granitemoe¶
Model type |
|
Class |
|
Task |
|
Source |
|
Description¶
GraniteMoE model: MoE with Granite scaling multipliers.
Combines MoE routing (TopK-then-softmax gating, per HuggingFace
GraniteMoeTopKGating) with Granite-specific scaling factors:
embedding_multiplier: scales embeddings after lookupattention_multiplier: replaces 1/sqrt(head_dim) attention scalelogits_scaling: divides final logitsresidual_multiplier: scales attention/MLP outputs before residual add
HuggingFace class: GraniteMoeForCausalLM.
Usage¶
mobius build --model <MODEL_ID> output_dir/
from mobius import build
model = build("<MODEL_ID>")