Gemma3-4b QNN example fixes

[qti-kromero]

Describe your changes

• Updates various handlers to handle Gemma3
• Adds custom module/custom class fields to HF model
• Olive recipe for Gemma3-4b

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.
• Is this PR including examples changes? If yes, please remember to update example documentation in a follow-up PR.

[xiaoyu-work]

Did you test the inference of the output model? I tested https://huggingface.co/google/gemma-3-270m but it failed with onnxruntime-genai when inferencing (this will be fixed soon), but I'm not sure if you got an error as well for Gemma3-4b with qnn.

[qti-kromero]

Did you test the inference of the output model? I tested https://huggingface.co/google/gemma-3-270m but it failed with onnxruntime-genai when inferencing (this will be fixed soon), but I'm not sure if you got an error as well for Gemma3-4b with qnn.

We have intelligible output from the text portion of the model within our KPIs but we have not successfully prepared the vision portion of the model or combined the two models together in a pipeline of OGA. Do you have details on what specific portion of the model you had issues with?

[xiaoyu-work]

Did you test the inference of the output model? I tested https://huggingface.co/google/gemma-3-270m but it failed with onnxruntime-genai when inferencing (this will be fixed soon), but I'm not sure if you got an error as well for Gemma3-4b with qnn.

We have intelligible output from the text portion of the model within our KPIs but we have not successfully prepared the vision portion of the model or combined the two models together in a pipeline of OGA. Do you have details on what specific portion of the model you had issues with?

I encountered a bug in model builder when testing gemma-3-270m, but it has now been fixed. Interestingly the 4B model works without issues, maybe there are some difference between their configs. If it is tested, I'm fine. Thanks.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[qti-kromero]

apparently I have to make a comment

[qti-kromero]

@jambayk @xiaoyu-work wondering if this could get reviewed when you have a chance

[xiaoyu-work]

Hi @qti-kromero , is this changed required for recipe microsoft/olive-recipes#125?

[qti-kromero]

Hi @qti-kromero , is this changed required for recipe microsoft/olive-recipes#125?

correct, although we may need to retest these changes since we haven't revisited them for awhile due to higher priority work

[xiaoyu-work]

Hi @qti-kromero , is this changed required for recipe microsoft/olive-recipes#125?

correct, although we may need to retest these changes since we haven't revisited them for awhile due to higher priority work

Thanks for confirming this! Let me know if they are ready for review and I will review both PRs together.

[jambayk]

Closing for now since the PR has become stale