Added a kaggle notebook for training

notebooks/qwen3_coder_codeflow_gguf_to_kaggle.ipynb

@@ -0,0 +1,205 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "# Fine-tune Qwen3-Coder and Upload Q3_K_XL GGUF to Kaggle\n",
+        "\n",
+        "This notebook runs the repo's QLoRA fine-tune, exports `out/qwen3-coder-codeflow-Q3_K_XL.gguf`, and uploads it to Kaggle as a dataset.\n",
+        "\n",
+        "Use a high-memory GPU runtime. A free T4/L4 Colab runtime is unlikely to have enough VRAM/RAM/disk for the 30B merge/export step. A100/H100-class runtime with high system RAM is the safer target.\n",
+        "\n",
+        "Before running the Kaggle upload cell, add these Colab Secrets:\n",
+        "\n",
+        "- `KAGGLE_USERNAME`\n",
+        "- `KAGGLE_KEY`\n",
+        "\n",
+        "Do not paste your Kaggle API key into the notebook itself."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Check GPU and disk before doing the expensive work.\n",
+        "!nvidia-smi\n",
+        "!df -h /content\n",
+        "!free -h"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Clone the repo. If you uploaded this notebook inside an already-cloned repo,\n",
+        "# skip this cell and set REPO_DIR to that path instead.\n",
+        "REPO_URL = \"https://github.com/rishijain27/codeflow-qwen-3-finetuning.git\"\n",
+        "REPO_DIR = \"/content/codeflow-qwen-3-finetuning\"\n",
+        "\n",
+        "import os\n",
+        "\n",
+        "if not os.path.exists(REPO_DIR):\n",
+        "    !git clone {REPO_URL} {REPO_DIR}\n",
+        "else:\n",
+        "    print(f\"Repo already exists at {REPO_DIR}\")\n",
+        "\n",
+        "%cd {REPO_DIR}"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# System build tools for llama.cpp. CUDA is provided by the Colab runtime.\n",
+        "!apt-get update\n",
+        "!apt-get install -y git build-essential ninja-build"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Create and populate the repo-local venv.\n",
+        "!python3 -m venv .venv\n",
+        "!.venv/bin/python -m pip install -U pip setuptools wheel\n",
+        "!.venv/bin/python -m pip install -r requirements.txt\n",
+        "!.venv/bin/python -m pip install kaggle"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Optional dry run: validates tokenization and dataset formatting without loading the 30B model.\n",
+        "!.venv/bin/python finetune.py --dry-run"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Fine-tune, merge, auto-build llama.cpp under .venv/llama.cpp if needed, and export Q3_K_XL GGUF.\n",
+        "# This is the long-running step.\n",
+        "!.venv/bin/python finetune.py \\\n",
+        "  --model Qwen/Qwen3-Coder-30B-A3B-Instruct \\\n",
+        "  --4bit \\\n",
+        "  --epochs 3 \\\n",
+        "  --batch-size 1 \\\n",
+        "  --grad-accum 16 \\\n",
+        "  --max-seq-len 2048 \\\n",
+        "  --output-dir out/qwen-mermaid-lora \\\n",
+        "  --merge-dir out/qwen-mermaid-merged \\\n",
+        "  --merge-device-map cpu \\\n",
+        "  --export-gguf \\\n",
+        "  --gguf-out out/qwen3-coder-codeflow-Q3_K_XL.gguf \\\n",
+        "  --gguf-quant Q3_K_XL"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Verify the GGUF exists.\n",
+        "!ls -lh out/qwen3-coder-codeflow-Q3_K_XL.gguf"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Configure Kaggle credentials from Colab Secrets.\n",
+        "from google.colab import userdata\n",
+        "import json\n",
+        "import os\n",
+        "from pathlib import Path\n",
+        "\n",
+        "kaggle_username = userdata.get(\"KAGGLE_USERNAME\")\n",
+        "kaggle_key = userdata.get(\"KAGGLE_KEY\")\n",
+        "if not kaggle_username or not kaggle_key:\n",
+        "    raise RuntimeError(\"Add KAGGLE_USERNAME and KAGGLE_KEY in Colab Secrets before running this cell.\")\n",
+        "\n",
+        "kaggle_dir = Path.home() / \".kaggle\"\n",
+        "kaggle_dir.mkdir(parents=True, exist_ok=True)\n",
+        "kaggle_json = kaggle_dir / \"kaggle.json\"\n",
+        "kaggle_json.write_text(json.dumps({\"username\": kaggle_username, \"key\": kaggle_key}))\n",
+        "kaggle_json.chmod(0o600)\n",
+        "print(f\"Configured Kaggle credentials for {kaggle_username}\")"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Stage the GGUF and Kaggle dataset metadata.\n",
+        "import json\n",
+        "import shutil\n",
+        "from pathlib import Path\n",
+        "\n",
+        "upload_dir = Path(\"kaggle_upload\")\n",
+        "upload_dir.mkdir(exist_ok=True)\n",
+        "gguf_src = Path(\"out/qwen3-coder-codeflow-Q3_K_XL.gguf\")\n",
+        "gguf_dst = upload_dir / gguf_src.name\n",
+        "if not gguf_src.exists():\n",
+        "    raise FileNotFoundError(gguf_src)\n",
+        "if not gguf_dst.exists():\n",
+        "    shutil.copy2(gguf_src, gguf_dst)\n",
+        "\n",
+        "metadata = {\n",
+        "    \"title\": \"Qwen3 Coder CodeFlow Q3_K_XL GGUF\",\n",
+        "    \"id\": \"rishijain27/qwen3-coder-codeflow-q3-k-xl-gguf\",\n",
+        "    \"licenses\": [{\"name\": \"other\"}],\n",
+        "}\n",
+        "(upload_dir / \"dataset-metadata.json\").write_text(json.dumps(metadata, indent=2))\n",
+        "print(f\"Staged files in {upload_dir.resolve()}\")\n",
+        "!ls -lh kaggle_upload"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Create the Kaggle dataset, or create a new version if it already exists.\n",
+        "!KAGGLE_CONFIG_DIR=$HOME/.kaggle .venv/bin/kaggle datasets create -p kaggle_upload --dir-mode skip || \\\n",
+        " KAGGLE_CONFIG_DIR=$HOME/.kaggle .venv/bin/kaggle datasets version -p kaggle_upload --dir-mode skip -m \"Update Qwen3 Coder CodeFlow Q3_K_XL GGUF\""
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "A100",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}