Conversation
Summary of ChangesHello @lukehinds, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request significantly expands DeepFabric's capabilities by introducing a complete training pipeline that integrates seamlessly with HuggingFace Transformers. Users can now leverage local LLMs for generating synthetic datasets and then fine-tune these models using advanced techniques like LoRA and quantization, all within a unified framework. This enhancement provides greater control and flexibility for users who prefer local model operations and end-to-end model development. Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a significant new feature: a complete training pipeline with support for local HuggingFace Transformers models. The changes include a new DeepFabricPipeline for high-level workflows, configuration models for SFT, LoRA, and quantization, and a DeepFabricSFTTrainer that integrates with TRL. The implementation is robust, with good error handling, dependency validation, and resource management for local models. The PR also adds comprehensive documentation, examples, and notebooks to guide users.
My review focuses on improving maintainability and performance in the new modules. I've suggested refactoring a conditional block to use a dictionary lookup for clarity, moving repeated imports out of a frequently called function to improve performance, and extracting a hardcoded multi-line string into a constant. Overall, this is a well-executed and substantial addition to the library.
| def transformers_callable(prompt: str, schema: type[BaseModel], **gen_kwargs): | ||
| # Import generator module to create logits processor | ||
| from outlines import generator # noqa: PLC0415 | ||
|
|
||
| # Create JSON schema string from Pydantic model | ||
| json_schema_str = schema.model_json_schema() | ||
| # Get JSON string representation | ||
| import json # noqa: PLC0415 | ||
|
|
||
| json_schema_json = json.dumps(json_schema_str) | ||
|
|
||
| # Convert to logits processor for transformers | ||
| logits_processor = generator.get_json_schema_logits_processor( | ||
| None, # backend_name (None for auto-detect) | ||
| outlines_model, | ||
| json_schema_json, | ||
| ) | ||
| # Generate using the model's generate method | ||
| return outlines_model.generate(prompt, output_type=logits_processor, **gen_kwargs) | ||
|
|
There was a problem hiding this comment.
For performance, these imports (outlines.generator and json) should be moved outside the transformers_callable function. Since this function will be called for every generation, the imports will be executed repeatedly, adding unnecessary overhead. They can be moved to the top of the if provider == 'transformers': block, for instance after line 161.
| if self.config.torch_dtype: | ||
| if self.config.torch_dtype == "auto": | ||
| model_kwargs["torch_dtype"] = "auto" | ||
| elif self.config.torch_dtype == "float16": | ||
| model_kwargs["torch_dtype"] = torch.float16 | ||
| elif self.config.torch_dtype == "bfloat16": | ||
| model_kwargs["torch_dtype"] = torch.bfloat16 | ||
| elif self.config.torch_dtype == "float32": | ||
| model_kwargs["torch_dtype"] = torch.float32 |
There was a problem hiding this comment.
This if/elif chain can be simplified by using a dictionary to map the string torch_dtype values to their corresponding torch objects. This would make the code more readable and easier to extend if more dtypes are supported in the future.
| if self.config.torch_dtype: | |
| if self.config.torch_dtype == "auto": | |
| model_kwargs["torch_dtype"] = "auto" | |
| elif self.config.torch_dtype == "float16": | |
| model_kwargs["torch_dtype"] = torch.float16 | |
| elif self.config.torch_dtype == "bfloat16": | |
| model_kwargs["torch_dtype"] = torch.bfloat16 | |
| elif self.config.torch_dtype == "float32": | |
| model_kwargs["torch_dtype"] = torch.float32 | |
| if self.config.torch_dtype: | |
| dtype_map = { | |
| "auto": "auto", | |
| "float16": torch.float16, | |
| "bfloat16": torch.bfloat16, | |
| "float32": torch.float32, | |
| } | |
| if self.config.torch_dtype in dtype_map: | |
| model_kwargs["torch_dtype"] = dtype_map[self.config.torch_dtype] |
| self.tokenizer.chat_template = ( | ||
| "{% for message in messages %}" | ||
| "{{ '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>\n' }}" | ||
| "{% endfor %}" | ||
| "{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" | ||
| ) |
There was a problem hiding this comment.
This hardcoded ChatML template string could be defined as a module-level constant for better readability and to avoid having a multi-line string in the middle of a method. For example: DEFAULT_CHATML_TEMPLATE = "..."
1 participant
No description provided.