Download code/hb_psi0/hb_train.py from humanoid-benchmark/psi0: direct link, hf CLI and curl.
- Browser
- Download file 2.48 kB
-
https://huggingface.co/humanoid-benchmark/psi0/resolve/main/code/hb_psi0/hb_train.py
- Command line
-
hf download hf://humanoid-benchmark/psi0/code/hb_psi0/hb_train.py
-
curl -L -o hb_train.py https://huggingface.co/humanoid-benchmark/psi0/resolve/main/code/hb_psi0/hb_train.py
2.48 kB
| """torchrun entry: register hb_psi0_config, apply hb_patches, run the pinned upstream scripts/train.py. | |
| Usage (via torchrun): hb_train.py <native Psi0 args ...> (config module name is inserted automatically) | |
| Env: PSI_SRC = Psi0 checkout root (contains scripts/train.py and .env). | |
| HB_PROBE_OUT = dir -> probe mode: no checkpoints, writes rank<R>.json with peak GPU memory + step time. | |
| """ | |
| import json | |
| import os | |
| import runpy | |
| import sys | |
| import time | |
| import traceback | |
| HERE = os.path.dirname(os.path.abspath(__file__)) | |
| sys.path.insert(0, HERE) | |
| def main() -> None: | |
| psi = os.environ["PSI_SRC"] | |
| import hb_psi0_config | |
| sys.modules["psi.config.train.hb_psi0_config"] = hb_psi0_config | |
| import hb_patches | |
| hb_patches.apply() | |
| try: # 2026-11-15 user: in-training open-loop action MSE every 250 steps (HB_INLOOP_EVERY=0 = off) | |
| import hb_inloop_psi0 | |
| hb_inloop_psi0.install() | |
| except Exception as e: # noqa: BLE001 - never block training | |
| print(f"HB_INLOOP disabled: {e!r}", flush=True) | |
| entry = os.path.join(psi, "scripts", "train.py") | |
| sys.argv = [entry, "hb_psi0_config"] + sys.argv[1:] | |
| probe_out = os.environ.get("HB_PROBE_OUT") | |
| t0 = time.time() | |
| ok, err = True, None | |
| try: | |
| runpy.run_path(entry, run_name="__main__") | |
| except BaseException as e: # noqa: BLE001 - probe must record OOM etc. | |
| ok, err = False, f"{type(e).__name__}: {e}"[:4000] | |
| if not probe_out: | |
| raise | |
| traceback.print_exc() | |
| if probe_out: | |
| import torch | |
| rank = int(os.environ.get("RANK", "0")) | |
| res = {"ok": ok, "error": err, "rank": rank, "wall_s": time.time() - t0} | |
| if torch.cuda.is_available(): | |
| dev = torch.cuda.current_device() | |
| res.update(peak_reserved=torch.cuda.max_memory_reserved(dev), | |
| peak_allocated=torch.cuda.max_memory_allocated(dev), | |
| total=torch.cuda.get_device_properties(dev).total_memory) | |
| res["peak_frac"] = res["peak_reserved"] / res["total"] | |
| t = hb_patches.PROBE["times"] | |
| res["steps_done"] = len(t) | |
| if len(t) >= 4: | |
| res["sec_per_step"] = (t[-1] - t[1]) / (len(t) - 2) # skip the first (warm-up) step | |
| os.makedirs(probe_out, exist_ok=True) | |
| with open(os.path.join(probe_out, f"rank{rank}.json"), "w") as f: | |
| json.dump(res, f, indent=2) | |
| sys.exit(0 if ok else 1) | |
| if __name__ == "__main__": | |
| main() | |