app.py

import gradio as gr
import subprocess
import os
import threading
import time

# --- Configuration ---
# IMPORTANT: Change this to your Hugging Face username
YOUR_HF_USERNAME = "Scropo"
# --------------------

HUB_MODEL_ID = f"{YOUR_HF_USERNAME}/gpt-oss-20b-mentalchat-finetuned"
OUTPUT_DIR = "./gpt-oss-20b-mentalchat-finetuned"

def get_training_command():
    """Builds the accelerate launch command."""
    hf_token = os.getenv("HF_TOKEN")
    if not hf_token:
        raise ValueError("HF_TOKEN secret is not set in the Space settings!")

    # This command is equivalent to the one you'd run in a terminal
    command = [
        "accelerate", "launch",
        "sft.py",
        "--model_name_or_path", "openai/gpt-oss-20b",
        "--dataset_name", "ShenLab/MentalChat16K",
        "--output_dir", OUTPUT_DIR,
        "--max_seq_length", "2048",
        "--num_train_epochs", "1",
        "--per_device_train_batch_size", "1",
        "--gradient_accumulation_steps", "8", # Adjust based on GPU memory
        "--learning_rate", "2e-5",
        "--logging_steps", "5",
        "--push_to_hub", "true",
        "--hub_model_id", HUB_MODEL_ID,
        "--hub_token", hf_token,
        "--peft_lora_r", "64",
        "--peft_lora_alpha", "16",
        "--bf16", "true", # Use bf16 for A100/H100
        "--gradient_checkpointing", "true",
    ]
    return command

def run_training():
    """Runs the training command and streams output."""
    command = get_training_command()
    yield "🚀 Starting training job...\n"
    yield f"Running command: {' '.join(command)}\n\n"

    # Use Popen to start the process and capture output in real-time
    process = subprocess.Popen(
        command,
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
        text=True,
        bufsize=1,
        universal_newlines=True,
    )

    # Stream the output
    for line in process.stdout:
        yield line

    process.wait()

    if process.returncode == 0:
        yield f"\n🎉 Training finished successfully! Model pushed to {HUB_MODEL_ID}"
    else:
        yield f"\n❌ Training failed with exit code {process.returncode}."

# --- Gradio Interface ---
with gr.Blocks() as demo:
    gr.Markdown("# 🚀 GPT-20B Supervised Fine-Tuning")
    gr.Markdown(
        "**Model:** `openai/gpt-oss-20b`\n\n"
        "**Dataset:** `ShenLab/MentalChat16K`\n\n"
        "Click the button below to launch the fine-tuning process. "
        "This will take a very long time. Monitor the logs for progress."
    )

    start_button = gr.Button("Start Fine-Tuning Job")
    log_output = gr.Textbox(
        label="Training Logs",
        lines=30,
        interactive=False,
        autoscroll=True
    )

    start_button.click(
        fn=run_training,
        inputs=[],
        outputs=[log_output],
    )

demo.queue().launch()