File size: 6,661 Bytes
fa8b708
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForLanguageModeling, get_scheduler
from huggingface_hub import HfApi, notebook_login
from datasets import load_dataset
from peft import LoraConfig, LoraModel, get_peft_model
from timm.scheduler import CosineLRScheduler
import wandb
import os
from accelerate import Accelerator
import numpy as np
import torch
import tqdm
import torch.nn as nn
import torch.optim as optim

lora_conf = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules="all-linear",
    modules_to_save=None,
)

model_id = "Qwen/Qwen2-1.5B-Instruct"
dataset_id = "HuggingFaceH4/orca-math-word-problems-200k"

model_kwargs = dict(
    use_cache=False,
    #attn_implementation="flash_attention_2",
    torch_dtype="auto",
    device_map="sequential",
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.model_max_length = 2048
model = AutoModelForCausalLM.from_pretrained(model_id, **model_kwargs)
model = get_peft_model(model, lora_conf)

def count_trainable_parameters(model):
    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
    params = sum([np.prod(p.size()) for p in model_parameters])
    return params

trainable_params = format(count_trainable_parameters(model), ",")

epochs = 1
per_dev_batch_size = 1
gradient_accumulation_steps = 20
dtype = torch.bfloat16
learning_rate = 1e-4

train_dataset = load_dataset(dataset_id, split="train_sft").select(range(150000))
test_dataset = load_dataset(dataset_id, split="test_sft").select(range(100))

# def apply_chat_template(example, tokenizer):
#     chat = []
#     convo = example['conversations']
#     for dic in convo:
#         if dic['from'] == 'human':
#             chat = [
#                 {"role": "user", "content": dic['value']},
#                 {"role": "assistant", "content": "skibbidy"}
#             ]

#     example['text'] = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=False, truncation=True)
#     return example

# train_dataset = raw_dataset.select(range(10000))
# test_dataset = raw_dataset.select(range(300))
# column_names = list(train_dataset.features)

def apply_chat_template(example, tokenizer):
    example['text'] = tokenizer.apply_chat_template(example['messages'], tokenize=True, add_generation_prompt=False, truncation=True)
    return example

column_names = list(train_dataset.features)

processed_train_dataset = train_dataset.map(
    apply_chat_template,
    # batched=True,
    # batch_size=20,
    fn_kwargs={"tokenizer": tokenizer},
    num_proc=10,
    remove_columns=column_names,
)

processed_test_dataset = test_dataset.map(
    apply_chat_template,
    # batched=True,
    # batch_size=20,
    fn_kwargs={"tokenizer": tokenizer},
    num_proc=10,
    remove_columns=column_names,
)

data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

train_dataloader = torch.utils.data.DataLoader( # 
    processed_train_dataset['text'],
    batch_size=per_dev_batch_size,
    shuffle=False,
    collate_fn=data_collator
)

test_dataloader = torch.utils.data.DataLoader(
    processed_test_dataset['text'],
    batch_size=per_dev_batch_size,
    shuffle=False,
    collate_fn=data_collator
)

global_step = 0
num_training_steps = epochs * len(train_dataloader)
warmup_ratio = 0.1
warmup_steps = 500
#warmup_steps = int(warmup_ratio * num_training_steps)

optimizer = optim.AdamW(model.parameters(), lr=learning_rate)
cross_entropy = nn.CrossEntropyLoss()

scheduler = get_scheduler(
    name="cosine",
    optimizer=optimizer, 
    num_warmup_steps=warmup_steps,
    num_training_steps=num_training_steps
)

acc = Accelerator(gradient_accumulation_steps=gradient_accumulation_steps)

if acc.is_main_process:
    wandb.init(
        project="qwen-math",
    
        config={
        "learning_rate": learning_rate,
        "dataset": dataset_id,
        "batch_size": per_dev_batch_size,
        "lora_r": lora_conf.r,
        "lora_alpha": lora_conf.lora_alpha,
        "lora_dropout": lora_conf.lora_dropout,
        "gradient_accumulation_steps": gradient_accumulation_steps,
        "warmup_ratio": warmup_ratio,
        "trainable_params": trainable_params,
        "num_training_steps": num_training_steps, 
        "model_name": "TinyLlama"
        }
    )   

optimizer, scheduler, train_dataloader, tokenizer, model, scheduler = acc.prepare(optimizer, scheduler, train_dataloader, tokenizer, model, scheduler)

def save_checkpoint():
    if acc.is_main_process:    
        save_path = os.path.join("checkpoint_math", f"step_{global_step}")
        model.module.save_pretrained(save_path)
    
        print(f"Saved model at step {global_step}")

def calc_metrics():
    model.eval()
    for batch in test_dataloader:
        pred = model(**batch)
        loss = pred.loss

        if acc.is_main_process:
            perplexity = torch.exp(loss)
            wandb.log({"eval_loss": loss.item(), "eval_perplexity": perplexity})

    model.train()

device = acc.device

model.train()
for epoch in range(epochs):
    for step, batch in enumerate(train_dataloader):
        #print(tokenizer.decode(batch['input_ids'][0]))
        
        # outputs = model(**batch)
        # loss = outputs.loss
        
        # acc.backward(loss)
        # wandb.log({"loss": loss.item(), "learning_rate": optimizer.param_groups[0]['lr'], "perplexity": perplexity})

        with acc.accumulate(model):
            #batch = {k: v.to(device) for k, v in batch.items()}
            outputs = model(**batch)
            loss = outputs.loss
            acc.backward(loss)
            optimizer.step()
            scheduler.step()
            optimizer.zero_grad()
                
            if acc.is_main_process:
                perplexity = torch.exp(loss)
                wandb.log({"loss": loss.item(), "learning_rate": optimizer.param_groups[0]['lr'], "perplexity": perplexity})
            
            global_step += 1

        if (step + 1) % 1000 == 0:
            save_checkpoint()
        
        # if (step + 1) % gradient_accumulation_steps == 0:
        #     optimizer.step()
        #     scheduler.step()
        #     optimizer.zero_grad()
        #     global_step += 1
    
        if (step + 1) % 100 == 0 and acc.is_main_process:
            print(f"Loss: {loss.item()}")
            
            
        if (step + 1) % 400 == 0:
            calc_metrics()

        if global_step > num_training_steps:
            break

    if global_step > num_training_steps:
        break

if acc.is_main_process:
    wandb.finish()
    save_checkpoint()