feat: Qwen3 VLM support #443
Draft
+259
−41
Conversation
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
* Set execution counts for code cells to reflect their order of execution. * Added HTML output styling for better visualization. * Included detailed output logs for model tracking and warnings related to imports.
* Updated the LocalBackend to include AutoImageProcessor for handling image inputs. * Modified tokenize_trajectory_groups to accept an image processor, enabling image tokenization. * Improved type annotations for better clarity and maintainability. * Adjusted error handling for image processing to ensure robustness.
* Added support for pixel_values and image_grid_thw in PackedTensors and DiskPackedTensors. * Updated packed_tensors_from_tokenized_results to handle new tensor types. * Improved type annotations for clarity and maintainability. * Refactored methods to ensure proper handling of image-related tensors throughout the preprocessing pipeline.
…a notebook * Updated execution counts for code cells to maintain proper order. * Enhanced output handling with detailed HTML and stream outputs for better visualization and tracking. * Refactored tensor handling in preprocessing to ensure correct data types for pixel_values and image_grid_thw.
* Updated the tokenization logic to use an offset for indexing, ensuring correct handling of image tokens. * Improved the efficiency of token replacement in the token_ids list.
* Added conditional loading of LoRA adapter in _set_lora method to handle cases where load_lora is not available. * Updated ModelState initialization to directly assign peft_model if it is already an instance of PeftModelForCausalLM, enhancing type safety and clarity.
* Updated pixel_values and image_grid_thw assignments to conditionally return [None] during warmup, improving flexibility in tensor management. * Ensured consistent handling of tensor data types for better integration in the processing pipeline.
* Introduced a new script to generate images based on yes/no/maybe prompts, utilizing PIL for rendering. * Created a training notebook that integrates OpenAI's API for model training with generated images. * Enhanced the image processing pipeline with improved font handling and text wrapping for better visual output. * Updated the math-vista notebook to reset execution counts and clean up outputs for consistency.
…swering * Added a new script to train a model using image and question pairs from the MathVista dataset. * Integrated asynchronous processing for efficient training and trajectory logging. * Enhanced image handling by saving decoded images to a temporary directory for model input. * Improved argument parsing for customizable training runs.
…ctory logging * Enhanced pixel_values and image_grid_thw assignments to conditionally return [None] during warmup, improving flexibility in tensor management. * Added type ignore comment for content conversion in trajectory_logging to suppress type errors. * Updated test notebook to reflect changes in tokenized results for better accuracy in outputs.
* Bumped versions of `unsloth` to 2025.10.8 and `unsloth-zoo` to 2025.10.9 in `pyproject.toml` and `uv.lock`. * Updated `transformers` version to 4.56.2 in both `pyproject.toml` and `uv.lock`. * Modified the training notebook to change the base model from `Qwen/Qwen2.5-VL-7B-Instruct` to `Qwen/Qwen3-VL-8B-Instruct` and updated execution counts for consistency.
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.
No description provided.