omitakahiro
commited on
Commit
•
1803d3a
1
Parent(s):
8f53238
Delete notebooks/QLoRA.ipynb
Browse files- notebooks/QLoRA.ipynb +0 -253
notebooks/QLoRA.ipynb
DELETED
@@ -1,253 +0,0 @@
|
|
1 |
-
{
|
2 |
-
"cells": [
|
3 |
-
{
|
4 |
-
"cell_type": "markdown",
|
5 |
-
"id": "1a884871-7a65-4501-9063-c85ad260d0da",
|
6 |
-
"metadata": {},
|
7 |
-
"source": [
|
8 |
-
"このnotebookはstockmark/stockmark-13bのモデルをkunishou/databricks-dolly-15k-jaのデータセットを用いてQLoRA tuningするためのコードの例です。ここでは主に、T4やV100などのGPUメモリの少ないGPUを用いることを想定しています。A100またはH100のGPUを用いている方は、本レポジトリのLoRA tuningのサンプルをお試しください。\n",
|
9 |
-
"\n",
|
10 |
-
"- モデル:https://huggingface.co/stockmark/stockmark-13b\n",
|
11 |
-
"- データ:https://github.com/kunishou/databricks-dolly-15k-ja\n",
|
12 |
-
"\n",
|
13 |
-
"以下の例では、学習から推論までの一連の流れを効率よく回すため、学習を0.1 epochのみ行います。T4 GPUで実行すると30分ほどかかります。\n",
|
14 |
-
"\n",
|
15 |
-
"また、ここで用いられているハイパーパラメータは最適化されたものではありませんので、必要に応じて調整してください。"
|
16 |
-
]
|
17 |
-
},
|
18 |
-
{
|
19 |
-
"cell_type": "markdown",
|
20 |
-
"id": "93b3f4b5-2825-4ef3-a0ee-7a60155aee5d",
|
21 |
-
"metadata": {},
|
22 |
-
"source": [
|
23 |
-
"# 準備\n",
|
24 |
-
"\n",
|
25 |
-
"Google Colaboratoryで実行される方は、以下のライブラリをインストールしてください。"
|
26 |
-
]
|
27 |
-
},
|
28 |
-
{
|
29 |
-
"cell_type": "code",
|
30 |
-
"execution_count": null,
|
31 |
-
"id": "7e05d465-4851-46fe-984d-942cab5e09a2",
|
32 |
-
"metadata": {},
|
33 |
-
"outputs": [],
|
34 |
-
"source": [
|
35 |
-
"!python3 -m pip install transformers accelerate datasets peft bitsandbytes"
|
36 |
-
]
|
37 |
-
},
|
38 |
-
{
|
39 |
-
"cell_type": "code",
|
40 |
-
"execution_count": null,
|
41 |
-
"id": "6a694ba9-a0fa-4f14-81cf-f35f683ba889",
|
42 |
-
"metadata": {},
|
43 |
-
"outputs": [],
|
44 |
-
"source": [
|
45 |
-
"import torch\n",
|
46 |
-
"import datasets\n",
|
47 |
-
"from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, BitsAndBytesConfig\n",
|
48 |
-
"from peft import get_peft_model, LoraConfig, PeftModel, PeftConfig, prepare_model_for_kbit_training\n",
|
49 |
-
"\n",
|
50 |
-
"model_name = \"stockmark/stockmark-13b\"\n",
|
51 |
-
"peft_model_name = \"stockmark-13b-adapter\"\n",
|
52 |
-
"\n",
|
53 |
-
"prompt_template = \"\"\"### Instruction:\n",
|
54 |
-
"{instruction}\n",
|
55 |
-
"\n",
|
56 |
-
"### Input:\n",
|
57 |
-
"{input}\n",
|
58 |
-
"\n",
|
59 |
-
"### Response:\n",
|
60 |
-
"\"\"\"\n",
|
61 |
-
"\n",
|
62 |
-
"def encode(sample):\n",
|
63 |
-
" prompt = prompt_template.format(instruction=sample[\"instruction\"], input=sample[\"input\"])\n",
|
64 |
-
" target = sample[\"output\"]\n",
|
65 |
-
" input_ids_prompt, input_ids_target = tokenizer([prompt, target]).input_ids\n",
|
66 |
-
" input_ids_target = input_ids_target + [ tokenizer.eos_token_id ]\n",
|
67 |
-
" input_ids = input_ids_prompt + input_ids_target\n",
|
68 |
-
" labels = input_ids.copy()\n",
|
69 |
-
" labels[:len(input_ids_prompt)] = [-100] * len(input_ids_prompt) # ignore label tokens in a prompt for loss calculation\n",
|
70 |
-
" return {\"input_ids\": input_ids, \"labels\": labels}\n",
|
71 |
-
"\n",
|
72 |
-
"def get_collator(tokenizer, max_length):\n",
|
73 |
-
" def collator(batch):\n",
|
74 |
-
" batch = [{ key: value[:max_length] for key, value in sample.items() } for sample in batch ]\n",
|
75 |
-
" batch = tokenizer.pad(batch)\n",
|
76 |
-
" batch[\"labels\"] = [ e + [-100] * (len(batch[\"input_ids\"][0]) - len(e)) for e in batch[\"labels\"] ]\n",
|
77 |
-
" batch = { key: torch.tensor(value) for key, value in batch.items() }\n",
|
78 |
-
" return batch\n",
|
79 |
-
"\n",
|
80 |
-
" return collator\n",
|
81 |
-
"\n",
|
82 |
-
"bnb_config = BitsAndBytesConfig(\n",
|
83 |
-
" load_in_4bit=True,\n",
|
84 |
-
" bnb_4bit_use_double_quant=True,\n",
|
85 |
-
" bnb_4bit_quant_type=\"nf4\",\n",
|
86 |
-
" bnb_4bit_compute_dtype=torch.float16\n",
|
87 |
-
")"
|
88 |
-
]
|
89 |
-
},
|
90 |
-
{
|
91 |
-
"cell_type": "markdown",
|
92 |
-
"id": "51e6cfcf-1ac1-400e-a4bc-ea64375d0f9e",
|
93 |
-
"metadata": {},
|
94 |
-
"source": [
|
95 |
-
"# データセットとモデルのロード"
|
96 |
-
]
|
97 |
-
},
|
98 |
-
{
|
99 |
-
"cell_type": "code",
|
100 |
-
"execution_count": null,
|
101 |
-
"id": "3ac80067-4e60-46c4-90da-05647cf96ccd",
|
102 |
-
"metadata": {},
|
103 |
-
"outputs": [],
|
104 |
-
"source": [
|
105 |
-
"# load_tokenizer\n",
|
106 |
-
"tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
|
107 |
-
"\n",
|
108 |
-
"# prepare dataset\n",
|
109 |
-
"dataset_name = \"kunishou/databricks-dolly-15k-ja\"\n",
|
110 |
-
"dataset = datasets.load_dataset(dataset_name)\n",
|
111 |
-
"dataset = dataset.map(encode)\n",
|
112 |
-
"dataset = dataset[\"train\"].train_test_split(0.1)\n",
|
113 |
-
"train_dataset = dataset[\"train\"]\n",
|
114 |
-
"eval_dataset = dataset[\"test\"]\n",
|
115 |
-
"\n",
|
116 |
-
"# load model\n",
|
117 |
-
"model = AutoModelForCausalLM.from_pretrained(model_name, device_map=\"auto\", quantization_config=bnb_config, torch_dtype=torch.float16)\n",
|
118 |
-
"\n",
|
119 |
-
"peft_config = LoraConfig(\n",
|
120 |
-
" task_type=\"CAUSAL_LM\",\n",
|
121 |
-
" inference_mode=False,\n",
|
122 |
-
" target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\"],\n",
|
123 |
-
" r=16,\n",
|
124 |
-
" lora_alpha=32,\n",
|
125 |
-
" lora_dropout=0.05\n",
|
126 |
-
")\n",
|
127 |
-
"\n",
|
128 |
-
"model = get_peft_model(model, peft_config)\n",
|
129 |
-
"model.print_trainable_parameters()"
|
130 |
-
]
|
131 |
-
},
|
132 |
-
{
|
133 |
-
"cell_type": "markdown",
|
134 |
-
"id": "9b471da0-7fba-4127-8b07-22da4cbee6a9",
|
135 |
-
"metadata": {},
|
136 |
-
"source": [
|
137 |
-
"# QLoRA Tuning"
|
138 |
-
]
|
139 |
-
},
|
140 |
-
{
|
141 |
-
"cell_type": "code",
|
142 |
-
"execution_count": null,
|
143 |
-
"id": "b9bafa12-538c-4abb-b8b3-bffeb0990b46",
|
144 |
-
"metadata": {},
|
145 |
-
"outputs": [],
|
146 |
-
"source": [
|
147 |
-
"training_args = TrainingArguments(\n",
|
148 |
-
" output_dir=\"./log_stockmark_13b\",\n",
|
149 |
-
" learning_rate=2e-4,\n",
|
150 |
-
" per_device_train_batch_size=1,\n",
|
151 |
-
" gradient_accumulation_steps=16,\n",
|
152 |
-
" per_device_eval_batch_size=1,\n",
|
153 |
-
" num_train_epochs=0.1,\n",
|
154 |
-
" logging_strategy='steps',\n",
|
155 |
-
" logging_steps=10,\n",
|
156 |
-
" save_strategy='epoch',\n",
|
157 |
-
" evaluation_strategy='epoch',\n",
|
158 |
-
" load_best_model_at_end=True,\n",
|
159 |
-
" metric_for_best_model=\"eval_loss\",\n",
|
160 |
-
" greater_is_better=False,\n",
|
161 |
-
" save_total_limit=2\n",
|
162 |
-
")\n",
|
163 |
-
"\n",
|
164 |
-
"trainer = Trainer(\n",
|
165 |
-
" model=model,\n",
|
166 |
-
" args=training_args,\n",
|
167 |
-
" train_dataset=train_dataset,\n",
|
168 |
-
" eval_dataset=eval_dataset,\n",
|
169 |
-
" data_collator=get_collator(tokenizer, 256)\n",
|
170 |
-
")\n",
|
171 |
-
"\n",
|
172 |
-
"# QLoRA tuning\n",
|
173 |
-
"trainer.train()\n",
|
174 |
-
"\n",
|
175 |
-
"# save model\n",
|
176 |
-
"model = trainer.model\n",
|
177 |
-
"model.save_pretrained(peft_model_name)"
|
178 |
-
]
|
179 |
-
},
|
180 |
-
{
|
181 |
-
"cell_type": "markdown",
|
182 |
-
"id": "a3f80a8e-1ac2-4bdc-8232-fe0ee18ffff5",
|
183 |
-
"metadata": {},
|
184 |
-
"source": [
|
185 |
-
"# 学習したモデルのロード(Optional)\n",
|
186 |
-
"異なるセッションでモデルを読み込む場合、まず最初の準備のセクションのコードを実行して、このコードを実行してください。"
|
187 |
-
]
|
188 |
-
},
|
189 |
-
{
|
190 |
-
"cell_type": "code",
|
191 |
-
"execution_count": null,
|
192 |
-
"id": "43241395-3035-4cb9-8c1c-45ffe8cd48be",
|
193 |
-
"metadata": {},
|
194 |
-
"outputs": [],
|
195 |
-
"source": [
|
196 |
-
"tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
|
197 |
-
"model = AutoModelForCausalLM.from_pretrained(model_name, device_map=\"auto\", quantization_config=bnb_config, torch_dtype=torch.float16)\n",
|
198 |
-
"model = PeftModel.from_pretrained(model, peft_model_name)"
|
199 |
-
]
|
200 |
-
},
|
201 |
-
{
|
202 |
-
"cell_type": "markdown",
|
203 |
-
"id": "2ce4db1f-9bad-4c8e-9c04-d1102b299f24",
|
204 |
-
"metadata": {},
|
205 |
-
"source": [
|
206 |
-
"# 推論"
|
207 |
-
]
|
208 |
-
},
|
209 |
-
{
|
210 |
-
"cell_type": "code",
|
211 |
-
"execution_count": null,
|
212 |
-
"id": "d7d6359b-e0ac-49df-a178-39bb9f79ca93",
|
213 |
-
"metadata": {},
|
214 |
-
"outputs": [],
|
215 |
-
"source": [
|
216 |
-
"prompt = prompt_template.format(instruction=\"自然言語処理とは?\", input=\"\")\n",
|
217 |
-
"\n",
|
218 |
-
"inputs = tokenizer(prompt, return_tensors=\"pt\").to(model.device)\n",
|
219 |
-
"with torch.no_grad():\n",
|
220 |
-
" tokens = model.generate(\n",
|
221 |
-
" **inputs,\n",
|
222 |
-
" max_new_tokens=128,\n",
|
223 |
-
" do_sample=True,\n",
|
224 |
-
" temperature=0.7\n",
|
225 |
-
" )\n",
|
226 |
-
"\n",
|
227 |
-
"output = tokenizer.decode(tokens[0], skip_special_tokens=True)\n",
|
228 |
-
"print(output)"
|
229 |
-
]
|
230 |
-
}
|
231 |
-
],
|
232 |
-
"metadata": {
|
233 |
-
"kernelspec": {
|
234 |
-
"display_name": "Python 3",
|
235 |
-
"language": "python",
|
236 |
-
"name": "python3"
|
237 |
-
},
|
238 |
-
"language_info": {
|
239 |
-
"codemirror_mode": {
|
240 |
-
"name": "ipython",
|
241 |
-
"version": 3
|
242 |
-
},
|
243 |
-
"file_extension": ".py",
|
244 |
-
"mimetype": "text/x-python",
|
245 |
-
"name": "python",
|
246 |
-
"nbconvert_exporter": "python",
|
247 |
-
"pygments_lexer": "ipython3",
|
248 |
-
"version": "3.8.10"
|
249 |
-
}
|
250 |
-
},
|
251 |
-
"nbformat": 4,
|
252 |
-
"nbformat_minor": 5
|
253 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|