Fix Cell 2: restart kernel after torchvision reinstall using sentinel file

Reinstalling torchvision in-session fixes disk but not memory -- the
mismatched CUDA 12.8 build stays loaded until kernel restart.

Cell 2 now:
1. Detects if sentinel /kaggle/working/.torchvision_fixed exists
2. If not: installs correct torchvision + all packages, creates
sentinel, restarts kernel via IPython.kernel.do_shutdown(restart=True)
3. If yes: skips restart (already fixed), prints torch/CUDA version

The sentinel persists in /kaggle/working for the session so the
restart only happens once -- re-running the cell is safe.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

notebooks/kaggle_master_trainer.ipynb

@@ -33,7 +33,7 @@
    "id": "cell-install",
    "metadata": {},
    "outputs": [],
-   "source": "# -- Cell 2: Install dependencies ---------------------------------------------\nimport subprocess, sys, torch\n\n# Reinstall torchvision to match the exact torch+CUDA version Kaggle provides.\n# torchvision compiled against a different CUDA major version breaks any import\n# that touches torchvision (including WhisperProcessor via image_utils).\ntorch_ver  = torch.__version__.split('+')[0]   # e.g. \"2.6.0\"\ncuda_tag   = 'cu' + torch.version.cuda.replace('.', '')[:3]  # e.g. \"cu130\"\ntv_wheel   = f'torchvision --index-url https://download.pytorch.org/whl/{cuda_tag}'\nprint(f'Reinstalling torchvision for torch={torch_ver}, cuda={cuda_tag} ...')\nsubprocess.check_call([sys.executable, '-m', 'pip', 'install', '-q',\n                       '--upgrade', 'torchvision',\n                       '--index-url', f'https://download.pytorch.org/whl/{cuda_tag}'])\nprint('torchvision reinstalled.')\n\npkgs = [\n    'transformers==5.5.0',\n    'datasets==4.8.4',\n    'accelerate==1.13.0',\n    'evaluate==0.4.2',\n    'huggingface-hub==1.9.0',\n    'peft==0.18.1',\n    'bitsandbytes==0.49.2',\n    'librosa==0.10.2',\n    'soundfile==0.12.1',\n    'jiwer==3.0.4',\n    'numpy==2.2.4',\n    'scipy==1.15.2',\n]\nsubprocess.check_call([sys.executable, '-m', 'pip', 'install', '-q', *pkgs])\nprint('All packages installed.')\nprint(f'torch  : {torch.__version__}')\nprint(f'CUDA   : {torch.version.cuda}')\nprint(f'GPU    : {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"none\"}')"
+   "source": "# -- Cell 2: Install dependencies + fix torchvision CUDA mismatch -------------\n# Kaggle ships torchvision built against CUDA 12.8 but torch uses CUDA 13.0.\n# Any import touching torchvision (WhisperProcessor, WhisperForConditionalGeneration)\n# raises RuntimeError. The fix requires reinstalling torchvision AND restarting\n# the kernel so the new version is loaded from scratch.\n#\n# This cell uses a sentinel file so it only restarts once:\n#   First run  -> installs correct torchvision -> restarts kernel\n#   Second run -> sentinel exists -> skips restart -> continues normally\n\nimport subprocess, sys, os\nfrom pathlib import Path\n\nSENTINEL = Path('/kaggle/working/.torchvision_fixed')\n\nif not SENTINEL.exists():\n    import torch\n    cuda_tag = 'cu' + torch.version.cuda.replace('.', '')[:3]   # e.g. cu130\n    print(f'Fixing torchvision for CUDA {cuda_tag} (will restart kernel) ...')\n\n    subprocess.check_call([\n        sys.executable, '-m', 'pip', 'install', '-q', '--upgrade',\n        'torchvision',\n        '--index-url', f'https://download.pytorch.org/whl/{cuda_tag}',\n    ])\n\n    pkgs = [\n        'transformers==5.5.0', 'datasets==4.8.4', 'accelerate==1.13.0',\n        'evaluate==0.4.2', 'huggingface-hub==1.9.0', 'peft==0.18.1',\n        'bitsandbytes==0.49.2', 'librosa==0.10.2', 'soundfile==0.12.1',\n        'jiwer==3.0.4', 'numpy==2.2.4', 'scipy==1.15.2',\n    ]\n    subprocess.check_call([sys.executable, '-m', 'pip', 'install', '-q', *pkgs])\n\n    SENTINEL.touch()\n    print('Packages installed. Restarting kernel to load new torchvision ...')\n    import IPython\n    IPython.Application.instance().kernel.do_shutdown(restart=True)\n\nelse:\n    print('torchvision already fixed (sentinel exists) -- skipping restart.')\n    import torch\n    print(f'torch     : {torch.__version__}')\n    print(f'CUDA      : {torch.version.cuda}')\n    print(f'GPU       : {torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"none\"}')\n    print('Ready.')"
   },
   {
    "cell_type": "code",