kimi_k2¶
kimi_k2
¶
Kimi K2 on vLLM (model_type kimi_k2): the text-only checkpoints, DeepSeek-V3's family under Moonshot's model type.
vLLM builds a text-only Kimi K2 checkpoint (Kimi-K2-Instruct and the other
DeepseekV3ForCausalLM ones) as its own DeepseekV2ForCausalLM, with the
stack at model.{embed_tokens, layers, norm}: DeepSeek-V3's tree, so its names
and envoys hold unchanged. Kimi K2.5's wrapper, which puts the language model
under language_model, is kimi_k25.