fix(kaggle): pin torch via constraints file in REPAIR cell

Step [3/6] (`pip install --force-reinstall bitsandbytes`) and step [4/6]
(`unsloth unsloth_zoo trl`) were re-resolving torch from default PyPI,
which is currently torch 2.11.0. That clobbered the cu128 torch 2.10.0
installed in step [2/6], producing:

torchvision 0.25.0+cu128 requires torch==2.10.0,
but you have torch 2.11.0 which is incompatible.

Fix:
* Write `/tmp/ermap_constraints.txt` immediately after step 2 pinning
`torch==2.10.0 / torchvision==0.25.0`.
* Step 3 (bnb) now uses `--no-deps` (bnb dlopens torch at runtime, so
install-time deps are unnecessary).
* Steps 4 + 5 pass `-c /tmp/ermap_constraints.txt` so pip can never
upgrade torch out from under us, regardless of what unsloth or trl
declare in their requirement lists.

Made-with: Cursor

kaggle/build_notebook.py

@@ -172,21 +172,33 @@ get_ipython().system('pip install -q --no-cache-dir --force-reinstall '
                      'torch==2.10.0 torchvision==0.25.0 '
                      '--index-url https://download.pytorch.org/whl/cu128')
 
+# Write a pip constraints file so subsequent installs (bnb, unsloth, trl, etc.)
+# 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.
+with open("/tmp/ermap_constraints.txt", "w") as _cf:
+    _cf.write("torch==2.10.0\\ntorchvision==0.25.0\\n")
+
 # 3. Reinstall bitsandbytes against the now-pinned torch.
-print("[3/6] Reinstalling bitsandbytes...")
-get_ipython().system('pip install -q --no-cache-dir --force-reinstall bitsandbytes')
+#    --no-deps because bnb just needs torch at RUNTIME (it dlopens torch's
+#    C++ lib) — its install-time deps don't include torch.
+print("[3/6] Reinstalling bitsandbytes (--no-deps to preserve torch)...")
+get_ipython().system('pip install -q --no-cache-dir --force-reinstall --no-deps bitsandbytes')
 
 # 4. Upgrade unsloth + unsloth_zoo + trl in lockstep. unsloth and
 #    unsloth_zoo are released as a matched pair; if pip pulls a fresh
 #    unsloth_zoo against an old unsloth you get
 #       ImportError: cannot import name 'create_gradient_checkpointing_buffer'
-print("[4/6] Upgrading unsloth + unsloth_zoo + trl...")
+#    The constraint file blocks them from moving torch.
+print("[4/6] Upgrading unsloth + unsloth_zoo + trl (constrained)...")
 get_ipython().system('pip install -q --upgrade --no-cache-dir '
+                     '-c /tmp/ermap_constraints.txt '
                      'unsloth unsloth_zoo "trl>=0.18.2"')
 
 # 5. ER-MAP runtime deps that aren't pre-installed on Kaggle.
-print("[5/6] Installing ER-MAP runtime deps...")
+print("[5/6] Installing ER-MAP runtime deps (constrained)...")
 get_ipython().system('pip install -q --no-cache-dir '
+                     '-c /tmp/ermap_constraints.txt '
                      '"groq>=0.18.0" "huggingface_hub>=0.25.0" '
                      '"gymnasium>=0.29.0" "openenv-core>=0.1.0"')
 

kaggle/train_ermap_grpo_kaggle.ipynb

@@ -120,21 +120,33 @@
     "                     'torch==2.10.0 torchvision==0.25.0 '\n",
     "                     '--index-url https://download.pytorch.org/whl/cu128')\n",
     "\n",
+    "# Write a pip constraints file so subsequent installs (bnb, unsloth, trl, etc.)\n",
+    "# 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",
+    "with open(\"/tmp/ermap_constraints.txt\", \"w\") as _cf:\n",
+    "    _cf.write(\"torch==2.10.0\\ntorchvision==0.25.0\\n\")\n",
+    "\n",
     "# 3. Reinstall bitsandbytes against the now-pinned torch.\n",
-    "print(\"[3/6] Reinstalling bitsandbytes...\")\n",
-    "get_ipython().system('pip install -q --no-cache-dir --force-reinstall bitsandbytes')\n",
+    "#    --no-deps because bnb just needs torch at RUNTIME (it dlopens torch's\n",
+    "#    C++ lib) — its install-time deps don't include torch.\n",
+    "print(\"[3/6] Reinstalling bitsandbytes (--no-deps to preserve torch)...\")\n",
+    "get_ipython().system('pip install -q --no-cache-dir --force-reinstall --no-deps bitsandbytes')\n",
     "\n",
     "# 4. Upgrade unsloth + unsloth_zoo + trl in lockstep. unsloth and\n",
     "#    unsloth_zoo are released as a matched pair; if pip pulls a fresh\n",
     "#    unsloth_zoo against an old unsloth you get\n",
     "#       ImportError: cannot import name 'create_gradient_checkpointing_buffer'\n",
-    "print(\"[4/6] Upgrading unsloth + unsloth_zoo + trl...\")\n",
+    "#    The constraint file blocks them from moving torch.\n",
+    "print(\"[4/6] Upgrading unsloth + unsloth_zoo + trl (constrained)...\")\n",
     "get_ipython().system('pip install -q --upgrade --no-cache-dir '\n",
+    "                     '-c /tmp/ermap_constraints.txt '\n",
     "                     'unsloth unsloth_zoo \"trl>=0.18.2\"')\n",
     "\n",
     "# 5. ER-MAP runtime deps that aren't pre-installed on Kaggle.\n",
-    "print(\"[5/6] Installing ER-MAP runtime deps...\")\n",
+    "print(\"[5/6] Installing ER-MAP runtime deps (constrained)...\")\n",
     "get_ipython().system('pip install -q --no-cache-dir '\n",
+    "                     '-c /tmp/ermap_constraints.txt '\n",
     "                     '\"groq>=0.18.0\" \"huggingface_hub>=0.25.0\" '\n",
     "                     '\"gymnasium>=0.29.0\" \"openenv-core>=0.1.0\"')\n",
     "\n",