app.py

#!/usr/bin/env python3
"""
FBMC Chronos-2 Forecasting API
HuggingFace Space Gradio Interface
Version: 1.6.0 - Extended Context Window (2,160 hours = 90 days / 3 months)
FORCE REBUILD: Optimized for 96GB VRAM with memory profiling diagnostics
"""

# CRITICAL: Set PyTorch memory allocator config BEFORE any imports
# This prevents memory fragmentation issues that cause OOM even with sufficient free memory
# Must be set before torch is imported the first time (including via gradio or other dependencies)
import os
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'

# Redirect ALL caches to /tmp to prevent 50GB storage limit exceeded
# This is the most common cause of "no logs" silent failures on A100 Spaces
# See: https://discuss.huggingface.co/t/how-to-fix-workload-evicted-storage-limit-exceeded-50g-error-in-huggingface-spaces/169258
os.environ['TORCH_HOME'] = '/tmp/torch_cache'
os.environ['HF_HOME'] = '/tmp/hf_home'
os.environ['TRANSFORMERS_CACHE'] = '/tmp/transformers_cache'
os.environ['HUB_DIR'] = '/tmp/torch_hub'

import sys
print(f"[STARTUP] Python version: {sys.version}", flush=True)
print(f"[STARTUP] Python path: {sys.path[:3]}", flush=True)
print(f"[STARTUP] PyTorch memory config: {os.environ.get('PYTORCH_CUDA_ALLOC_CONF')}", flush=True)

import gradio as gr
from datetime import datetime

print("[STARTUP] Basic imports successful", flush=True)

try:
    from src.forecasting.chronos_inference import run_inference
    print("[STARTUP] chronos_inference import successful", flush=True)
except Exception as e:
    print(f"[ERROR] Failed to import chronos_inference: {e}", flush=True)
    import traceback
    traceback.print_exc()
    run_inference = None


# Global configuration
FORECAST_TYPES = {
    "smoke_test": "Smoke Test (1 border × 7 days)",
    "full_14day": "Full Forecast (All borders × 14 days)"
}

print("[STARTUP] Configuration loaded", flush=True)


def forecast_api(run_date_str, forecast_type):
    """
    API endpoint for triggering forecasts.

    Args:
        run_date_str: Date in YYYY-MM-DD format
        forecast_type: 'smoke_test' or 'full_14day'

    Returns:
        Path to downloadable forecast results file
    """
    try:
        # Validate run date
        run_date = datetime.strptime(run_date_str, "%Y-%m-%d")

        # Run inference
        result_path = run_inference(
            run_date=run_date_str,
            forecast_type=forecast_type,
            output_dir="/tmp"
        )

        return result_path

    except Exception as e:
        error_msg = f"Error: {str(e)}"
        print(error_msg)
        # Return error message as text file
        error_path = "/tmp/error.txt"
        with open(error_path, 'w') as f:
            f.write(error_msg)
        return error_path


# Build Gradio interface
with gr.Blocks(title="FBMC Chronos-2 Forecasting") as demo:
    gr.Markdown("""
    # FBMC Chronos-2 Zero-Shot Forecasting API

    **Flow-Based Market Coupling** electricity flow forecasting using Amazon Chronos-2.

    This Space provides GPU-accelerated zero-shot inference for cross-border electricity flows.
    """)

    with gr.Row():
        with gr.Column():
            gr.Markdown("### Configuration")

            run_date_input = gr.Textbox(
                label="Run Date (YYYY-MM-DD)",
                value="2025-09-30",
                placeholder="2025-09-30",
                info="Date when forecast is made (data up to this date is historical)"
            )

            forecast_type_input = gr.Radio(
                choices=list(FORECAST_TYPES.keys()),
                value="smoke_test",
                label="Forecast Type",
                info="Smoke test: Quick validation (1 border, 7 days). Full: Production forecast (all borders, 14 days)"
            )

            submit_btn = gr.Button("Run Forecast", variant="primary")

        with gr.Column():
            gr.Markdown("### Results")

            output_file = gr.File(
                label="Download Forecast Results",
                type="filepath"
            )

            gr.Markdown("""
            **Output format**: Parquet file with columns:
            - `timestamp`: Hourly timestamps (D+1 to D+7 or D+14)
            - `{border}_median`: Median forecast (MW)
            - `{border}_q10`: 10th percentile (MW)
            - `{border}_q90`: 90th percentile (MW)

            **Inference environment**:
            - GPU: HuggingFace Space GPU (accelerated inference)
            - Model: amazon/chronos-2 (120M parameters)
            - Precision: bfloat16
            """)

    # Wire up the interface
    submit_btn.click(
        fn=forecast_api,
        inputs=[run_date_input, forecast_type_input],
        outputs=output_file
    )

    gr.Markdown("""
    ---
    ### About

    **Zero-shot forecasting**: No model training required. The pre-trained Chronos-2 model
    generalizes directly to FBMC cross-border flows using historical patterns and future covariates.

    **Features**:
    - 2,514 engineered features using past-only covariate masking
    - Known-future: weather, generation, load forecasts (615 features)
    - Past-only masked: CNEC outages, volatility, flows (1,899 features)
    - 24-month historical context (Oct 2023 - Oct 2025)
    - Time-aware extraction (prevents data leakage)
    - Probabilistic forecasts (9 quantiles: 1st/5th/10th/25th/50th/75th/90th/95th/99th)

    **Performance**:
    - Smoke test: ~30 seconds (1 border × 168 hours)
    - Full forecast: ~5 minutes (38 borders × 336 hours)

    **Project**: FBMC Flow Forecasting MVP | **Author**: Evgueni Poloukarov
    """)

# Launch the app
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )