Jm/hotfix#26
Conversation
- Introduced `convert_img_to_tensor.py` for converting images to preprocessed tensors for calibration. - Updated Korean and English README files to include instructions for using the new tensor conversion script. - Modified `model_compile.py` and other compilation scripts to standardize target device argument handling with choices for `regulus-ra`, `regulus-rb`, and `aries-rb`. - Enhanced documentation for clarity on model preparation, calibration dataset preparation, and compilation steps.
- Refine language and formatting in the README files for oriented bounding boxes, pose estimation, speech-to-text, and vision language model compilation. - Standardize code block formatting for commands across all tutorials. - Clarify prerequisites and steps for model compilation processes. - Enhance explanations of calibration data generation and model compilation stages. - Update target device specifications and notes regarding REGULUS support.
- Refine instructions for running inference with Whisper STT model on Mobilint NPU hardware. - Clarify prerequisites and installation steps for required packages. - Enhance model preparation and inference execution sections for better clarity. - Standardize command line arguments and expected outputs across tutorials. - Update Korean and English README files for consistency and improved readability.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 165c621d78
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… from Qwen2 to Qwen3 in VLM runtime and compilation documentation
|
@mobilint-review Review this PR. |
|
🤖 Received @mobilint-review. The self-hosted Codex agent is responding now. |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Codex review
Requested by @parkjinman98.
Verdict
I found several consistency issues where newly added argparse choices allow regulus-ra for workflows that the PR documentation itself says are unsupported on older REGULUS hardware. These should be tightened before merging so the scripts fail early with the same guidance as the tutorials.
Suggested next steps
- Align each script's
--target-devicechoices with the support matrix documented in the adjacent README. - Re-check the English and Korean README counterparts after the parser changes so the supported device tables remain synchronized.
Trigger: issue_comment
Source: #26 (comment)
|
@mobilint-review Review this PR. |
|
🤖 Received @mobilint-review. The self-hosted Codex agent is responding now. |
There was a problem hiding this comment.
Codex review
Requested by @parkjinman98.
Verdict
Reviewed PR #26. I found one actionable issue: the new optional tensor-calibration documentation describes a path that the checked-in model_compile.py scripts do not currently support as documented.
Suggested next steps
- Add a
model_compile.pyoption that disablespreprocessing_configandUint8InputConfigwhen--calib-data-pathpoints to prepared tensors, or remove the CLI-style tensor path guidance. - Apply the same correction across the matching English and Korean tutorial pages that added Step 2-1/3-1 tensor calibration sections.
Trigger: issue_comment
Source: #26 (comment)
| - `--save-path`: Path to save the MXQ model (onnx -> mxq output) | ||
| - `--mblt-path`: Path to save the MBLT intermediate graph (onnx -> mblt output) | ||
| - `--target-device` (required): Target NPU. See the table below. The inference scheme is derived automatically (ARIES = `all`, REGULUS = `single`). | ||
| If you already prepared `.npy` tensor files, you can use that directory as `calib_data_path` instead of supplying raw image files and a preprocessing pipeline. |
There was a problem hiding this comment.
Prepared tensor path is not wired into model_compile.py
This new optional section says users can point calib_data_path at preprocessed .npy tensors instead of raw images and a preprocessing pipeline, but the actual model_compile.py still always passes preprocessing_config=preprocessing_config and uint8_input_config=Uint8InputConfig(...) into mxq_compile. That means following this guidance with the generated tensor directory either still applies the image preprocessing path or fails depending on how qbcompiler handles .npy input. Please add a real script mode/flag for tensor calibration, or mark this as an API-only code example and not something supported by the documented CLI. The same pattern appears in the other newly added tensor-calibration sections.
2 participants
No description provided.