Update Gemma3 multimodal SFT Jupyter notebook

.github/workflows/run_jupyter_notebooks.yml

@@ -123,6 +123,11 @@ jobs:
             echo "------------------------------------------------------"
 
             $PAPERMILL_EXE "$notebook" "$output_name" -k maxtext_venv
+
+            # Clean up any checkpoint directories created by the notebook to avoid filling up disk space
+            echo "Post-notebook disk cleanup for $filename ..."
+            rm -rf "$MAXTEXT_PKG_DIR"/sft_*_output "$MAXTEXT_PKG_DIR"/rl_*_output
+            rm -rf "$HOME/.cache/huggingface/hub"
           done
       - name: Upload Outputs
         if: always()

docs/tutorials/posttraining/multimodal.md

@@ -6,7 +6,7 @@ This document provides a guide to use the multimodal functionalities in MaxText
 - **Multimodal Decode**: Inference with text+images as input.
 - **Supervised Fine-Tuning (SFT)**: Apply SFT to the model using a visual-question-answering dataset.
 
-We also provide a [colab](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/examples/multimodal_gemma3_demo.ipynb) for multimodal features demonstration. The following table provides a list of models and modalities we currently support:
+We also provide a [colab](https://github.com/AI-Hypercomputer/maxtext/blob/main/src/maxtext/examples/sft_multimodal_gemma3_demo.ipynb) for multimodal features demonstration. The following table provides a list of models and modalities we currently support:
 
 | Models                                         | Input Modalities | Output Modalities |
 | :--------------------------------------------- | :--------------- | :---------------- |

src/maxtext/examples/rl_llama3_demo.ipynb

@@ -222,6 +222,12 @@
     "        check=True,\n",
     "        env=env\n",
     "    )\n",
+    "\n",
+    "    # The HF model cache is no longer needed after conversion to MaxText format.\n",
+    "    import shutil\n",
+    "    hf_cache = epath.Path(os.path.expanduser(\"~\")) / \".cache\" / \"huggingface\" / \"hub\"\n",
+    "    if hf_cache.exists():\n",
+    "        shutil.rmtree(str(hf_cache))\n",
     "    \n",
     "    MODEL_CHECKPOINT_PATH = os.path.join(MODEL_CHECKPOINT_PATH, \"0/items\")\n",
     "else:\n",

src/maxtext/examples/sft_llama3_demo_tpu.ipynb

@@ -236,6 +236,12 @@
     "        env=env\n",
     "    )\n",
     "\n",
+    "    # The HF model cache is no longer needed after conversion to MaxText format.\n",
+    "    import shutil\n",
+    "    hf_cache = epath.Path(os.path.expanduser(\"~\")) / \".cache\" / \"huggingface\" / \"hub\"\n",
+    "    if hf_cache.exists():\n",
+    "        shutil.rmtree(str(hf_cache))\n",
+    "\n",
     "    MODEL_CHECKPOINT_PATH = os.path.join(MODEL_CHECKPOINT_PATH, \"0/items\")\n",
     "else:\n",
     "    print(f\"Model checkpoint exists at {MODEL_CHECKPOINT_PATH}\")"