app.py

import gradio as gr
from transformers import PreTrainedTokenizerFast, AutoConfig
from safetensors.torch import load_model  # Import load_model for safetensors
import torch
from EnhancedTransformerModel import EnhancedTransformerModel  # Custom model class

# --- Load Model and Tokenizer ---
MODEL_PATH = "model.safetensors"  # Path to the safetensors model file
CONFIG_PATH = "config.json"       # Path to the model configuration
TOKENIZER_PATH = "tokenizer"      # Path to the tokenizer directory

# Load configuration
config = AutoConfig.from_pretrained(CONFIG_PATH)

# Load tokenizer
tokenizer = PreTrainedTokenizerFast.from_pretrained(TOKENIZER_PATH)

# Initialize the custom model
model = EnhancedTransformerModel(
    vocab_size=config.vocab_size,
    max_seq_length=config.max_position_embeddings,
    d_model=config.hidden_size,
    num_heads=config.num_attention_heads,
    num_layers=config.num_hidden_layers,
    ff_dim=config.intermediate_size,
    dropout=0.1
)

# Load the model weights from safetensors
state_dict = load_model(MODEL_PATH)
model.load_state_dict(state_dict)
model.eval()
model.to("cuda" if torch.cuda.is_available() else "cpu")

# --- Inference Function ---
def generate_text(prompt, max_length=50):
    """
    Generate text based on the input prompt using the trained model.
    """
    # Tokenize the input prompt
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, padding=True, max_length=384)
    input_ids = inputs["input_ids"].to(model.device)
    attention_mask = inputs["attention_mask"].to(model.device)

    # Generate output tokens
    with torch.no_grad():
        outputs = model(input_ids, attention_mask)
        logits = outputs[:, -1, :]  # Get the logits for the last token
        next_token = torch.argmax(logits, dim=-1)  # Greedy decoding

    # Decode the generated token
    generated_text = tokenizer.decode(next_token, skip_special_tokens=True)
    return generated_text

# --- Gradio Interface ---
with gr.Blocks() as demo:
    gr.Markdown("# Snowflake-G0-stable Language Model")
    gr.Markdown("This is an enhanced transformer language model trained on the DialogMLM-50K dataset. Try it out below!")

    with gr.Row():
        input_prompt = gr.Textbox(label="Input Prompt", placeholder="Enter your text here...")
        output_text = gr.Textbox(label="Generated Text")

    submit_button = gr.Button("Generate")

    def on_submit(prompt):
        return generate_text(prompt)

    submit_button.click(on_submit, inputs=input_prompt, outputs=output_text)

# Launch the app
demo.launch()