modernbert-decoder¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
ModernBERT decoder (causal LM variant).
Extends CausalLMModel with:
Embedding LayerNorm (via _ModernBertDecoderTextModel)
Three-component LM head: dense → LayerNorm → decoder (with bias)
Replicates HuggingFace ModernBertForCausalLM.
HF weight naming differences vs standard CausalLMModel:
Separate q_proj/k_proj/v_proj (not fused Wqkv)
Fused MLP Wi → split into gate_proj + up_proj
Custom layer-norm names (attn_norm / mlp_norm / final_norm)
Embedding LayerNorm at model.embeddings.norm
Three-part LM head (lm_head.dense + lm_head.norm + decoder)
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")