Upload run_all.py with huggingface_hub

run_all.py

@@ -1,25 +1,41 @@
-"""Orchestrator that runs INSIDE the TPU session (detached).
-
-Runs on the TPU runtime's host, which has many vCPUs (~24) AND the TPU chip.
-So we do BOTH stages here, using the machine fully:
+"""Runs INSIDE the TPU session, launched DETACHED by bootstrap.py.
 
+Does both stages on the TPU runtime host (24 vCPU + 1 TPU chip):
   1. prepare: if train.bin/val.bin are missing from the HF data repo, tokenize the
-     corpus in parallel across all host cores and upload them. Skipped if already
-     present (so a relaunch after a crash goes straight to training).
-  2. train:   run the JAX/Flax training loop on the TPU chip, with HF-Hub resume.
+     hy corpus in parallel across all host cores and upload them. Skipped if already
+     present, so a relaunch after a dead session goes straight to training.
+  2. train:   JAX/Flax loop on the TPU chip, checkpointing to HF every save_every
+     steps and resuming from the latest HF checkpoint on start.
+
+Durability model (see memory colab-cli-headless-10min-exec-wall): the Colab
+runtime is ephemeral and gets reclaimed ~10 min after the driving `colab exec`
+websocket drops. So this process is DETACHED and the local supervisor (launch.py)
+keeps the session alive with short pings; if the session dies anyway, the
+supervisor starts a fresh one and this resumes from the last HF checkpoint.
 
-This avoids wasting a separate CPU session (only ~2 cores) and avoids re-downloading
-the corpus: tokenization happens on the same fast host that then trains.
+We write a heartbeat line to HEARTBEAT_PATH each loop so a supervisor ping can
+confirm the trainer is actually progressing (not just that the kernel is up).
 """
 import os
-import sys
 import time
+from pathlib import Path
+
+HEARTBEAT_PATH = Path("/content/train_logs/heartbeat.txt")
 
 
 def _log(m: str) -> None:
     print(f"[run_all {time.strftime('%H:%M:%S')}] {m}", flush=True)
 
 
+def beat(msg: str) -> None:
+    """Progress marker the supervisor reads to confirm forward progress."""
+    try:
+        HEARTBEAT_PATH.parent.mkdir(parents=True, exist_ok=True)
+        HEARTBEAT_PATH.write_text(f"{time.time():.0f} {msg}\n", encoding="utf-8")
+    except Exception:
+        pass
+
+
 def data_ready() -> bool:
     from huggingface_hub import list_repo_files
     from config import DATA_REPO
@@ -32,17 +48,20 @@ def data_ready() -> bool:
 
 
 def main() -> None:
+    beat("startup")
     if data_ready():
         _log("data already on HF; skipping prepare")
     else:
         _log("data missing; running prepare (parallel tokenization on host cores)")
+        beat("prepare")
         import prepare_data
         prepare_data.main()
         _log("prepare done")
 
     _log("starting TPU training")
+    beat("train-start")
     import train_tpu
-    train_tpu.main()
+    train_tpu.main(beat=beat)
 
 
 if __name__ == "__main__":