[Python][Relax] Fix YaRN correction dim calculation

[akaashrp]

Replace expressions with TIR constructs

[gemini-code-assist]

Summary of Changes

Hello @akaashrp, 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 addresses an issue in the YaRN correction dimension calculation within the TVM Relax frontend. The primary goal is to align the mathematical operations with TVM's Tensor IR (TIR) framework by replacing native Python math functions with their TIR equivalents. This modification is crucial for ensuring that these calculations are correctly represented and optimized when compiled by TVM, thereby improving the overall robustness and performance of the YaRN position embedding implementation.

Highlights

• YaRN Correction Dimension Calculation: The yarn_find_correction_dim function has been updated to replace standard Python math.log calls with TVM's Tensor IR (TIR) constructs, specifically tir.log and tir.const. This change ensures that the numerical operations are properly integrated into the TVM compilation graph.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The 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 /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

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 .gemini/ folder in the base of the repository. Detailed instructions can be found here.

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

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request refactors the yarn_find_correction_dim function to use TIR constructs (tir.log, tir.const) instead of Python's math functions. This change ensures that the logarithmic computations are represented within the TIR expression graph, which is the correct approach for building symbolic computations in TVM. The change is correct and improves the consistency of the codebase. I have no further comments.

[Liberatedwinner]

I noticed an issue with this change.
When using Metal, I encountered the following error:
tvm.error.InternalError: Fail to compile metal source:program_source:20:61: error: call to 'log' is ambiguous
Could you take a look?

[akaashrp]

@Liberatedwinner are there specific models or commands for which you run into this issue?

[Liberatedwinner]

Please run the code file I attached.
test_tir_log.py

The result of test code is:

============================================================
PR version (tir.log)
============================================================
low: T.max(T.float32(128.0) * T.log(T.float32(20.371832715762604)) / (T.float32(2.0) * T.log(T.float32(10000.0))), T.float32(0.0))
high: T.min(T.float32(128.0) * T.log(T.float32(651.89864690440334)) / (T.float32(2.0) * T.log(T.float32(10000.0))), T.float32(127.0))

============================================================
Original version (math.log)
============================================================
low: T.float32(20.944481620636051)
high: T.float32(45.026881273754547)

Using tir.log instead of math.log in yarn_find_correction_dim causes Metal codegen to fail with call to 'log' is ambiguous.

tir.log creates a symbolic expression (e.g., T.log(T.float32(20.3718...))) that remains in the IR and gets compiled into Metal shader code.
Metal Shading Language has multiple overloads for math functions (float, half, etc.), and the compiler cannot resolve which one to use.
See: Metal Shading Language Specification (Section 6.5 "Math Functions" and Table 6.4)

[akaashrp]

I'm able to see that the outputs look different for the PR version and the original version. However, I'm also able to compile models that use YaRN without issue on metal. Could you point me to a codegen path I could run so that I can arrive at the call to 'log' is ambiguous error you mentioned?

[Liberatedwinner]

Would you mind checking the error with this file?
test_tir_log.py

[akaashrp]

@Liberatedwinner, I apologize for the delay. To deal with the merge conflicts, I created a new PR that should address the issues on Metal by precomputing the denominator: #18661.

FWIW, I ran into a Metal error (as expected) when initially testing with your file but am getting an unrelated MLC-LLM pipeline error now (both before and after changes).