inference.py

import torch
from model import SupernovaNepaliEncoder, SupernovaEncoderConfig

def load_model(model_path='.'):
    config = SupernovaEncoderConfig.from_pretrained(model_path)
    model = SupernovaNepaliEncoder(config)
    # Load weights if necessary, or use from_pretrained
    return model

def get_sana_embeddings(model, input_ids, attention_mask=None):
    model.eval()
    with torch.no_grad():
        # Returns [batch, seq_len, 2304]
        embeddings = model(input_ids, attention_mask=attention_mask)
    return embeddings