| ---
|
| license: bsd-3-clause
|
| tags:
|
| - co-attention
|
| - constant-warmup
|
| - generation
|
| - giant
|
| - groupnorm
|
| - kaiming
|
| - lamb
|
| - multi-query
|
| - relu
|
| - vit
|
| ---
|
|
|
| # model_035528848_vit_giant.py
|
|
|
| ## Model Overview
|
|
|
| A **giant**-scale implementation of the **vit** architecture, built for **generation** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: vit
|
| - **Scale**: giant
|
| - **Attention**: multi query
|
| - **Fusion strategy**: co attention
|
| - **Task head**: generation
|
| - **Activation**: relu
|
| - **Normalization**: groupnorm
|
| - **Initialization**: kaiming
|
|
|
| ## Training
|
|
|
| - **Optimizer**: lamb
|
| - **LR scheduler**: constant warmup
|
|
|
| ## Files
|
|
|
| - `model_035528848_vit_giant.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
