Expand description
2026-09-25: Loads the DeepSeek-V4 multi-token-prediction (MTP) draft module
from the mtp.0.* tensors.
The body is a V4 transformer layer built by assemble_layer from the
mtp.0 prefix. The MTP-specific pieces are the input combiner
(enorm/hnorm, e_proj/h_proj), the final norm and the module’s own
HC head. The token embedding and the LM head are not loaded here: the
proposer, crate::deepseek_v4_mtp::DeepseekV4MtpHead, takes the target
model’s at build time.
Owner: model-arch weight loader. Invariants: none beyond the types.
Structs§
- Deepseek
V4Mtp Module - 2026-09-25: A loaded DeepSeek-V4 MTP draft module: the V4 layer body plus the MTP input combiner, final norm and HC head.
Functions§
- load_
v4_ mtp_ module - 2026-09-25: Loads the DeepSeek-V4 MTP draft module.