File size: 23,790 Bytes
feee6eb
 
 
 
4ec4a3c
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ec4a3c
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
4ec4a3c
feee6eb
4ec4a3c
 
 
 
 
 
 
 
 
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ec4a3c
feee6eb
4ec4a3c
 
 
 
 
 
 
 
 
 
 
 
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ec4a3c
e51648a
4ec4a3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e51648a
 
 
 
 
 
4ec4a3c
e51648a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e51648a
 
 
 
 
 
 
feee6eb
e51648a
feee6eb
 
 
 
e51648a
feee6eb
e51648a
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e51648a
feee6eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "from trl import SFTTrainer\n",
    "from peft import LoraConfig, get_peft_model\n",
    "\n",
    "import os\n",
    "from uuid import uuid4\n",
    "import pandas as pd\n",
    "\n",
    "import subprocess\n",
    "import evaluate\n",
    "import transformers\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "def max_token_len(dataset):\n",
    "    max_seq_length = 0\n",
    "    for row in dataset:\n",
    "        tokens = len(tokenizer(row['text'])['input_ids'])\n",
    "        if tokens > max_seq_length:\n",
    "            max_seq_length = tokens\n",
    "    return max_seq_length"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model Max Length: 1000000000000000019884624838656\n"
     ]
    }
   ],
   "source": [
    "# model_name='TinyLlama/TinyLlama-1.1B-Chat-v0.1'\n",
    "model_name = 'mistralai/Mistral-7B-v0.1'\n",
    "# model_name = 'distilbert-base-uncased'\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "model_max_length = tokenizer.model_max_length\n",
    "print(\"Model Max Length:\", model_max_length)\n",
    "\n",
    "# dataset = load_dataset(\"imdb\", split=\"train\")\n",
    "dataset_name = 'ai-aerospace/ams_data_train_generic_v0.1_100'\n",
    "dataset = load_dataset(dataset_name)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Max token length train: 1121\n",
      "Max token length validation: 38\n",
      "Block size: 2242\n",
      "{'project_name': './llms/ams_data_train-100_91a45e55-876a-4b93-a9e7-70d26238cd33', 'model_name': 'mistralai/Mistral-7B-v0.1', 'repo_id': 'ai-aerospace/ams-data-train-100-81dbb7fc-16f6-4870-a898-c1840e33430d', 'train_data': 'train_data', 'validation_data': 'validation_data', 'data_directory': './fine_tune_data/', 'block_size': 2242, 'model_max_length': 1121, 'logging_steps': -1, 'evaluation_strategy': 'epoch', 'save_total_limit': 1, 'save_strategy': 'epoch', 'mixed_precision': 'fp16', 'lr': 3e-05, 'epochs': 3, 'batch_size': 2, 'warmup_ratio': 0.1, 'gradient_accumulation': 1, 'optimizer': 'adamw_torch', 'scheduler': 'linear', 'weight_decay': 0, 'max_grad_norm': 1, 'seed': 42, 'quantization': 'int4', 'lora_r': 16, 'lora_alpha': 32, 'lora_dropout': 0.05}\n"
     ]
    }
   ],
   "source": [
    "# Write dataset files into data directory\n",
    "data_directory = './fine_tune_data/'\n",
    "\n",
    "# Create the data directory if it doesn't exist\n",
    "os.makedirs(data_directory, exist_ok=True)\n",
    "\n",
    "# Write the train data to a CSV file\n",
    "train_data='train_data'\n",
    "train_filename = os.path.join(data_directory, train_data)\n",
    "dataset['train'].to_pandas().to_csv(train_filename+'.csv', columns=['text'], index=False)\n",
    "max_token_length_train=max_token_len(dataset['train'])\n",
    "print('Max token length train: '+str(max_token_length_train))\n",
    "\n",
    "# Write the validation data to a CSV file\n",
    "validation_data='validation_data'\n",
    "validation_filename = os.path.join(data_directory, validation_data)\n",
    "dataset['validation'].to_pandas().to_csv(validation_filename+'.csv', columns=['text'], index=False)\n",
    "max_token_length_validation=max_token_len(dataset['validation'])\n",
    "print('Max token length validation: '+str(max_token_length_validation))\n",
    "      \n",
    "max_token_length=max(max_token_length_train,max_token_length_validation)\n",
    "# max_token_length=max_token_length_train\n",
    "if max_token_length > model_max_length:\n",
    "    raise ValueError(\"Maximum token length exceeds model limits.\")\n",
    "block_size=2*max_token_length\n",
    "print('Block size: '+str(block_size))\n",
    "\n",
    "# Define project parameters\n",
    "username='ai-aerospace'\n",
    "project_name='./llms/'+'ams_data_train-100_'+str(uuid4())\n",
    "repo_name='ams-data-train-100-'+str(uuid4())\n",
    "\n",
    "model_params={\n",
    "  \"project_name\": project_name,\n",
    "  \"model_name\": model_name,\n",
    "  \"repo_id\": username+'/'+repo_name,\n",
    "  \"train_data\": train_data,\n",
    "  \"validation_data\": validation_data,\n",
    "  \"data_directory\": data_directory,\n",
    "  \"block_size\": block_size,\n",
    "  \"model_max_length\": max_token_length,\n",
    "  \"logging_steps\": -1,\n",
    "  \"evaluation_strategy\": \"epoch\",\n",
    "  \"save_total_limit\": 1,\n",
    "  \"save_strategy\": \"epoch\",\n",
    "  \"mixed_precision\": \"fp16\",\n",
    "  \"lr\": 0.00003,\n",
    "  \"epochs\": 3,\n",
    "  \"batch_size\": 2,\n",
    "  \"warmup_ratio\": 0.1,\n",
    "  \"gradient_accumulation\": 1,\n",
    "  \"optimizer\": \"adamw_torch\",\n",
    "  \"scheduler\": \"linear\",\n",
    "  \"weight_decay\": 0,\n",
    "  \"max_grad_norm\": 1,\n",
    "  \"seed\": 42,\n",
    "  \"quantization\": \"int4\",\n",
    "  \"lora_r\": 16,\n",
    "  \"lora_alpha\": 32,\n",
    "  \"lora_dropout\": 0.05\n",
    "}\n",
    "for key, value in model_params.items():\n",
    "  os.environ[key] = str(value)\n",
    "\n",
    "print(model_params)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "ename": "ValueError",
     "evalue": "FP16 Mixed precision training with AMP or APEX (`--fp16`) and FP16 half precision evaluation (`--fp16_full_eval`) can only be used on CUDA or NPU devices or certain XPU devices (with IPEX).",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mValueError\u001b[0m                                Traceback (most recent call last)",
      "Cell \u001b[0;32mIn[27], line 1\u001b[0m\n\u001b[0;32m----> 1\u001b[0m args_custom\u001b[38;5;241m=\u001b[39m\u001b[43mtransformers\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mTrainingArguments\u001b[49m\u001b[43m(\u001b[49m\n\u001b[1;32m      2\u001b[0m \u001b[43m    \u001b[49m\u001b[43mper_device_train_batch_size\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mbatch_size\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      3\u001b[0m \u001b[43m    \u001b[49m\u001b[43mper_device_eval_batch_size\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mbatch_size\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      4\u001b[0m \u001b[43m    \u001b[49m\u001b[43mgradient_accumulation_steps\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mgradient_accumulation\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      5\u001b[0m \u001b[43m    \u001b[49m\u001b[43mwarmup_ratio\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mwarmup_ratio\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      6\u001b[0m \u001b[43m    \u001b[49m\u001b[43mnum_train_epochs\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mepochs\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      7\u001b[0m \u001b[43m    \u001b[49m\u001b[43mlearning_rate\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mlr\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m      8\u001b[0m \u001b[43m    \u001b[49m\u001b[43mfp16\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[1;32m      9\u001b[0m \u001b[43m    \u001b[49m\u001b[43mlogging_steps\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mlogging_steps\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     10\u001b[0m \u001b[43m    \u001b[49m\u001b[43msave_total_limit\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43msave_total_limit\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     11\u001b[0m \u001b[43m    \u001b[49m\u001b[43mevaluation_strategy\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mevaluation_strategy\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     12\u001b[0m \u001b[43m    \u001b[49m\u001b[43mmetric_for_best_model\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[38;5;124;43mf1\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[43m,\u001b[49m\n\u001b[1;32m     13\u001b[0m \u001b[43m    \u001b[49m\u001b[43moutput_dir\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mmodel_outputs\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m,\u001b[49m\n\u001b[1;32m     14\u001b[0m \u001b[43m    \u001b[49m\u001b[43mlogging_dir\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mmodel_outputs\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m,\u001b[49m\n\u001b[1;32m     15\u001b[0m \u001b[43m    \u001b[49m\u001b[43moptim\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43moptimizer\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     16\u001b[0m \u001b[43m    \u001b[49m\u001b[43mmax_grad_norm\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mmax_grad_norm\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     17\u001b[0m \u001b[43m    \u001b[49m\u001b[43mweight_decay\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mweight_decay\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m     18\u001b[0m \u001b[43m    \u001b[49m\u001b[43mlr_scheduler_type\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mmodel_params\u001b[49m\u001b[43m[\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mscheduler\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m]\u001b[49m\n\u001b[1;32m     19\u001b[0m \u001b[43m)\u001b[49m\n\u001b[1;32m     21\u001b[0m \u001b[38;5;66;03m# Args from medium article\u001b[39;00m\n\u001b[1;32m     22\u001b[0m args_medium\u001b[38;5;241m=\u001b[39mtransformers\u001b[38;5;241m.\u001b[39mTrainingArguments(\n\u001b[1;32m     23\u001b[0m     per_device_train_batch_size\u001b[38;5;241m=\u001b[39m\u001b[38;5;241m8\u001b[39m,\n\u001b[1;32m     24\u001b[0m     per_device_eval_batch_size\u001b[38;5;241m=\u001b[39m\u001b[38;5;241m32\u001b[39m,\n\u001b[0;32m   (...)\u001b[0m\n\u001b[1;32m     40\u001b[0m     report_to\u001b[38;5;241m=\u001b[39m\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mwandb\u001b[39m\u001b[38;5;124m'\u001b[39m  \u001b[38;5;66;03m# enable logging to W&B\u001b[39;00m\n\u001b[1;32m     41\u001b[0m )\n",
      "File \u001b[0;32m<string>:121\u001b[0m, in \u001b[0;36m__init__\u001b[0;34m(self, output_dir, overwrite_output_dir, do_train, do_eval, do_predict, evaluation_strategy, prediction_loss_only, per_device_train_batch_size, per_device_eval_batch_size, per_gpu_train_batch_size, per_gpu_eval_batch_size, gradient_accumulation_steps, eval_accumulation_steps, eval_delay, learning_rate, weight_decay, adam_beta1, adam_beta2, adam_epsilon, max_grad_norm, num_train_epochs, max_steps, lr_scheduler_type, lr_scheduler_kwargs, warmup_ratio, warmup_steps, log_level, log_level_replica, log_on_each_node, logging_dir, logging_strategy, logging_first_step, logging_steps, logging_nan_inf_filter, save_strategy, save_steps, save_total_limit, save_safetensors, save_on_each_node, save_only_model, no_cuda, use_cpu, use_mps_device, seed, data_seed, jit_mode_eval, use_ipex, bf16, fp16, fp16_opt_level, half_precision_backend, bf16_full_eval, fp16_full_eval, tf32, local_rank, ddp_backend, tpu_num_cores, tpu_metrics_debug, debug, dataloader_drop_last, eval_steps, dataloader_num_workers, past_index, run_name, disable_tqdm, remove_unused_columns, label_names, load_best_model_at_end, metric_for_best_model, greater_is_better, ignore_data_skip, fsdp, fsdp_min_num_params, fsdp_config, fsdp_transformer_layer_cls_to_wrap, deepspeed, label_smoothing_factor, optim, optim_args, adafactor, group_by_length, length_column_name, report_to, ddp_find_unused_parameters, ddp_bucket_cap_mb, ddp_broadcast_buffers, dataloader_pin_memory, dataloader_persistent_workers, skip_memory_metrics, use_legacy_prediction_loop, push_to_hub, resume_from_checkpoint, hub_model_id, hub_strategy, hub_token, hub_private_repo, hub_always_push, gradient_checkpointing, gradient_checkpointing_kwargs, include_inputs_for_metrics, fp16_backend, push_to_hub_model_id, push_to_hub_organization, push_to_hub_token, mp_parameters, auto_find_batch_size, full_determinism, torchdynamo, ray_scope, ddp_timeout, torch_compile, torch_compile_backend, torch_compile_mode, dispatch_batches, split_batches, include_tokens_per_second, include_num_input_tokens_seen, neftune_noise_alpha)\u001b[0m\n",
      "File \u001b[0;32m~/Repositories/HuggingFace/fine-tuning-playground/.venv/lib/python3.11/site-packages/transformers/training_args.py:1499\u001b[0m, in \u001b[0;36mTrainingArguments.__post_init__\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m   1488\u001b[0m         \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m--optim adamw_torch_fused with --fp16 requires PyTorch>2.0\u001b[39m\u001b[38;5;124m\"\u001b[39m)\n\u001b[1;32m   1490\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m (\n\u001b[1;32m   1491\u001b[0m     \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mframework \u001b[38;5;241m==\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mpt\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1492\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m is_torch_available()\n\u001b[0;32m   (...)\u001b[0m\n\u001b[1;32m   1497\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mfp16 \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mfp16_full_eval)\n\u001b[1;32m   1498\u001b[0m ):\n\u001b[0;32m-> 1499\u001b[0m     \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\n\u001b[1;32m   1500\u001b[0m         \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mFP16 Mixed precision training with AMP or APEX (`--fp16`) and FP16 half precision evaluation\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1501\u001b[0m         \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m (`--fp16_full_eval`) can only be used on CUDA or NPU devices or certain XPU devices (with IPEX).\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1502\u001b[0m     )\n\u001b[1;32m   1504\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m (\n\u001b[1;32m   1505\u001b[0m     \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mframework \u001b[38;5;241m==\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mpt\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1506\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m is_torch_available()\n\u001b[0;32m   (...)\u001b[0m\n\u001b[1;32m   1513\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mbf16 \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mbf16_full_eval)\n\u001b[1;32m   1514\u001b[0m ):\n\u001b[1;32m   1515\u001b[0m     \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\n\u001b[1;32m   1516\u001b[0m         \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mBF16 Mixed precision training with AMP (`--bf16`) and BF16 half precision evaluation\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1517\u001b[0m         \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m (`--bf16_full_eval`) can only be used on CUDA, XPU (with IPEX), NPU or CPU/TPU/NeuronCore devices.\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m   1518\u001b[0m     )\n",
      "\u001b[0;31mValueError\u001b[0m: FP16 Mixed precision training with AMP or APEX (`--fp16`) and FP16 half precision evaluation (`--fp16_full_eval`) can only be used on CUDA or NPU devices or certain XPU devices (with IPEX)."
     ]
    }
   ],
   "source": [
    "args_custom=transformers.TrainingArguments(\n",
    "    per_device_train_batch_size=model_params['batch_size'],\n",
    "    per_device_eval_batch_size=model_params['batch_size'],\n",
    "    gradient_accumulation_steps=model_params['gradient_accumulation'],\n",
    "    warmup_ratio=model_params['warmup_ratio'],\n",
    "    num_train_epochs=model_params['epochs'],\n",
    "    learning_rate=model_params['lr'],\n",
    "    fp16=True,\n",
    "    logging_steps=model_params['logging_steps'],\n",
    "    save_total_limit=model_params['save_total_limit'],\n",
    "    evaluation_strategy=model_params['evaluation_strategy'],\n",
    "    metric_for_best_model=\"f1\",\n",
    "    output_dir='model_outputs',\n",
    "    logging_dir='model_outputs',\n",
    "    optim=model_params['optimizer'],\n",
    "    max_grad_norm=model_params['max_grad_norm'],\n",
    "    weight_decay=model_params['weight_decay'],\n",
    "    lr_scheduler_type=model_params['scheduler']\n",
    ")\n",
    "\n",
    "# Args from medium article\n",
    "args_medium=transformers.TrainingArguments(\n",
    "    per_device_train_batch_size=8,\n",
    "    per_device_eval_batch_size=32,\n",
    "    gradient_accumulation_steps=4,\n",
    "    warmup_steps=100,\n",
    "    max_steps=12276,\n",
    "    learning_rate=2e-4,\n",
    "    fp16=True,\n",
    "    eval_steps= 1000,\n",
    "    logging_steps=1000,\n",
    "    save_steps=1000,\n",
    "    evaluation_strategy=\"steps\",\n",
    "    do_eval=True,\n",
    "    load_best_model_at_end=True,\n",
    "    metric_for_best_model=\"f1\",\n",
    "    output_dir='model_outputs',\n",
    "    logging_dir='model_outputs',\n",
    "    remove_unused_columns =False, \n",
    "    report_to='wandb'  # enable logging to W&B\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Start trainer\n",
    "# trainer = SFTTrainer(\n",
    "#     model_name,\n",
    "#     train_dataset=dataset,\n",
    "#     dataset_text_field=\"text\",\n",
    "#     max_seq_length=512,\n",
    "# )\n",
    "\n",
    "peft_config = LoraConfig(\n",
    "    r=model_params['lora_r'],\n",
    "    lora_alpha=model_params['lora_alpha'],\n",
    "    lora_dropout=model_params['lora_dropout']\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load the model\n",
    "model = AutoModelForCausalLM.from_pretrained(\n",
    "    model_name,\n",
    "    load_in_4bit=True\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Setting up the LoRA model\n",
    "# import os\n",
    "# os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\"\n",
    "# from transformers import AutoModelForSequenceClassification\n",
    "# from peft import LoraConfig, get_peft_model, TaskType\n",
    "\n",
    "# MODEL =\"xlm-roberta-large\"\n",
    "\n",
    "# config = LoraConfig(\n",
    "#     task_type=\"SEQ_CLS\",\n",
    "#     r=16,\n",
    "#     lora_alpha=16,\n",
    "#     target_modules=[\"query\", \"value\"],  # Targets the attention blocks in the model\n",
    "#     lora_dropout=0.1,\n",
    "#     bias=\"none\",\n",
    "#     modules_to_save=[\"classifier\"],\n",
    "# )\n",
    "\n",
    "# model = AutoModelForSequenceClassification.from_pretrained(\n",
    "#     MODEL,\n",
    "#     num_labels=len(unique_subissues),\n",
    "#     id2label=id2label,\n",
    "#     label2id=label2id,\n",
    "#     ignore_mismatched_sizes=True\n",
    "# )  \n",
    "\n",
    "lora_model = get_peft_model(model, peft_config)\n",
    "lora_model.print_trainable_parameters()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# trainer = SFTTrainer(\n",
    "#     model,\n",
    "#     train_dataset=dataset,\n",
    "#     dataset_text_field=\"text\",\n",
    "#     peft_config=peft_config,\n",
    "#     max_seq_length=model_params['model_max_length']\n",
    "# )\n",
    "\n",
    "# trainer.train()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "f1_metric = evaluate.load(\"f1\")\n",
    "recall_metric = evaluate.load(\"recall\")\n",
    "accuracy_metric = evaluate.load(\"accuracy\")\n",
    "precision_metric = evaluate.load(\"precision\")\n",
    "\n",
    "def compute_metrics(eval_pred):\n",
    "    logits, labels = eval_pred\n",
    "    predictions = np.argmax(logits, axis=-1)\n",
    "    results = {}\n",
    "    results.update(f1_metric.compute(predictions=predictions, references = labels, average=\"macro\"))\n",
    "    results.update(recall_metric.compute(predictions=predictions, references = labels, average=\"macro\"))\n",
    "    results.update(accuracy_metric.compute(predictions=predictions, references = labels))\n",
    "    results.update(precision_metric.compute(predictions=predictions, references = labels, average=\"macro\"))\n",
    "\n",
    "    return results\n",
    "\n",
    "# See https://towardsdatascience.com/fine-tune-your-llm-without-maxing-out-your-gpu-db2278603d78 for details\n",
    "trainer = transformers.Trainer(\n",
    "    model=lora_model,\n",
    "    train_dataset=model_params['train_data'],\n",
    "    eval_dataset=model_params['validation_data'],\n",
    "    compute_metrics=compute_metrics,\n",
    "    args=args_custom\n",
    ")\n",
    "trainer.train()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".venv",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}