Wiola-360M

Wiola-360M is a 360 million parameter decoder-only language model developed by OSCOWL AI.

Architecture

Wiola introduces several architectural improvements over a standard Transformer decoder:

Grouped Query Attention (GQA)
Spiral Rotary Positional Encoding (SRPE)
Adaptive Token Merging (ATM)
Gated Cross Layer Attention (GCLA)
Dual Stream Feed Forward (DSFF)

Model

Parameters: 360M
Context Length: 2048
Vocabulary Size: 32000
Attention Heads: 16
Key/Value Heads: 4

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "oscowlai/Wiola360M",
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    "oscowlai/Wiola360M",
    trust_remote_code=True,
)

License

Apache-2.0

Downloads last month: 17

Safetensors

Model size

0.3B params

Tensor type

F32