Add learning rate config support for Moonshot Kimi models #200
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
joschu
reviewed
Dec 20, 2025
tinker_cookbook/hyperparam_utils.py
Outdated
| "meta-llama/Llama-3.3-70B-Instruct": 8192, | ||
| }[model_name] | ||
|
|
||
| if "moonshotai" in model_name: |
Collaborator
There was a problem hiding this comment.
can you make it check for the full name, in case we add another moonshot model later?
Contributor
Author
There was a problem hiding this comment.
@joschu ah nice catch. Updated this and the DeepSeek PR
bledden
force-pushed
the
fix/moonshot-lr-configs
branch
from
c7119f3 to
b84afb0
Compare
The get_lr() function was raising an AssertionError for moonshotai models because they weren't included in the model name checks. This adds: - Hidden size lookup for Kimi-K2 models (hidden_size=7168) - Learning rate scaling exponent for Kimi (0.0775, same as other MoE models) Related to thinking-machines-lab/tinker-feedback#49
bledden
force-pushed
the
fix/moonshot-lr-configs
branch
from
b84afb0 to
fc8b62d
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Summary
The
get_lr()function inhyperparam_utils.pywas missing support for Moonshot Kimi models, causing it to fail with anAssertionErrorwhen users tried to get learning rate recommendations formoonshotai/Kimi-K2-Thinking.This PR adds the missing configuration:
Uses specific model name matching (
moonshotai/Kimi-K2andkimi-k2) to avoid conflicts if Moonshot releases other model families in the future.Test plan
get_lr("moonshotai/Kimi-K2-Thinking")returns a valid learning rate instead of raising an errorNote on exponent value
The exponent value (0.0775) is based on Qwen's MoE models since Kimi-K2 is also an MoE architecture. The existing exponents for Llama (0.781) and Qwen (0.0775) were empirically derived through hyperparameter sweeps. A follow-up eval run would help validate this is optimal for Kimi-K2 specifically.
Related to thinking-machines-lab/tinker-feedback#49