app.py

import gradio as gr
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# Device configuration (prioritize GPU if available)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "Phearion/bigbrain-v0.0.1"

bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16
        )

# Load models and tokenizer efficiently
config = PeftConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, quantization_config=bnb_config)

# Load the Lora model
model = PeftModel.from_pretrained(model, model_id)

def greet(text):
    with torch.no_grad():
        # Include EOS token for better context
        input_text = f"<s>### User:\n{text}\n\n### Assistant:\n"
        
        batch = tokenizer(input_text, return_tensors='pt', add_special_tokens=True).to(device)

        with torch.cuda.amp.autocast():
            output_tokens = model.generate(
                **batch,
                max_new_tokens=25,  # Limit response length
                do_sample=True,  # Sample from the distribution
                pad_token_id=tokenizer.eos_token_id,  # Stop at EOS
            )

        # Decode only the generated tokens
        response = tokenizer.decode(output_tokens[0][len(batch['input_ids'][0]):], skip_special_tokens=True)

        # Additional stopping condition at next "### Response:"
        response_parts = response.split("### Assistant:")
        return response_parts[0]  # Return only the first part

iface = gr.Interface(fn=greet, inputs="text", outputs="text"
, title="PEFT Model for Big Brain")
iface.launch()  # Share directly to Gradio Space