Fix notebook for T4/free tier: auto GPU detection + bitsandbytes fix

- Auto-selects model: 7B (T4/15GB), 14B (A100/40GB), 32B (A100/80GB)
- Upgrades bitsandbytes to fix triton.ops error on Colab free tier
- Training hyperparams tuned per GPU tier

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

notebooks/ch_trader_finetune.ipynb

@@ -22,7 +22,7 @@
    "cell_type": "markdown",
    "id": "title",
    "metadata": {},
-   "source": "# StockEx Clearing House — LLM Fine-Tuning\n\nFine-tunes **Qwen/Qwen2.5-32B-Instruct** with QLoRA to act as a clearing house trading agent.\n\nGiven a member's capital, holdings, and live market BBO, the model outputs a valid JSON trading decision.\n\n**Output model:** `RayMelius/stockex-ch-trader` on HuggingFace Hub\n\n---\n**Runtime:** GPU → A100 80GB (Colab Pro+)  \n**Required secret:** Add `HF_TOKEN` in Colab → Secrets (🔑 icon in left sidebar)"
+   "source": "# StockEx Clearing House — LLM Fine-Tuning\n\nFine-tunes a Qwen2.5 Instruct model with QLoRA to act as a clearing house trading agent.\n\n**Auto-selects model size based on available VRAM:**\n| GPU | VRAM | Model |\n|-----|------|-------|\n| T4 (free) | 15 GB | Qwen2.5-7B-Instruct |\n| A100 40 GB | 40 GB | Qwen2.5-14B-Instruct |\n| A100 80 GB | 80 GB | Qwen2.5-32B-Instruct |\n\n**Output model:** `RayMelius/stockex-ch-trader` on HuggingFace Hub\n\n**Required secret:** Add `HF_TOKEN` in Colab → Secrets (🔑 icon in left sidebar)"
   },
   {
    "cell_type": "code",
@@ -30,17 +30,7 @@
    "id": "install",
    "metadata": {},
    "outputs": [],
-   "source": [
-    "# ── Install dependencies ───────────────────────────────────────────────────────\n",
-    "!pip install -q \\\n",
-    "    transformers==4.46.3 \\\n",
-    "    peft==0.13.2 \\\n",
-    "    trl==0.12.1 \\\n",
-    "    datasets==3.1.0 \\\n",
-    "    accelerate==1.1.1 \\\n",
-    "    bitsandbytes==0.44.1 \\\n",
-    "    huggingface_hub"
-   ]
+   "source": "# ── Install dependencies ───────────────────────────────────────────────────────\n# Reinstall bitsandbytes with proper CUDA support (fixes triton.ops error on Colab)\n!pip install -q -U bitsandbytes\n!pip install -q \\\n    \"transformers>=4.46.3\" \\\n    \"peft>=0.13.2\" \\\n    \"trl>=0.12.1\" \\\n    \"datasets>=3.1.0\" \\\n    \"accelerate>=1.1.1\" \\\n    huggingface_hub\nprint(\"Dependencies installed.\")"
   },
   {
    "cell_type": "code",
@@ -71,7 +61,7 @@
    "id": "config",
    "metadata": {},
    "outputs": [],
-   "source": "# ── Configuration ─────────────────────────────────────────────────────────────\nBASE_MODEL   = \"Qwen/Qwen2.5-32B-Instruct\"\nOUTPUT_REPO  = \"RayMelius/stockex-ch-trader\"\nOUTPUT_DIR   = \"./stockex-ch-trader\"\n\n# LoRA\nLORA_R       = 16\nLORA_ALPHA   = 32\nLORA_DROPOUT = 0.05\n\n# Training  (32B needs smaller batch; effective batch = 1 × 16 = 16)\nNUM_EPOCHS   = 3\nBATCH_SIZE   = 1\nGRAD_ACCUM   = 16\nLR           = 1e-4\nMAX_SEQ_LEN  = 512\nDATASET_SIZE = 2500\n\n# HuggingFace token — reads from Colab Secrets first, then env var\ntry:\n    from google.colab import userdata\n    HF_TOKEN = userdata.get(\"HF_TOKEN\")\n    print(\"HF_TOKEN loaded from Colab Secrets\")\nexcept Exception:\n    HF_TOKEN = os.getenv(\"HF_TOKEN\", \"\")\n\nif not HF_TOKEN:\n    raise ValueError(\"HF_TOKEN not found. Add it in Colab → Secrets (🔑).\")\n\nfrom huggingface_hub import login\nlogin(token=HF_TOKEN)\nprint(\"Logged in to HuggingFace Hub\")"
+   "source": "# ── Auto-select model based on available VRAM ─────────────────────────────────\nimport torch\n\nassert torch.cuda.is_available(), \"No GPU found — change runtime to GPU.\"\nvram_gb = torch.cuda.get_device_properties(0).total_memory / 1e9\ngpu_name = torch.cuda.get_device_name(0)\nprint(f\"GPU: {gpu_name}  |  VRAM: {vram_gb:.1f} GB\")\n\nif vram_gb >= 70:\n    BASE_MODEL = \"Qwen/Qwen2.5-32B-Instruct\"\n    BATCH_SIZE, GRAD_ACCUM, LR = 1, 16, 1e-4\nelif vram_gb >= 35:\n    BASE_MODEL = \"Qwen/Qwen2.5-14B-Instruct\"\n    BATCH_SIZE, GRAD_ACCUM, LR = 2, 8,  1e-4\nelse:                                          # T4 / 15 GB\n    BASE_MODEL = \"Qwen/Qwen2.5-7B-Instruct\"\n    BATCH_SIZE, GRAD_ACCUM, LR = 4, 4,  2e-4\n\nprint(f\"Selected model: {BASE_MODEL}\")\n\n# ── Fixed config ───────────────────────────────────────────────────────────────\nOUTPUT_REPO  = \"RayMelius/stockex-ch-trader\"\nOUTPUT_DIR   = \"./stockex-ch-trader\"\nLORA_R       = 16\nLORA_ALPHA   = 32\nLORA_DROPOUT = 0.05\nNUM_EPOCHS   = 3\nMAX_SEQ_LEN  = 512\nDATASET_SIZE = 2500\n\n# ── HuggingFace login ─────────────────────────────────────────────────────────\nimport os\ntry:\n    from google.colab import userdata\n    HF_TOKEN = userdata.get(\"HF_TOKEN\")\n    print(\"HF_TOKEN loaded from Colab Secrets\")\nexcept Exception:\n    HF_TOKEN = os.getenv(\"HF_TOKEN\", \"\")\n\nif not HF_TOKEN:\n    raise ValueError(\"HF_TOKEN not found. Add it in Colab → Secrets (🔑).\")\n\nfrom huggingface_hub import login\nlogin(token=HF_TOKEN)\nprint(\"Logged in to HuggingFace Hub\")"
   },
   {
    "cell_type": "markdown",