ucast_inference.py

"""U-Cast (#10467) rescue: run the REAL ERA5 forecasting benchmark on Modal.

Uses the paper's standalone inference script with the released checkpoint,
streams ERA5 from public WeatherBench2 GCS, computes RMSE + CRPS (--score).
This targets U-Cast's actual performance claim, not just its param count.
"""
import modal

REPO = "u-cast"
image = (modal.Image.debian_slim(python_version="3.11")
         .pip_install("torch", "xarray", "netCDF4", "zarr<3", "einops", "tqdm",
                      "pyyaml", "huggingface_hub", "gcsfs", "numpy", "scipy", "dask")
         .add_local_dir(REPO, f"/root/{REPO}", copy=True))
app = modal.App("ucast-inference", image=image)


@app.function(gpu="A10G", timeout=3600)
def run_infer(ic_date="2020-01-01", ensemble=5, horizon=10):
    import subprocess, os, re
    os.chdir(f"/root/{REPO}")
    cmd = [
        "python", "run_inference_standalone.py",
        "--ckpt-path", "hf:salv47/u-cast/ucast.ckpt",
        "--data-dir", "gs://weatherbench2/datasets/era5",
        "--config-path", "configs/config_inference.yaml",
        "--ic-start-dates", ic_date,
        "--ensemble-size", str(ensemble),
        "--prediction-horizon", str(horizon),
        "--score",
    ]
    p = subprocess.run(cmd, capture_output=True, text=True, timeout=3200)
    out = p.stdout + "\n" + p.stderr
    # capture score lines (RMSE / CRPS)
    scores = [l for l in out.splitlines() if re.search(r"RMSE|CRPS|crps|rmse|score", l, re.I)]
    return {"returncode": p.returncode, "score_lines": scores[-40:], "tail": out.splitlines()[-40:]}


@app.local_entrypoint()
def main():
    import json
    r = run_infer.remote()
    print(json.dumps(r, indent=2))
    with open("ucast_inference_results.json", "w") as f:
        json.dump(r, f, indent=2)