[AMD] Fix ZeroDivisionError in compute_normalized_prompt_logprobs#14721
[AMD] Fix ZeroDivisionError in compute_normalized_prompt_logprobs#14721sunxxuns wants to merge 1 commit intosgl-project:mainfrom
Conversation
On AMD/ROCm, input_token_logprobs can be empty in some cases, causing ZeroDivisionError when computing normalized prompt logprobs for the select operation. Changes: 1. Return -inf when no valid logprobs (choice should not be selected) 2. Skip token healing adjustment if no logprobs available 3. Guard against division by zero when num_tokens <= 1 This fixes test_hellaswag_select and test_select failures on AMD.
Summary of ChangesHello @sunxxuns, 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 resolves a critical 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 successfully addresses a ZeroDivisionError in compute_normalized_prompt_logprobs and select. The changes are generally good, but there's a logic issue in the select method's fix. When only one token is present for token healing, the code avoids the crash but leaves normalized_prompt_logprobs in an inconsistent state with a stale value. I've provided a suggestion to set the logprob to negative infinity in this case, which aligns with how empty logprobs are handled elsewhere in the code.
| if num_tokens > 1: | ||
| normalized_prompt_logprobs[i] = ( | ||
| normalized_prompt_logprobs[i] * num_tokens - healed_token_logprob | ||
| ) / (num_tokens - 1) |
There was a problem hiding this comment.
While this change correctly prevents a ZeroDivisionError when num_tokens is 1, it leaves normalized_prompt_logprobs[i] with a stale value. After the healed token is removed on line 295, input_token_logprobs[i] becomes empty, but the corresponding normalized_prompt_logprobs[i] is not updated, leading to an inconsistent state. To align with the logic in compute_normalized_prompt_logprobs which returns -inf for empty logprobs, it would be more correct to set normalized_prompt_logprobs[i] to -inf when num_tokens is 1.
| if num_tokens > 1: | |
| normalized_prompt_logprobs[i] = ( | |
| normalized_prompt_logprobs[i] * num_tokens - healed_token_logprob | |
| ) / (num_tokens - 1) | |
| if num_tokens > 1: | |
| normalized_prompt_logprobs[i] = ( | |
| normalized_prompt_logprobs[i] * num_tokens - healed_token_logprob | |
| ) / (num_tokens - 1) | |
| else: | |
| normalized_prompt_logprobs[i] = float("-inf") |
1 participant
No description provided.