"""Decode every mp4 of an arm's LeRobot dataset once -> /workspace/isr/cache/<arm>/<video_key>/chunk-XXX_file-XXX.npy (uint8 [T,H,W,3]).
Page cache (503 GB RAM) keeps them hot; training then reads frames by index with zero video decode.
    python cache_videos.py --arm raw --procs 40"""
import argparse, glob, os, multiprocessing as mp
import numpy as np
def one(args):
    src, dst = args
    if os.path.exists(dst): return dst, 0
    from torchcodec.decoders import VideoDecoder
    d = VideoDecoder(src); n = d.metadata.num_frames
    arr = d.get_frames_in_range(0, n).data.permute(0, 2, 3, 1).contiguous().numpy()
    os.makedirs(os.path.dirname(dst), exist_ok=True); np.save(dst + ".tmp.npy", arr); os.rename(dst + ".tmp.npy", dst)
    return dst, len(arr)
if __name__ == "__main__":
    ap = argparse.ArgumentParser(); ap.add_argument("--arm", required=True); ap.add_argument("--procs", type=int, default=40); a = ap.parse_args()
    root = f"/workspace/isr/datasets/{a.arm}"; jobs = []
    for src in sorted(glob.glob(f"{root}/videos/*/chunk-*/file-*.mp4")):
        rel = os.path.relpath(src, f"{root}/videos"); key, chunk, f = rel.split("/")
        jobs.append((src, f"/workspace/isr/cache/{a.arm}/{key}/{chunk}_{f[:-4]}.npy"))
    with mp.Pool(a.procs) as p: tot = sum(n for _, n in p.imap_unordered(one, jobs, chunksize=4))
    print(f"CACHE_OK {a.arm} files={len(jobs)} frames={tot}", flush=True)