fix(kaggle): align pip-managed numpy with kernel's loaded numpy

Symptom: Cell 5 (post-restart verify) blew up with
RuntimeError: numpy was upgraded mid-session
(loaded: 2.4.4, installed: 2.4.3)
because Kaggle ships numpy 2.4.4 at /usr/lib/python3/dist-packages while
pip writes to /usr/local/lib/python3.12/dist-packages — and torch's
cu128 install pulled numpy 2.4.3 as a transitive dep into the latter.
unsloth_zoo's startup does a strict string-equality check between
`numpy.__version__` (loaded via Python's import path priority) and
importlib.metadata.version("numpy") (reads pip records) and bails on any
mismatch, even patch-level.

Fix in REPAIR cell:

* Detect the kernel's actual loaded numpy version up front and append
`numpy=={detected}` to the pip constraints file, so unsloth/trl/bnb
cannot resolve a different one.
* Add a 5b realign step that does
pip install --force-reinstall --no-deps numpy=={detected}
to overwrite whatever pip put in /usr/local/lib with the version the
kernel actually loads. After this, importlib.metadata and numpy.__version__
agree regardless of which dist-packages path takes precedence.

Made-with: Cursor

kaggle/build_notebook.py

@@ -176,8 +176,19 @@ get_ipython().system('pip install -q --no-cache-dir --force-reinstall '
 # can NEVER pull a different torch from default PyPI. Without this, step 3's
 # `--force-reinstall bitsandbytes` and step 4's `unsloth` upgrade re-resolve
 # torch from PyPI (currently 2.11.0), which breaks the cu128 torchvision pair.
+#
+# Also pin numpy to whatever Kaggle's kernel already has loaded — Kaggle's
+# image puts numpy at /usr/lib/python3/dist-packages while pip writes to
+# /usr/local/lib/python3.12/dist-packages, so any version drift between the
+# two paths trips unsloth_zoo's strict loaded-vs-installed check at import.
+import subprocess as _sp
+_kernel_numpy = _sp.check_output(
+    [sys.executable, "-c", "import numpy; print(numpy.__version__)"],
+    text=True,
+).strip()
+print(f"      detected kernel numpy = {_kernel_numpy} (will pin)")
 with open("/tmp/ermap_constraints.txt", "w") as _cf:
-    _cf.write("torch==2.10.0\\ntorchvision==0.25.0\\n")
+    _cf.write(f"torch==2.10.0\\ntorchvision==0.25.0\\nnumpy=={_kernel_numpy}\\n")
 
 # 3. Reinstall bitsandbytes against the now-pinned torch.
 #    --no-deps because bnb just needs torch at RUNTIME (it dlopens torch's
@@ -202,6 +213,16 @@ get_ipython().system('pip install -q --no-cache-dir '
                      '"groq>=0.18.0" "huggingface_hub>=0.25.0" '
                      '"gymnasium>=0.29.0" "openenv-core>=0.1.0"')
 
+# 5b. Realign the pip-managed numpy with whatever the Kaggle kernel actually
+#     has loaded. This force-rewrites /usr/local/lib/.../numpy at the exact
+#     version reported by the running interpreter, so importlib.metadata
+#     and `numpy.__version__` agree even if Kaggle ships its base numpy at
+#     a different dist-packages path.
+print(f"[5b/6] Realigning pip-managed numpy to {_kernel_numpy}...")
+get_ipython().system(
+    f'pip install -q --force-reinstall --no-deps "numpy=={_kernel_numpy}"'
+)
+
 # 6. Verify in a SUBPROCESS (so the parent kernel never imports any of these
 #    while pip is mid-flight, which is what causes the
 #       'numpy was upgraded mid-session (loaded: X, installed: Y)' RuntimeError

kaggle/train_ermap_grpo_kaggle.ipynb

@@ -124,8 +124,19 @@
     "# can NEVER pull a different torch from default PyPI. Without this, step 3's\n",
     "# `--force-reinstall bitsandbytes` and step 4's `unsloth` upgrade re-resolve\n",
     "# torch from PyPI (currently 2.11.0), which breaks the cu128 torchvision pair.\n",
+    "#\n",
+    "# Also pin numpy to whatever Kaggle's kernel already has loaded — Kaggle's\n",
+    "# image puts numpy at /usr/lib/python3/dist-packages while pip writes to\n",
+    "# /usr/local/lib/python3.12/dist-packages, so any version drift between the\n",
+    "# two paths trips unsloth_zoo's strict loaded-vs-installed check at import.\n",
+    "import subprocess as _sp\n",
+    "_kernel_numpy = _sp.check_output(\n",
+    "    [sys.executable, \"-c\", \"import numpy; print(numpy.__version__)\"],\n",
+    "    text=True,\n",
+    ").strip()\n",
+    "print(f\"      detected kernel numpy = {_kernel_numpy} (will pin)\")\n",
     "with open(\"/tmp/ermap_constraints.txt\", \"w\") as _cf:\n",
-    "    _cf.write(\"torch==2.10.0\\ntorchvision==0.25.0\\n\")\n",
+    "    _cf.write(f\"torch==2.10.0\\ntorchvision==0.25.0\\nnumpy=={_kernel_numpy}\\n\")\n",
     "\n",
     "# 3. Reinstall bitsandbytes against the now-pinned torch.\n",
     "#    --no-deps because bnb just needs torch at RUNTIME (it dlopens torch's\n",
@@ -150,6 +161,16 @@
     "                     '\"groq>=0.18.0\" \"huggingface_hub>=0.25.0\" '\n",
     "                     '\"gymnasium>=0.29.0\" \"openenv-core>=0.1.0\"')\n",
     "\n",
+    "# 5b. Realign the pip-managed numpy with whatever the Kaggle kernel actually\n",
+    "#     has loaded. This force-rewrites /usr/local/lib/.../numpy at the exact\n",
+    "#     version reported by the running interpreter, so importlib.metadata\n",
+    "#     and `numpy.__version__` agree even if Kaggle ships its base numpy at\n",
+    "#     a different dist-packages path.\n",
+    "print(f\"[5b/6] Realigning pip-managed numpy to {_kernel_numpy}...\")\n",
+    "get_ipython().system(\n",
+    "    f'pip install -q --force-reinstall --no-deps \"numpy=={_kernel_numpy}\"'\n",
+    ")\n",
+    "\n",
     "# 6. Verify in a SUBPROCESS (so the parent kernel never imports any of these\n",
     "#    while pip is mid-flight, which is what causes the\n",
     "#       'numpy was upgraded mid-session (loaded: X, installed: Y)' RuntimeError\n",