Module deepseek_v4 

2026-09-25: DeepSeek-V4 weight loader: layers, embedding, final norm and lm_head.

Owner: model-arch weight loader. Invariants:

• supports_tp is false: V4 layers shard experts by EP only.
• load_mtp_weights returns None; the V4 MTP module is loaded separately by mtp::load_v4_mtp_module.

Modules

mtp

2026-09-25: Loads the DeepSeek-V4 multi-token-prediction (MTP) draft module from the mtp.0.* tensors.

Structs

DeepSeekV4WeightLoader