phi¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
Phi-1/2 causal language model with parallel attention.
Differences from the Llama-style CausalLMModel:
Single-norm parallel residual (like GPT-J, not sequential like Llama)
Non-gated FCMLP instead of gated MLP
Full LayerNorm throughout (not RMSNorm)
LM head has a bias term (
lm_head.bias)Final norm is
model.final_layernorm(matches HF attribute)
Replicates HuggingFace’s PhiForCausalLM.
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")