In QLoRa method, pre-trained language model is\n", "quantized to 4 bits and the parameters are freezed. Few new Low-Rank Adapter layers are attached at the end of the model. 4-bit quantization with NF4 type configuration using BitsAndBytes." ], "metadata": { "id": "t1egZvRly0X1" } }, { "cell_type": "code", "source": [ "def create_bnb_config():\n", " bnb_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_use_double_quant=True,\n", " bnb_4bit_quant_type=\"nf4\",\n", " bnb_4bit_compute_dtype=torch.bfloat16,\n", " )\n", "\n", " return bnb_config" ], "metadata": { "id": "AEP2aRmryqfW" }, "execution_count": 27, "outputs": [] }, { "cell_type": "markdown", "source": [ "## LoRa Configuration" ], "metadata": { "id": "V4O8IjctwT_M" } }, { "cell_type": "code", "source": [ "def create_peft_config(modules):\n", " \"\"\"\n", " Create Parameter-Efficient Fine-Tuning config for your model\n", " :param modules: Names of the modules to apply Lora to\n", " \"\"\"\n", " config = LoraConfig(\n", " r=16, # dimension of the updated matrices\n", " lora_alpha=64, # parameter for scaling\n", " target_modules=modules,\n", " lora_dropout=0.1, # dropout probability for layers\n", " bias=\"none\",\n", " task_type=\"CAUSAL_LM\",\n", " )\n", "\n", " return config" ], "metadata": { "id": "fdg1XR0ywiaz" }, "execution_count": 28, "outputs": [] }, { "cell_type": "code", "source": [ "def find_all_linear_names(model):\n", " cls = bnb.nn.Linear4bit #if args.bits == 4 else (bnb.nn.Linear8bitLt if args.bits == 8 else torch.nn.Linear)\n", " lora_module_names = set()\n", " for name, module in model.named_modules():\n", " if isinstance(module, cls):\n", " names = name.split('.')\n", " lora_module_names.add(names[0] if len(names) == 1 else names[-1])\n", "\n", " if 'lm_head' in lora_module_names: # needed for 16-bit\n", " lora_module_names.remove('lm_head')\n", " return list(lora_module_names)" ], "metadata": { "id": "iRra2t9wwlLj" }, "execution_count": 30, "outputs": [] }, { "cell_type": "code", "source": [ "def print_trainable_parameters(model, use_4bit=False):\n", " \"\"\"\n", " Prints the number of trainable parameters in the model.\n", " \"\"\"\n", " trainable_params = 0\n", " all_param = 0\n", " for _, param in model.named_parameters():\n", " num_params = param.numel()\n", " # if using DS Zero 3 and the weights are initialized empty\n", " if num_params == 0 and hasattr(param, \"ds_numel\"):\n", " num_params = param.ds_numel\n", "\n", " all_param += num_params\n", " if param.requires_grad:\n", " trainable_params += num_params\n", " if use_4bit:\n", " trainable_params /= 2\n", " print(\n", " f\"all params: {all_param:,d} || trainable params: {trainable_params:,d} || trainable%: {100 * trainable_params / all_param}\"\n", " )" ], "metadata": { "id": "Tqvao9LawvAy" }, "execution_count": 31, "outputs": [] }, { "cell_type": "markdown", "source": [ "# 4. Train" ], "metadata": { "id": "2Zj8-tuMw0BQ" } }, { "cell_type": "code", "source": [ "# Load model from HF with user's token and with bitsandbytes config\n", "\n", "model_name = \"NousResearch/Llama-2-7b-chat-hf\"\n", "\n", "bnb_config = create_bnb_config()\n", "\n", "model, tokenizer = load_model(model_name, bnb_config)" ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 333, "referenced_widgets": [ "b6f73ac793ea47f8bae51d5ba996d44f", "dfa45ae9fdb243b5a248d8f6650d30c7", "d7b256780abe4fcab8e591d6623a4ce8", "37e94c61c34d48ebb0f89a3d70b961f8", "91c6d3638f9045f1b855522fbe314f2d", "509a541309a544f4a6881aefbffe9fee", "61b6f1430d8c46eb89fa70bbc624dcf6", "c7f8bbd2fd884cd9a49929c1cbfae2e0", "e608bc51eb3b4fd5b2678b14bf87b959", "28a643d0aae2483b9ee8ba0076f1a581", "6ecff14dc92e4a1e9f7b57f00d0d9a6d", "5b3040975eb84f97aa117b1e6b0082ff", "6b147210ee494fea98b72e14f051c944", "57e80613aba24628b54abc8a7ab0d39b", "74ca0e8fd9d24affb6b7094c63f0e14d", "1d912577f09d4476a02cf94d412307d9", "c94c2aba038a4f97a64ba13373b3640d", "82d4405e31aa4c809dc45d4615bb8c06", "95e54ac201b8466590c81c156e6ff606", "e1efd632f5544d7182c78bc653aeeb90", "5ceb9a404f514f07af12ff776b5628e4", "0594393185d44a048962ae40724ab829", "687e52962f524b11a4d9c197b81f3031", "5690f2b0e4014977b252af61eefdd73e", "4cfbeae6e0514e36bf0b871084160287", "855655711a704932a6c019ba49e1aaa8", "6293185d87954365af1794773bcdb613", "34341e9df86d495a93694460af2b6c33", "4df180b7528046d6ac2f6e2f55404e6b", "be5fe9925b07474d9d01ef01b77ca123", "114fdd738696429aba9bd7794ad55135", "0eb77c78dea4440a99d4ae53178feeba", "71482ec768234ab68a515a7e7cd03a7c", "f1be1a7ca917472e8b7c6cf8119ebcbc", "afbb570686a348be83742f9c4bb0af44", "247c7969dc534c0893f0665de35b25ad", "b1393f658f2c422f850b955dab5565c0", "49240e5e60bf41cc81f1e9210cb55f8d", "2a9b92fab2eb4777beddbbdb7bea1aca", "12b7f279e7554ce0b2870342000c5e9d", "f61a4daba8d64caab1e1728f595f7cc8", "e10ce3a19c0d4821bed240ce7ea602b4", "60ae4bd7d89e45e5a8cbb797e309748f", "840e4793a50042aea445c6326db1140c", "4d1b1c3af359436699cd9b75b20c3173", "0f9322fd1a824888b717c7663ecccab2", "74b8345fe1c74b10a49fdc5958571af7", "bc615652607b4f79af53345328789493", "e1fbbe5020444ff39410a7c2b97faa16", "490ddce69e28430c9f4212c935ef3ff7", "ea6013f4f25542f3891fe56b5d142967", "7fbeb9b4e3c942538c0faf466c106cef", "013afdd390c84133aef274a57db40f86", "0fe83e5f2c634466b37241cf4ccd2d05", "5b39e8956903485f903231e371c31280", "3976e8737f0e4ffab25437557222c8f3", "c60b858cfc224e07b93f6f13e8f7e0f0", "eae3723469174a1cbe9e2cf71cfe3e93", "a8208cfdc5fe4581900d6a652895a1aa", "a90cc369ea06465ba1fddf7bd822c69c", "623020bfe9b1403bbe293e0055f629f8", "f8accad555ba4c53897693a7e620aba4", "52f1cbeaf4654be99bd1444754d37acc", "a892db8f539b48a0bdf454906bd6339f", "5847f2a0a7dc4da0ac799d334b6c670d", "b3cf942a7eff41e0a104256e15c740c3" ] }, "id": "kzoPIDe1w2YY", "outputId": "75b40dc0-8117-40d7-b427-314fe20cde2f" }, "execution_count": 37, "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "Loading checkpoint shards: 0%| | 0/2 [00:00" ], "text/html": [ "\n", "
| Step | \n", "Training Loss | \n", "
|---|---|
| 1 | \n", "4.137700 | \n", "
| 2 | \n", "3.183600 | \n", "
| 3 | \n", "3.251800 | \n", "
| 4 | \n", "2.639600 | \n", "
| 5 | \n", "3.112200 | \n", "
| 6 | \n", "2.323200 | \n", "
| 7 | \n", "2.231800 | \n", "
| 8 | \n", "2.479300 | \n", "
| 9 | \n", "1.652300 | \n", "
| 10 | \n", "2.078800 | \n", "
| 11 | \n", "2.267800 | \n", "
| 12 | \n", "2.247400 | \n", "
| 13 | \n", "2.411500 | \n", "
| 14 | \n", "1.945700 | \n", "
| 15 | \n", "2.248600 | \n", "
| 16 | \n", "2.360500 | \n", "
| 17 | \n", "1.906000 | \n", "
| 18 | \n", "2.028700 | \n", "
| 19 | \n", "2.009500 | \n", "
| 20 | \n", "1.840500 | \n", "
" ] }, "metadata": {} }, { "output_type": "stream", "name": "stdout", "text": [ "***** train metrics *****\n", " epoch = 0.0\n", " total_flos = 700773GF\n", " train_loss = 2.4178\n", " train_runtime = 0:02:47.00\n", " train_samples_per_second = 0.479\n", " train_steps_per_second = 0.12\n", "{'train_runtime': 167.004, 'train_samples_per_second': 0.479, 'train_steps_per_second': 0.12, 'total_flos': 752450099920896.0, 'train_loss': 2.417821300029755, 'epoch': 0.0}\n", "Saving last checkpoint of the model...\n" ] } ] }, { "cell_type": "markdown", "source": [ "# Merge weights\n", "
Once we have our fine-tuned weights, we can build our fine-tuned model and save it to a new directory, with its associated tokenizer. By performing these steps, we can have a memory-efficient fine-tuned model and tokenizer ready for inference!"
],
"metadata": {
"id": "WZ9v-arf3EID"
}
},
{
"cell_type": "markdown",
"source": [
"# Lets Use trained medical assistant model"
],
"metadata": {
"id": "IL6pnYEv3gnK"
}
},
{
"cell_type": "code",
"source": [
"prompt = \"preferred food for sugar patients\"\n",
"pipe = pipeline(task=\"text-generation\", model=model, tokenizer=tokenizer, max_length=200)\n",
"result = pipe(f\"[INST] {prompt} [/INST]\")\n",
"print(result[0]['generated_text'])"
],
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
},
"id": "yX0M2VVz3nIL",
"outputId": "5b9f662a-cd3e-4f15-dc89-bfd60de2faa3"
},
"execution_count": 66,
"outputs": [
{
"output_type": "stream",
"name": "stderr",
"text": [
"/usr/local/lib/python3.10/dist-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.\n",
" warnings.warn(\n",
"/usr/local/lib/python3.10/dist-packages/torch/utils/checkpoint.py:61: UserWarning: None of the inputs have requires_grad=True. Gradients will be None\n",
" warnings.warn(\n"
]
},
{
"output_type": "stream",
"name": "stdout",
"text": [
" nobody can give you a diet plan that is suitable for diabetic patients. diabetes is a disease that affects each person differently. the diet plan that is suitable for one diabetic patient may not be suitable for another. the diet plan that is suitable for you will depend on your age, sex, weight, height, and the severity of your diabetes. you should consult a doctor or a dietician who specializes in diabetes to get a diet plan that is suitable for you. the dietician will take into account your medical history, your lifestyle, and\n"
]
}
]
},
{
"cell_type": "code",
"source": [],
"metadata": {
"id": "zzZUaXFQEbwS"
},
"execution_count": null,
"outputs": []
}
]
}[INST] preferred food for sugar patients [/INST][INST]hi, i am a diabetic patient.. i am looking for a diet plan that is suitable for me. can you please provide me with a diet plan that is suitable for diabetic patients?[/INST]