app.py

import gradio as gr
import pandas as pd
from datasets import Dataset
from transformers import AutoTokenizer, TrainingArguments, Trainer, AutoModelForCausalLM
import torch
import zerogpu

# Initialize ZeroGPU
zerogpu.initialize()

def train_model(file, hf_token):
    try:
        # Basic data loading test
        df = pd.read_csv(file.name)
        print(f"Loaded CSV with {len(df)} rows")
        
        # Load tokenizer and model
        model_name = "facebook/opt-125m"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            device_map=None,  # Force simple device mapping
            low_cpu_mem_usage=True
        )
        model = model.to('cpu')  # Explicitly move to CPU
        
        # Basic dataset creation
        dataset = Dataset.from_pandas(df)
        
        args = TrainingArguments(
            output_dir="./results",
            per_device_train_batch_size=1,
            num_train_epochs=1,
            no_cuda=True,
            local_rank=-1
        )
        
        trainer = Trainer(
            model=model,
            args=args,
            train_dataset=dataset,
            tokenizer=tokenizer
        )
        
        return f"Setup successful! Loaded {len(df)} rows"
        
    except Exception as e:
        return f"Error: {str(e)}\nType: {type(e)}"

demo = gr.Interface(
    fn=train_model,
    inputs=[
        gr.File(label="Upload CSV file"),
        gr.Textbox(label="HF Token", type="password")
    ],
    outputs="text",
    title="Product Classifier Training (CPU)",
)

if __name__ == "__main__":
    demo.launch(debug=True)  # Enable debug mode