hunyuan_v1_moe¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
HunYuan MoE V1 causal language model.
Extends the generic MoE model with:
QK-norm: per-head RMSNorm on Q and K projections after linear projection.
Ungated shared expert (
shared_mlp): runs on every token, added to routed expert output without sigmoid gating.Custom gate naming: HF uses
gate.wg.weight; we usegate.weight.
HuggingFace class: HunYuanMoEV1ForCausalLM
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")