feat(autotune): serve-ready Kaggle notebook — flat-adapter self-homing, tunnel URL relay

Training v10 COMPLETE (500/500, train_loss 0.026, eval_loss 0.066); adapter
verified (safetensors header == file size) and uploaded as dataset
liovinajo/nlsql-adapter. Kaggle unpacks zips flat, so serve resolves the
adapter dir by rglob and refuses to silently serve the base model. Live
kernel logs are unreadable via API, so the serve cell publishes the
cloudflared URL through a private nlsql-tunnel dataset (creds via
nlsql-relay). Serve push itself is blocked until the weekly GPU quota
resets.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: JuliaEdom

scripts/autotune/kaggle/kaggle_autotune.ipynb

@@ -20,7 +20,24 @@
    "id": "acae54e37e7d407bbb7b55eff062a284",
    "metadata": {},
    "outputs": [],
-   "source": "MODE = \"train\"  # \"train\" | \"serve\"\n\nBASE_MODEL = \"Qwen/Qwen2.5-Coder-7B-Instruct\"\nDATA_DIR = \"/kaggle/input/nlsql-autotune-data\"  # train.jsonl, val.jsonl, train_qlora.py\nOUT = \"/kaggle/working/qlora_out\"\n\n# train mode\nEPOCHS = 1.0\nMAX_SEQ = 4096\nSMOKE_STEPS = 0  # >0: run this many steps end-to-end first, then the real run (0 = straight in)\nRESUME_INPUT = \"auto\"  # non-empty: rglob /kaggle/input for checkpoint-* and resume; \"\" = fresh\n\n# serve mode\nADAPTER_DIR = \"/kaggle/input/nlsql-adapter/adapter\"  # absent path => serve plain base model\nSERVE_HOURS = 5.0\nPORT = 8000"
+   "source": [
+    "MODE = \"serve\"  # \"train\" | \"serve\"\n",
+    "\n",
+    "BASE_MODEL = \"Qwen/Qwen2.5-Coder-7B-Instruct\"\n",
+    "DATA_DIR = \"/kaggle/input/nlsql-autotune-data\"  # train.jsonl, val.jsonl, train_qlora.py\n",
+    "OUT = \"/kaggle/working/qlora_out\"\n",
+    "\n",
+    "# train mode\n",
+    "EPOCHS = 1.0\n",
+    "MAX_SEQ = 4096\n",
+    "SMOKE_STEPS = 0  # >0: run this many steps end-to-end first, then the real run (0 = straight in)\n",
+    "RESUME_INPUT = \"auto\"  # non-empty: rglob /kaggle/input for checkpoint-* and resume; \"\" = fresh\n",
+    "\n",
+    "# serve mode\n",
+    "ADAPTER_DIR = \"/kaggle/input/nlsql-adapter/adapter\"  # absent path => serve plain base model\n",
+    "SERVE_HOURS = 8.0\n",
+    "PORT = 8000"
+   ]
   },
   {
    "cell_type": "code",
@@ -37,6 +54,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "import json\n",
+    "import os\n",
     "import re\n",
     "import subprocess\n",
     "import sys\n",
@@ -56,7 +75,15 @@
     "    import torch\n",
     "\n",
     "    tp = 2 if torch.cuda.device_count() >= 2 else 1\n",
-    "    with_lora = Path(ADAPTER_DIR).is_dir()\n",
+    "    # Kaggle unpacks an uploaded zip FLAT (v9/v10 lesson) -- self-home on the config\n",
+    "    hits = sorted(Path(\"/kaggle/input/nlsql-adapter\").rglob(\"adapter_config.json\"))\n",
+    "    adapter_dir = str(hits[0].parent) if hits else ADAPTER_DIR\n",
+    "    with_lora = Path(adapter_dir, \"adapter_model.safetensors\").exists()\n",
+    "    if not with_lora:\n",
+    "        raise RuntimeError(\n",
+    "            \"no adapter under /kaggle/input/nlsql-adapter -- refusing to silently serve base\"\n",
+    "        )\n",
+    "    print(f\"adapter_dir resolved: {adapter_dir}\", flush=True)\n",
     "    vllm_cmd = [\n",
     "        \"vllm\",\n",
     "        \"serve\",\n",
@@ -78,7 +105,7 @@
     "            \"--max-lora-rank\",\n",
     "            \"16\",\n",
     "            \"--lora-modules\",\n",
-    "            f\"sqltuned={ADAPTER_DIR}\",\n",
+    "            f\"sqltuned={adapter_dir}\",\n",
     "        ]\n",
     "    print(\" \".join(vllm_cmd), flush=True)\n",
     "    with open(\"/kaggle/working/vllm.log\", \"w\") as vllm_log:\n",
@@ -109,6 +136,35 @@
     "            url = m.group(0)\n",
     "            break\n",
     "    Path(\"/kaggle/working/tunnel_url.txt\").write_text(url or \"NONE\")\n",
+    "    # Live kernel logs are unreadable via API, so push the URL out through a\n",
+    "    # private dataset the Windows side can poll. Creds ride in nlsql-relay.\n",
+    "    creds = json.loads(Path(\"/kaggle/input/nlsql-relay/kaggle.json\").read_text())\n",
+    "    relay_env = {**os.environ, \"KAGGLE_USERNAME\": creds[\"username\"], \"KAGGLE_KEY\": creds[\"key\"]}\n",
+    "    relay = Path(\"/kaggle/working/relay\")\n",
+    "    relay.mkdir(exist_ok=True)\n",
+    "    (relay / \"tunnel_url.txt\").write_text(url or \"NONE\")\n",
+    "    (relay / \"dataset-metadata.json\").write_text(\n",
+    "        json.dumps(\n",
+    "            {\n",
+    "                \"title\": \"nlsql-tunnel\",\n",
+    "                \"id\": \"liovinajo/nlsql-tunnel\",\n",
+    "                \"licenses\": [{\"name\": \"CC0-1.0\"}],\n",
+    "            }\n",
+    "        )\n",
+    "    )\n",
+    "    for verb in ([\"version\", \"-m\", \"tunnel-url\"], [\"create\"]):\n",
+    "        rr = subprocess.run(\n",
+    "            [\"kaggle\", \"datasets\", *verb, \"-p\", str(relay)],\n",
+    "            capture_output=True,\n",
+    "            text=True,\n",
+    "            env=relay_env,\n",
+    "        )\n",
+    "        print(f\"relay {verb[0]}: rc={rr.returncode}\", flush=True)\n",
+    "        if rr.returncode == 0:\n",
+    "            break\n",
+    "    else:\n",
+    "        print(\"relay FAILED; url only in log/output\", flush=True)\n",
+    "        print(rr.stdout[-400:], rr.stderr[-400:], flush=True)\n",
     "    model_name = \"sqltuned\" if with_lora else BASE_MODEL\n",
     "    print(\n",
     "        f\"\\n{'=' * 60}\\nTUNNEL: {url}\\nbase_url for .env: {url}/v1\\n--sql-model {model_name}\\n{'=' * 60}\\n\",\n",

scripts/autotune/kaggle/kernel-metadata.json

@@ -8,7 +8,10 @@
   "enable_gpu": true,
   "machine_shape": "NvidiaTeslaT4",
   "enable_internet": true,
-  "dataset_sources": ["liovinajo/nlsql-autotune-data", "liovinajo/nlsql-autotune-prev"],
+  "dataset_sources": [
+    "liovinajo/nlsql-adapter",
+    "liovinajo/nlsql-relay"
+  ],
   "competition_sources": [],
   "kernel_sources": []
-}
+}