONNX input/output types.
def get_dtype_from_session(session: ort.InferenceSession) -> torch.dtype:
echarlaix
approved these changes
May 16, 2025
echarlaix
left a comment
echarlaix
left a comment
Collaborator
There was a problem hiding this comment.
LGTM! Thanks a lot for the great PR @IlyasMoutawwakil 🔥
| self.output_dtypes = {output.name: output.type for output in session.get_outputs()} | ||
|
|
||
| self.model_path = Path(session._model_path) | ||
| self.model_name = self.model_path.name |
Collaborator
There was a problem hiding this comment.
yes talking about model_name which seems to be used in the tests (and can be easily extracted from model_path) so we can remove imo, not specific to this PR though so feel free to close!
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
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?
This PR tries to further extend the io binding feature to multi part pipelines and enables it more generally through two classes:
ORTSessionMixin: should be used whenever a class has an underlying inference session, it enables inference with and without io/binding, and handles the torch/onnxruntime interface like provider/device/dtype.ORTSessionsWrapper: should be used when a class has many models/parts (like encoder-decoders or diffusion pipelines), it usinfies the interface to these parts without compromising flexibility, for example one on can setpipeline.use_io_bindingto True/False which will set it to all parts, but can also target a specific partpipeline.unet.use_io_bindingdifferently. Parts can also be on different devices.I also added a special feature to diffusion pipelines and especially Unet/Tranformers component so that it reuses the same input tensor as output buffer (instead of creating a new one), this makes the diffusion operation more of an in-place operation, which should make it faster (theoretically), or at least uses less memory.
Before submitting
Who can review?