Enhance embedding introspection for GenAI config - #321
Merged
Conversation
Performance Comparison
|
|
The author of this PR, apsonawane, is not an activated member of this organization on Codecov. |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR addresses Gemma4 multimodal ORT GenAI export issues by aligning vision input dtype handling, ensuring embedding outputs are represented in genai_config, and passing the selected execution provider through the example export flow.
Changes:
- Declares Gemma4 vision
pixel_valuesas float32 and casts internally to model dtype. - Adds embedding output-name overrides and graph-output introspection for ORT GenAI config generation.
- Propagates
--devicetomobius build --epin the Gemma4 example script.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
src/mobius/tasks/_gemma4.py |
Changes Gemma4 vision encoder pixel_values input dtype to float32. |
src/mobius/models/gemma4.py |
Casts float32 vision inputs to model dtype inside the vision encoder. |
src/mobius/integrations/ort_genai/genai_config.py |
Adds embedding_output_names support to vision config generation. |
src/mobius/integrations/ort_genai/auto_export.py |
Introspects embedding model outputs and forwards them into genai_config. |
examples/gemma4_genai.py |
Passes the selected device as the Mobius CLI execution provider. |
Member
|
Wasn't the decision to cast inputs in genai? I just merged a PR for that behavior. Or is there a technical difficulty with vision (overflow etc. )? |
…d pass EP to export script
justinchuby
approved these changes
May 28, 2026
justinchuby
left a comment
justinchuby
left a comment
Member
There was a problem hiding this comment.
Approving to unblock. LMK the state of the input contract
apsonawane
force-pushed
the
asonawane/gemma
branch
from
May 28, 2026 17:58
de60ba3 to
0a9feb0
Compare
Contributor
Author
|
yes, sorry for the confusion. Casting is happening in genai. I reverted it here |
justinchuby
approved these changes
May 29, 2026
justinchuby
enabled auto-merge (squash)
May 29, 2026 14:48
justinchuby
changed the title
justinchuby
disabled auto-merge
May 29, 2026 14:52
justinchuby
enabled auto-merge (squash)
May 29, 2026 14:52
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.
Summary
Two fixes for Gemma4 multimodal export and genai_config generation.
Introspect embedding model outputs in genai_config.json
The Gemma4 embedding model produces
per_layer_inputsalongsideinputs_embeds, butgenai_config.jsononly listedinputs_embedsin the embedding outputs section. This caused onnxruntime-genai to not bind theper_layer_inputsoutput, so ORT allocated a separate buffer and the decoder received uninitialized data.Added
_introspect_outputs()to auto-discover embedding model outputs from the ONNX graph, andembedding_output_namesparameter toGenaiConfigGenerator.with_vision().Files:
integrations/ort_genai/auto_export.py,integrations/ort_genai/genai_config.pyPass
--epto mobius CLI in export scriptbuild_and_export()ingemma4_genai.pydid not pass the execution provider tomobius build. This caused all exports to use the default EP (which has emptygqa_dtypes), so KV-shared layers fell back to standard Attention instead of GQA — leading topast_present_share_bufferfailures at runtime.Now
--device cudacorrectly maps to--ep cuda, producing all-GQA models.Files:
examples/gemma4_genai.pyTesting