Add kaggle/train_grpo.py

kaggle/train_grpo.py

@@ -47,15 +47,8 @@ def _truncate_preview(obj: Any, n: int = 400) -> str:
 
 
 def _ensure_package() -> None:
+    """Always refresh deepagents_rl from HF so Kaggle never runs a stale copy."""
     _vlog("bootstrap.ensure_package.begin")
-    try:
-        import deepagents_rl  # noqa: F401
-
-        _vlog("bootstrap.ensure_package.already_importable")
-        return
-    except ImportError:
-        _vlog("bootstrap.ensure_package.download_from_hf")
-
     token = os.environ.get("HF_TOKEN")
     if not token:
         raise RuntimeError("HF_TOKEN required to download deepagents_rl sources")
@@ -63,6 +56,7 @@ def _ensure_package() -> None:
     from huggingface_hub import snapshot_download
 
     root = Path("/kaggle/working/deepagents_rl_src")
+    _vlog("bootstrap.ensure_package.download_from_hf", local_dir=str(root))
     snapshot_download(
         repo_id=os.getenv("HF_REPO_ID", "vanishingradient/deepagents-rl-coding"),
         repo_type="model",
@@ -72,11 +66,22 @@ def _ensure_package() -> None:
         force_download=True,
     )
     pkg_parent = root / "src"
-    if str(pkg_parent) not in sys.path:
-        sys.path.insert(0, str(pkg_parent))
+    # Prefer the freshly downloaded tree over any earlier import.
+    while str(pkg_parent) in sys.path:
+        sys.path.remove(str(pkg_parent))
+    sys.path.insert(0, str(pkg_parent))
+    for mod in list(sys.modules):
+        if mod == "deepagents_rl" or mod.startswith("deepagents_rl."):
+            del sys.modules[mod]
     import deepagents_rl  # noqa: F401
+    from deepagents_rl.train import swe_env as _swe_env
 
-    _vlog("bootstrap.ensure_package.done", path=str(pkg_parent))
+    _vlog(
+        "bootstrap.ensure_package.done",
+        path=str(pkg_parent),
+        swe_env=str(Path(_swe_env.__file__).resolve()),
+        has_phantom_alias=hasattr(_swe_env, "_PHANTOM_DIRS"),
+    )
 
 
 # ---------------------------------------------------------------------------