Refactor IO config to yaml in assets

[xiaoyu-work]

This pull request introduces a major refactor of the HuggingFace ONNX I/O configuration system in Olive. The changes move away from hardcoded Python class registries and instead adopt a YAML-driven approach for specifying input/output configurations for both tasks and diffusers pipelines. This enables easier extensibility and clearer, data-driven configuration management. Additionally, the update adds default dummy input shapes and rich configuration templates for tasks and diffusers components.

The most important changes are:

YAML-based IO Configuration System:

• Added new YAML files (defaults.yaml, tasks.yaml, diffusers.yaml) in olive/assets/io_configs/ that define default dummy input shapes, task templates, and diffusers component/pipeline specifications for ONNX export. These files replace the previous Python-based registries and provide a more flexible and maintainable configuration system. [1] [2] [3]

Refactor of IO Config Python Interface:

• Replaced the contents of olive/common/hf/io_config/__init__.py to remove all hardcoded class registries and mapping logic, and instead import new YAML-driven utility functions (generate_dummy_inputs, get_io_config, get_diffusers_io_config, etc.). This greatly simplifies the interface and shifts responsibility to the new YAML configuration system.

Legal and Organizational Improvements:

• Added copyright and license headers to olive/assets/__init__.py and olive/assets/io_configs/__init__.py.## Describe your changes

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

(Optional) Issue link

[Copilot]

Pull request overview

This pull request introduces a major refactor of the HuggingFace ONNX I/O configuration system, migrating from hardcoded Python class registries to a flexible YAML-driven approach. The changes enable easier extensibility and clearer configuration management for model optimization workflows.

Changes:

• Replaced Python-based OnnxConfig class registries with YAML configuration files for tasks and diffusers components
• Introduced unified dummy input generators that read specifications from YAML rather than class hierarchies
• Consolidated I/O configuration generation into a single task-driven API with automatic optional input filtering based on model signatures

Reviewed changes

Copilot reviewed 29 out of 34 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
olive/assets/io_configs/tasks.yaml	Defines input/output specifications for 15+ HuggingFace tasks
olive/assets/io_configs/diffusers.yaml	Defines I/O specs for diffusers components (SD, SDXL, SD3, Flux, Sana)
olive/assets/io_configs/defaults.yaml	Provides default dimension values and attribute aliases for config resolution
olive/common/hf/io_config/yaml_loader.py	YAML loading and config resolution utilities with LRU caching
olive/common/hf/io_config/task_config.py	New unified API for generating I/O configs and dummy inputs
olive/common/hf/io_config/input_generators.py	Refactored to use YAML-driven input generation
olive/common/hf/model_io.py	Updated to use new task-driven API
olive/common/hf/wrapper.py	Migrated to use resolve_alias for config attribute access
test/common/hf/io_config/*.py	Updated tests for new YAML-driven system

Comments suppressed due to low confidence (1)

olive/common/hf/io_config/task_config.py:1

• Missing check for whether 'with_past' key exists in task_spec before accessing it. This will raise a KeyError if a task doesn't define with_past overrides. Add a check: if self.use_past_in_inputs and 'with_past' in self.task_spec:

# -------------------------------------------------------------------------

[shaahji]

The previous entries include the key in the value i.e. name is included in the list of aliases. These two entries don't. Should they? Or should the above 4 be updated to exclude the key name?

[xiaoyu-work]

Good call! No need to put itself in the list. I'll update this.

[shaahji]

Same as above!

[shaahji]

Same problem here too. height and weight are not included in the list of values.

[shaahji]

Rename this file something more specific - yaml_loader is too generic and says nothing about the actual functionality it provides.

YAML implementation is anyways private to the file. Public API seems more like utility functions.

[xiaoyu-work]

updated to io_resolver.

[shaahji]

Wouldn't importlib require user to create a whl and install the package for the system to pickup the change to config files.
(Unless of course if the user has the cloned repo in editable mode.)

[xiaoyu-work]

Yes, we move assets folder into olive package, so this change is for developers. If user has its own io config, they can specify it in the config file.

[shaahji]

Suggestion: Current implementation requires user to clone the git repo and modify it. This approach has challenges, most importantly dev vs. released version. Instead, rather than piling all the configuration in a single file, have one file hold configuration for one task. This will allow you to accept new task configuration on command line without explicitly modifying cloned repo. This approach also allows working with released version - which is always recommended compared to using a git clone.

P.S. That's the approach both lm-eval and bb implements.

[xiaoyu-work]

Suggestion: Current implementation requires user to clone the git repo and modify it. This approach has challenges, most importantly dev vs. released version. Instead, rather than piling all the configuration in a single file, have one file hold configuration for one task. This will allow you to accept new task configuration on command line without explicitly modifying cloned repo. This approach also allows working with released version - which is always recommended compared to using a git clone.

P.S. That's the approach both lm-eval and bb implements.

We now have 22 tasks, having separate config files for each task makes maintenance difficult. Most use cases would be adding new task IO rather than modifying existing ones. How about keeping the current configuration files, but also allowing override with external config files?

[shaahji]

Suggestion: Current implementation requires user to clone the git repo and modify it. This approach has challenges, most importantly dev vs. released version. Instead, rather than piling all the configuration in a single file, have one file hold configuration for one task. This will allow you to accept new task configuration on command line without explicitly modifying cloned repo. This approach also allows working with released version - which is always recommended compared to using a git clone.
P.S. That's the approach both lm-eval and bb implements.

We now have 22 tasks, having separate config files for each task makes maintenance difficult. Most use cases would be adding new task IO rather than modifying existing ones. How about keeping the current configuration files, but also allowing override with external config files?

How often do we expect to change any of these ones they are committed? For all intents and purposes, this is readonly data, unlikely to change ever. But if it makes it simpler, all for it.

On the implementation front, you could basically assume any number of tasks per file. So, there shouldn't be any difference whether you are loading the Olive released or the user provided one. Basically, the code path should remain the same.

[xiaoyu-work]

Suggestion: Current implementation requires user to clone the git repo and modify it. This approach has challenges, most importantly dev vs. released version. Instead, rather than piling all the configuration in a single file, have one file hold configuration for one task. This will allow you to accept new task configuration on command line without explicitly modifying cloned repo. This approach also allows working with released version - which is always recommended compared to using a git clone.
P.S. That's the approach both lm-eval and bb implements.

We now have 22 tasks, having separate config files for each task makes maintenance difficult. Most use cases would be adding new task IO rather than modifying existing ones. How about keeping the current configuration files, but also allowing override with external config files?

How often do we expect to change any of these ones they are committed? For all intents and purposes, this is readonly data, unlikely to change ever. But if it makes it simpler, all for it.

On the implementation front, you could basically assume any number of tasks per file. So, there shouldn't be any difference whether you are loading the Olive released or the user provided one. Basically, the code path should remain the same.

i think in real life, it is rare that user wants to provide his own io config for his own task. if a task is not supported by Olive, then we should add support it.