The entire file is source L1–L1:
from .stable_latent_moe import StableLatentMoE
This enables the decoder to import StableLatentMoE from kimi_k3.moe. The class, its QuantileBalancer, grouped expert dispatch and mutable balancing buffers are implemented in moe/stable_latent_moe.py. The initializer creates no experts, makes no routing choices and declares no __all__.