|
| 1 | +import os |
| 2 | + |
| 3 | +import torch |
| 4 | +import torch_npu |
| 5 | +from transformers import GPT2Tokenizer, GPT2LMHeadModel, TrainingArguments, Trainer |
| 6 | +from datasets import load_dataset |
| 7 | +from transformers import DataCollatorForLanguageModeling |
| 8 | + |
| 9 | + |
| 10 | +# 固定随机种子 |
| 11 | +def set_seed(seed=42): |
| 12 | + torch.manual_seed(seed) |
| 13 | + if torch.npu.is_available(): |
| 14 | + torch.npu.manual_seed_all(seed) |
| 15 | + |
| 16 | + |
| 17 | +# 训练并比较 CPU 和 GPU 的训练损失 |
| 18 | +def train_and_compare_gpt2(model_name): |
| 19 | + set_seed() |
| 20 | + |
| 21 | + def train_on_device(use_cpu=False): |
| 22 | + # 加载 GPT-2 模型和 tokenizer |
| 23 | + model = GPT2LMHeadModel.from_pretrained(model_name) |
| 24 | + tokenizer = GPT2Tokenizer.from_pretrained(model_name) |
| 25 | + tokenizer.pad_token = tokenizer.eos_token # GPT-2 没有 pad_token，需要将 eos_token 作为 pad_token |
| 26 | + |
| 27 | + # 加载 wikitext-2 数据集 |
| 28 | + train_dataset = load_dataset('wikitext', 'wikitext-2-raw-v1', split='train', verification_mode="no_checks") |
| 29 | + val_dataset = load_dataset('wikitext', 'wikitext-2-raw-v1', split='validation', verification_mode="no_checks") |
| 30 | + |
| 31 | + def preprocess_function(examples): |
| 32 | + return tokenizer(examples['text'], padding="max_length", truncation=True, max_length=128) |
| 33 | + |
| 34 | + train_dataset = train_dataset.map(preprocess_function, batched=True) |
| 35 | + val_dataset = val_dataset.map(preprocess_function, batched=True) |
| 36 | + |
| 37 | + train_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask']) |
| 38 | + val_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask']) |
| 39 | + |
| 40 | + # 设置训练参数 |
| 41 | + training_args = TrainingArguments( |
| 42 | + output_dir='./results', |
| 43 | + per_device_train_batch_size=4, |
| 44 | + per_device_eval_batch_size=4, |
| 45 | + num_train_epochs=1, |
| 46 | + logging_dir='./logs', |
| 47 | + logging_steps=10, |
| 48 | + eval_strategy='epoch', |
| 49 | + save_strategy='epoch', |
| 50 | + report_to="none", |
| 51 | + use_cpu=use_cpu |
| 52 | + ) |
| 53 | + |
| 54 | + data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False) |
| 55 | + |
| 56 | + # 创建 Trainer |
| 57 | + trainer = Trainer( |
| 58 | + data_collator=data_collator, |
| 59 | + model=model, |
| 60 | + args=training_args, |
| 61 | + train_dataset=train_dataset, |
| 62 | + eval_dataset=val_dataset |
| 63 | + ) |
| 64 | + |
| 65 | + # 训练模型 |
| 66 | + trainer.train() |
| 67 | + |
| 68 | + # 评估模型 |
| 69 | + metrics = trainer.evaluate() |
| 70 | + |
| 71 | + # 返回评估损失 |
| 72 | + return metrics['eval_loss'] |
| 73 | + |
| 74 | + # 在 GPU 上训练（如果有 GPU） |
| 75 | + if torch.npu.is_available(): |
| 76 | + print(f"Training on NPU") |
| 77 | + gpu_loss = train_on_device(False) |
| 78 | + print(f"GPU Training Loss: {gpu_loss:.4f}") |
| 79 | + else: |
| 80 | + gpu_loss = None |
| 81 | + print("No GPU available for training.") |
| 82 | + |
| 83 | + # 在 CPU 上训练 |
| 84 | + if os.getenv("IS_CI"): |
| 85 | + # Skip training when running in CI because it's too slow |
| 86 | + cpu_loss = 3.0 |
| 87 | + else: |
| 88 | + print(f"Training on CPU") |
| 89 | + cpu_loss = train_on_device(True) |
| 90 | + |
| 91 | + print(f"CPU Training Loss: {cpu_loss:.4f}") |
| 92 | + |
| 93 | + return cpu_loss, gpu_loss |
| 94 | + |
| 95 | + |
| 96 | +# 推理并比较 CPU 和 GPU 的推理损失 |
| 97 | +def infer_and_compare_gpt2(model_name): |
| 98 | + set_seed() |
| 99 | + |
| 100 | + def infer_on_device(device: torch.device): |
| 101 | + # 加载 GPT-2 模型和 tokenizer |
| 102 | + model = GPT2LMHeadModel.from_pretrained(model_name).to(device) |
| 103 | + tokenizer = GPT2Tokenizer.from_pretrained(model_name) |
| 104 | + |
| 105 | + # 设置 pad_token 为 eos_token |
| 106 | + tokenizer.pad_token = tokenizer.eos_token |
| 107 | + |
| 108 | + # 推理测试句子 |
| 109 | + test_sentence = "The quick brown fox jumps over the lazy dog." |
| 110 | + inputs = tokenizer(test_sentence, return_tensors="pt", padding=True, truncation=True).to(device) |
| 111 | + |
| 112 | + with torch.no_grad(): |
| 113 | + outputs = model(**inputs, labels=inputs["input_ids"]) |
| 114 | + |
| 115 | + # 计算损失 |
| 116 | + loss = outputs.loss.item() |
| 117 | + return loss |
| 118 | + |
| 119 | + # 在 GPU 上推理（如果有 GPU） |
| 120 | + if torch.npu.is_available(): |
| 121 | + gpu_device = torch.device('npu') |
| 122 | + gpu_loss = infer_on_device(gpu_device) |
| 123 | + print(f"GPU Inference Loss: {gpu_loss:.4f}") |
| 124 | + else: |
| 125 | + gpu_loss = None |
| 126 | + print("No GPU available for inference.") |
| 127 | + |
| 128 | + # 在 CPU 上推理 |
| 129 | + cpu_device = torch.device('cpu') |
| 130 | + cpu_loss = infer_on_device(cpu_device) |
| 131 | + |
| 132 | + print(f"CPU Inference Loss: {cpu_loss:.4f}") |
| 133 | + |
| 134 | + return cpu_loss, gpu_loss |
| 135 | + |
| 136 | + |
| 137 | +# 主函数 |
| 138 | +if __name__ == "__main__": |
| 139 | + model_name = "gpt2" |
| 140 | + |
| 141 | + # 训练并比较训练损失 |
| 142 | + print("Comparing Training Loss:") |
| 143 | + cpu_train_loss, gpu_train_loss = train_and_compare_gpt2(model_name) |
| 144 | + |
| 145 | + # 推理并比较推理损失 |
| 146 | + print("\nComparing Inference Loss:") |
| 147 | + cpu_infer_loss, gpu_infer_loss = infer_and_compare_gpt2(model_name) |
0 commit comments