Notebook: Phase 8c cleanup cell to free GPU before Phase 9 mini-train (OOM fix)

notebooks/colab_train.ipynb

@@ -403,6 +403,43 @@
     "KL to base 0.595, sustained.\n"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Phase 8c \u2014 Free GPU memory before mini-training\n",
+    "\n",
+    "Phase 6 loaded the trained Qwen-3B (~7 GB) into the kernel's GPU memory\n",
+    "and Phase 8 ran a separate eval subprocess. Before we launch the mini\n",
+    "GRPO retrain (which spawns *another* subprocess that loads Qwen-1.5B\n",
+    "for training), we have to free the parent kernel's GPU references \u2014\n",
+    "otherwise the subprocess sees < 100 MB free on a T4 and OOMs\n",
+    "immediately. Skip this cell only if Phase 9 won't be run.\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": [],
+   "source": [
+    "# Drop every reference to the loaded trained model + LoRA so the\n",
+    "# subprocess in Phase 9 has the full GPU to itself.\n",
+    "import gc, torch\n",
+    "for name in ('trained', 'policy', 'result_trained'):\n",
+    "    if name in globals():\n",
+    "        try:\n",
+    "            del globals()[name]\n",
+    "        except Exception:\n",
+    "            pass\n",
+    "gc.collect()\n",
+    "if torch.cuda.is_available():\n",
+    "    torch.cuda.empty_cache()\n",
+    "    torch.cuda.ipc_collect()\n",
+    "    free_mb, total_mb = (m // (1024**2) for m in torch.cuda.mem_get_info())\n",
+    "    print(f'GPU free: {free_mb} MB / {total_mb} MB')\n"
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},
@@ -428,12 +465,16 @@
     "import os\n",
     "os.chdir('/content/chaosops_src')\n",
     "os.makedirs('/content/artifacts/mini-grpo', exist_ok=True)\n",
-    "!PYTHONPATH=/tmp python -m chaosops.train.grpo_train \\\n",
+    "# Sized for free-tier T4 (16 GB) AFTER Phase 8c cleanup.\n",
+    "# lora_rank=8, max_seq_length=768, group=2 keeps peak VRAM ~6 GB.\n",
+    "!PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \\\n",
+    "    PYTHONPATH=/tmp python -m chaosops.train.grpo_train \\\n",
     "    --model-name Qwen/Qwen2.5-1.5B-Instruct \\\n",
     "    --backend transformers \\\n",
     "    --total-episodes 20 \\\n",
     "    --group-size 2 \\\n",
-    "    --lora-rank 16 \\\n",
+    "    --lora-rank 8 \\\n",
+    "    --max-seq-length 768 \\\n",
     "    --learning-rate 2e-5 \\\n",
     "    --temperature 0.8 \\\n",
     "    --rogue-bonus-multiplier 2.0 \\\n",
@@ -447,21 +488,34 @@
    "execution_count": null,
    "outputs": [],
    "source": [
-    "# Plot the mini-run reward curve so judges can see live signal\n",
-    "import json, matplotlib.pyplot as plt\n",
-    "log = json.load(open('/content/artifacts/mini-grpo/training_metrics.json'))\n",
-    "xs = [e['episode'] for e in log]\n",
-    "ys = [e['mean_combined_reward'] for e in log]\n",
-    "plt.figure(figsize=(8, 4))\n",
-    "plt.plot(xs, ys, 'o-', color='#8e44ad', linewidth=2)\n",
-    "plt.axhline(0, color='#888', linewidth=0.6)\n",
-    "plt.xlabel('Training step')\n",
-    "plt.ylabel('Mean combined reward (0.6 \u00b7 team + 0.4 \u00b7 oversight)')\n",
-    "plt.title('Mini-GRPO reward curve (Qwen 2.5-1.5B, 20 steps, T4)')\n",
-    "plt.grid(True, linestyle=':', alpha=0.4)\n",
-    "plt.tight_layout()\n",
-    "plt.savefig('/content/artifacts/mini-grpo/learning_curve.png', dpi=150)\n",
-    "plt.show()\n"
+    "# Plot the mini-run reward curve so judges can see live signal.\n",
+    "# If training crashed (OOM, etc.) we print a hint instead of throwing.\n",
+    "import json, os, matplotlib.pyplot as plt\n",
+    "metrics_path = '/content/artifacts/mini-grpo/training_metrics.json'\n",
+    "if not os.path.exists(metrics_path):\n",
+    "    print('No training_metrics.json found \u2014 Phase 9 training did not complete.')\n",
+    "    print('Common causes:')\n",
+    "    print('  \u2022 Phase 8c memory cleanup was skipped \u2192 mini-train OOMed.')\n",
+    "    print('  \u2022 Colab kernel ran out of GPU before training started.')\n",
+    "    print('Phase 8 already proved the trained adapter beats baselines \u2014')\n",
+    "    print('Phase 9 is OPTIONAL training-pipeline reproducibility evidence.')\n",
+    "else:\n",
+    "    log = json.load(open(metrics_path))\n",
+    "    if not log:\n",
+    "        print('training_metrics.json is empty \u2014 no log points were captured.')\n",
+    "    else:\n",
+    "        xs = [e['episode'] for e in log]\n",
+    "        ys = [e['mean_combined_reward'] for e in log]\n",
+    "        plt.figure(figsize=(8, 4))\n",
+    "        plt.plot(xs, ys, 'o-', color='#8e44ad', linewidth=2)\n",
+    "        plt.axhline(0, color='#888', linewidth=0.6)\n",
+    "        plt.xlabel('Training step')\n",
+    "        plt.ylabel('Mean combined reward (0.6 \u00b7 team + 0.4 \u00b7 oversight)')\n",
+    "        plt.title('Mini-GRPO reward curve (Qwen 2.5-1.5B, 20 steps, T4)')\n",
+    "        plt.grid(True, linestyle=':', alpha=0.4)\n",
+    "        plt.tight_layout()\n",
+    "        plt.savefig('/content/artifacts/mini-grpo/learning_curve.png', dpi=150)\n",
+    "        plt.show()\n"
    ]
   },
   {