[Apple GPU]系列一之在苹果M芯片上微调30亿参数模型

pip install --upgrade transformerspip install --upgrade datasetspip install --upgrade acceleratepip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

import torchimport randomrandom.seed(42)torch.manual_seed(42)from transformers import LlamaTokenizer, LlamaForCausalLM 
# 验证环境无误if torch.backends.mps.is_available():    mps_device = torch.device("mps")    x = torch.ones(1, device=mps_device)    print (x)else:    print ("MPS device not found.")

# 出现这个结果说明当前设备支持tensor([1.], device='mps:0')

model_path ='openlm-research/open_llama_3b_v2'tokenizer = LlamaTokenizer.from_pretrained(model_path, legacy=True)base_model =LlamaForCausalLM.from_pretrained(model_path)

from peft import LoraConfig, PeftModellora_config = LoraConfig(    r=64,    lora_alpha=32,    lora_dropout=0.05,    bias="none",    task_type="CAUSAL_LM")model = PeftModel(base_model,lora_config,adapter_name="Shakespeare")device = torch.device("mps")model.to(device)

# 配置数据集import os import requestsfile_name="shakespeare.txt"url = "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"if not os.path.isfile(file_name):    data = requests.get(url)    with open(file_name, 'w') as f:        f.write(data.text)from transformers import TextDatasettrain_dataset = TextDataset(tokenizer=tokenizer, file_path=file_name, block_size=128)[:256]

# trainingfrom transformers import Trainer, TrainingArgumentstraining_args = TrainingArguments(    output_dir="out_put",    overwrite_output_dir=True,    num_train_epochs=10,    per_device_eval_batch_size=32,    evaluation_strategy='no')

from transformers import DataCollatorForLanguageModelingdata_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,mlm=False)trainer = Trainer(    model=model,    args=training_args,    data_collator=data_collator,    train_dataset=train_dataset,)

def generate_response(prompt_text, model, tokenizer, max_length=30, num_return_sequences=1):    input_ids = tokenizer.encode(prompt_text, return_tensors="pt").to(device)    attention_mask = torch.ones(input_ids.shape, device=input_ids.device)        output_sequences = model.generate(        input_ids=input_ids,        max_length=max_length,        attention_mask=attention_mask,        num_return_sequences=num_return_sequences,        no_repeat_ngram_size=2,    )        responses =[]    for response_id in output_sequences:        response = tokenizer.decode(response_id, skip_special_tokens=True)        responses.append(response)            return responses
prompt_text = "Uneasy lies the head that wears a crown."responses = generate_response(prompt_text, model, tokenizer)for response in responses:    print(response)

tensor([1.], device='mps:0')Uneasy lies the head that wears a crown.- William ShakespeareThe head of the family is the most important person in the world

trainer.train()

StepTraining Loss--------------------Out[4]:TrainOutput(global_step=320, training_loss=2.709520721435547, metrics={'train_runtime': 1315.7902, 'train_samples_per_second': 1.946, 'train_steps_per_second': 0.243, 'total_flos': 6577270554624000.0, 'train_loss': 2.709520721435547, 'epoch': 10.0})

prompt_text = "Uneasy lies the head that wears a crown."responses = generate_response(prompt_text, model, tokenizer)for response in responses:    print(response)

Uneasy lies the head that wears a crown.MENENIUS :I'll not be a king. I'd rather beA poor man in my country than a princeIn a foreign land.

save_path ="merged_fine_tuned_openllama2_3b_shakespeare_m1max"tokenizer.save_pretrained(save_path)merged_model = model.merge_and_unload()merged_model.save_pretrained(save_path)

https://github.com/AtomGradient/OpenWWDC24