Expand description
2026-09-25: DeepSeek-V4 weight loader: layers, embedding, final norm and lm_head.
Owner: model-arch weight loader. Invariants:
supports_tpis false: V4 layers shard experts by EP only.load_mtp_weightsreturnsNone; the V4 MTP module is loaded separately bymtp::load_v4_mtp_module.
Modules§
- mtp
- 2026-09-25: Loads the DeepSeek-V4 multi-token-prediction (MTP) draft module
from the
mtp.0.*tensors.