Upload code/run_training.sh with huggingface_hub

code/run_training.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+# Training entry point inside an HF Jobs container.
+# Mounted by launch_hf_training.sh: the code snapshot at /job (written by the
+# launcher via stdin? no -- this file is uploaded to the artifacts repo too).
+set -euo pipefail
+
+echo "[job] host: $(hostname)"
+nvidia-smi || echo "[job] WARNING: no GPU visible"
+
+# --- deps: jax + CUDA 12, mujoco playground, wandb --------------------------
+pip install --quiet --no-cache-dir \
+    "jax[cuda12]==0.11.1" jaxlib==0.11.1 \
+    mujoco==3.12.0 mujoco-mjx==3.12.0 mujoco-playground==0.2.0 \
+    brax==0.14.2 mediapy tensorboardX wandb
+
+export MUJOCO_GL=egl
+export XLA_PYTHON_CLIENT_PREALLOCATE=false
+export XLA_PYTHON_CLIENT_MEM_FRACTION=0.85
+export XLA_FLAGS="--xla_gpu_triton_gemm_any=True"
+
+# --- unpack the code snapshot (pre-uploaded to the artifacts dataset repo) ---
+mkdir -p /work/repo && cd /work/repo
+tar xzf /artifacts/code/code.tar.gz
+
+echo "[job] training starting: $TRAIN_FLAGS"
+python rl/scripts/train_jax_ppo.py $TRAIN_FLAGS --logdir /work/logs
+
+# --- upload artifacts back to the Hub ---------------------------------------
+python - <<'PY'
+import os
+from huggingface_hub import HfApi
+api = HfApi()
+repo = os.environ["ARTIFACTS_REPO"]
+exp = os.environ["EXP_NAME"]
+api.create_repo(repo, repo_type="dataset", private=True, exist_ok=True)
+api.upload_folder(
+    repo_id=repo, repo_type="dataset",
+    folder_path="/work/logs",
+    path_in_repo=f"runs/{exp}",
+)
+print("[job] artifacts uploaded to", repo, "runs/", exp)
+PY
+echo "[job] done."