gpt_oss¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
GPT-OSS causal language model with MoE FFN and attention sinks.
Architecture highlights:
Alternating sliding/full attention layers (
config.layer_types)Attention sinks: learned per-head scalar extends softmax denominator
GQA with YaRN RoPE and attention projection biases
MoE FFN: top-k routing with softmax scores and additive router bias
Custom activation: (up+1) * silu_alpha(gate), alpha=1.702
HuggingFace model_type: gpt_oss.
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")