Refactor InferenceModel
#485
Conversation
Summary of ChangesHello @chenyushuo, 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 refactors the inference model architecture by introducing a new 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 solid refactoring by creating a BaseInferenceModel to abstract away common logic from TinkerModel and vLLMRolloutModel. This change significantly reduces code duplication and enhances maintainability. The implementation is well-executed. I have a couple of minor suggestions to further improve the readability of the new base class.
| token_ids = token_ids[: self.config.max_prompt_tokens + 1] # leave one for response | ||
| return [ | ||
| Experience( | ||
| tokens=token_ids, | ||
| logprobs=torch.zeros(1, dtype=torch.float32), | ||
| prompt_length=len(token_ids) - 1, | ||
| prompt_text=self.tokenizer.decode(token_ids[:-1]), | ||
| response_text=self.tokenizer.decode(token_ids[-1]), | ||
| truncate_status="prompt_truncated", | ||
| reward=0.0, | ||
| ) | ||
| for _ in range(kwargs.get("n", 1)) | ||
| ], False |
There was a problem hiding this comment.
The logic for handling prompt truncation and creating a dummy Experience object is a bit dense. Using an explicit variable for the prompt token count would make the code more self-documenting and easier to understand at a glance.
| token_ids = token_ids[: self.config.max_prompt_tokens + 1] # leave one for response | |
| return [ | |
| Experience( | |
| tokens=token_ids, | |
| logprobs=torch.zeros(1, dtype=torch.float32), | |
| prompt_length=len(token_ids) - 1, | |
| prompt_text=self.tokenizer.decode(token_ids[:-1]), | |
| response_text=self.tokenizer.decode(token_ids[-1]), | |
| truncate_status="prompt_truncated", | |
| reward=0.0, | |
| ) | |
| for _ in range(kwargs.get("n", 1)) | |
| ], False | |
| prompt_token_count = self.config.max_prompt_tokens | |
| token_ids = token_ids[:prompt_token_count + 1] | |
| return [ | |
| Experience( | |
| tokens=token_ids, | |
| logprobs=torch.zeros(1, dtype=torch.float32), | |
| prompt_length=prompt_token_count, | |
| prompt_text=self.tokenizer.decode(token_ids[:-1]), | |
| response_text=self.tokenizer.decode(token_ids[-1]), | |
| truncate_status="prompt_truncated", | |
| reward=0.0, | |
| ) | |
| for _ in range(kwargs.get("n", 1)) | |
| ], False |
| if self.config.max_model_len is not None and self.config.max_model_len > 0: | ||
| if len(token_ids) > self.config.max_model_len - 1: | ||
| truncate_status = "response_truncated" | ||
| self.logger.warning( | ||
| f"Warning: {len(token_ids)=} exceeds the length limit {(self.config.max_model_len - 1)=}" | ||
| ) | ||
| token_ids = token_ids[: self.config.max_model_len - 1] | ||
| action_mask = action_mask[: self.config.max_model_len - 1] |
There was a problem hiding this comment.
The use of max_model_len - 1 is a bit of a magic number. Introducing a local variable for the maximum length and adding a comment to explain the -1 would improve code clarity and maintainability, making it easier for future developers to understand the rationale behind this truncation logic.
| if self.config.max_model_len is not None and self.config.max_model_len > 0: | |
| if len(token_ids) > self.config.max_model_len - 1: | |
| truncate_status = "response_truncated" | |
| self.logger.warning( | |
| f"Warning: {len(token_ids)=} exceeds the length limit {(self.config.max_model_len - 1)=}" | |
| ) | |
| token_ids = token_ids[: self.config.max_model_len - 1] | |
| action_mask = action_mask[: self.config.max_model_len - 1] | |
| if self.config.max_model_len is not None and self.config.max_model_len > 0: | |
| # The -1 is to leave space for at least one token to be generated for logprobs calculation. | |
| max_len = self.config.max_model_len - 1 | |
| if len(token_ids) > max_len: | |
| truncate_status = "response_truncated" | |
| self.logger.warning( | |
| f"Warning: {len(token_ids)=} exceeds the length limit {max_len=}" | |
| ) | |
| token_ids = token_ids[:max_len] | |
| action_mask = action_mask[:max_len] |
|
/unittest-module-common |
Summary
Skipped
Tests
Github Test Reporter by CTRF 💚 |
2 participants
Description
As the title says.
Checklist
Please check the following items before code is ready to be reviewed.