Fix MAX_SEQ_LENGTH: 1024 was too small for prompt+completion, bump to 2048

training/train_grpo.ipynb

@@ -128,27 +128,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "import torch\n",
-    "\n",
-    "MAX_STEPS = 300\n",
-    "MAX_SEQ_LENGTH = 1024\n",
-    "LORA_RANK = 32\n",
-    "\n",
-    "gpu_name = torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"cpu\"\n",
-    "print(f\"GPU: {gpu_name}\")\n",
-    "\n",
-    "if any(x in gpu_name.upper() for x in [\"H100\",\"A100\",\"L40\",\"4090\"]):\n",
-    "    MODEL_NAME = \"unsloth/Qwen3-4B\"\n",
-    "    LOAD_IN_4BIT = False; FAST_INFERENCE = True; NUM_GENERATIONS = 4; LR = 5e-5\n",
-    "    print(\"Config: Qwen3-4B instruct, BF16 + vLLM\")\n",
-    "else:\n",
-    "    MODEL_NAME = \"unsloth/Qwen3-4B-unsloth-bnb-4bit\"\n",
-    "    LOAD_IN_4BIT = True; FAST_INFERENCE = False; NUM_GENERATIONS = 2; LR = 2e-4\n",
-    "    print(\"Config: Qwen3-4B instruct, 4-bit (T4)\")\n",
-    "\n",
-    "print(f\"Model: {MODEL_NAME} | Steps: {MAX_STEPS} | Gens: {NUM_GENERATIONS}\")"
-   ]
+   "source": "import torch\n\nMAX_STEPS = 300\nMAX_SEQ_LENGTH = 2048  # Must be >= max_prompt_length + max_completion_length (512+1024)\nLORA_RANK = 32\n\ngpu_name = torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"cpu\"\nprint(f\"GPU: {gpu_name}\")\n\nif any(x in gpu_name.upper() for x in [\"H100\",\"A100\",\"L40\",\"4090\"]):\n    MODEL_NAME = \"unsloth/Qwen3-4B\"\n    LOAD_IN_4BIT = False; FAST_INFERENCE = True; NUM_GENERATIONS = 4; LR = 5e-5\n    print(\"Config: Qwen3-4B instruct, BF16 + vLLM\")\nelse:\n    MODEL_NAME = \"unsloth/Qwen3-4B-unsloth-bnb-4bit\"\n    LOAD_IN_4BIT = True; FAST_INFERENCE = False; NUM_GENERATIONS = 2; LR = 2e-4\n    print(\"Config: Qwen3-4B instruct, 4-bit (T4)\")\n\nprint(f\"Model: {MODEL_NAME} | Steps: {MAX_STEPS} | Gens: {NUM_GENERATIONS} | SeqLen: {MAX_SEQ_LENGTH}\")"
   },
   {
    "cell_type": "markdown",