imagegpt¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
GPT-2 causal language model.
Differences from Llama-style:
Pre-norm with LayerNorm (not RMSNorm)
Absolute positional embeddings (not RoPE)
Combined QKV projection in HF weights (split during preprocess)
Conv1D weights are transposed vs Linear
Tied word embeddings (lm_head = wte)
Replicates HuggingFace’s GPT2LMHeadModel.
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")