[model] support hy_v3#53
Conversation
There was a problem hiding this comment.
Code Review
This pull request introduces support for the Hunyuan (hy_v3) model by updating the documentation, adding the model type to the constants, and implementing the HyV3Bridge class. Review feedback suggests refining the HuggingFace weight keys to ensure correct loading and expanding the list of supported model types to improve automatic recognition of standard Hunyuan checkpoints.
|
/gemini review |
There was a problem hiding this comment.
Code Review
This pull request introduces support for the Hunyuan (hy_v3) model by adding the HyV3Bridge class, updating model constants, and expanding the configuration parser to support new parameters such as moe_router_use_sigmoid, router_scaling_factor, and qk_norm. A critical logic error was found in the configuration parser where the moe_router_score_function is unconditionally set to 'sigmoid' if the moe_router_use_sigmoid key exists, regardless of its boolean value.
No description provided.