使用 LLMs 生成¶

pip install transformers bitsandbytes>=0.39.0 -q

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
)

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")

generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

tokenizer.pad_token = tokenizer.eos_token  # 大多数LLM默认情况下没有设置填充标记
model_inputs = tokenizer(
    ["A list of colors: red, blue", "Portugal is"], return_tensors="pt", padding=True
).to("cuda")
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
tokenizer.pad_token = tokenizer.eos_token  # 大多数LLM默认情况下没有填充标记
model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
)

model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")

# 默认情况下，输出将最多包含20个标记。
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# 设置`max_new_tokens`来控制输出的新tokens的最大长度
generated_ids = model.generate(**model_inputs, max_new_tokens=50)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# set_seed 方法在 transformers 库中用于设置随机数生成器的种子。其作用是确保实验的可重复性（reproducibility），即在相同的条件下运行代码时，可以得到相同的结果。
from transformers import set_seed
set_seed(42)

model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")

# LLM + 贪婪解码 = 重复、乏味的输出
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# 通过设置采样`do_sample=True`，输出变得更有创意！
generated_ids = model.generate(**model_inputs, do_sample=True)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# 上面初始化的分词器在默认情况下启用了右填充：
# 第一个序列较短，在序列右侧进行填充，模型将填充标记误认为是有效信息，导致输出的结果不佳。
model_inputs = tokenizer(
    ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
).to("cuda")
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# 使用左填充，模型生成了预期的结果！
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
tokenizer.pad_token = tokenizer.eos_token  # 大多数LLM默认情况下没有填充标记
model_inputs = tokenizer(
    ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
).to("cuda")
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
model = AutoModelForCausalLM.from_pretrained(
    "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
)
set_seed(0)
prompt = """How many helicopters can a human eat in one sitting? Reply as a thug."""
model_inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
input_length = model_inputs.input_ids.shape[1]
generated_ids = model.generate(**model_inputs, max_new_tokens=20)
print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
# 哦，它没有按照我们的指示以暴徒的风格进行回复！
# 让我们看看当我们写一个更好的提示并为这个模型使用正确的模板时会发生什么（通过`tokenizer.apply_chat_template`）

set_seed(0)
messages = [
    {
        "role": "system",
        "content": "You are a friendly chatbot who always responds in the style of a thug",
    },
    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]
model_inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
input_length = model_inputs.shape[1]
generated_ids = model.generate(model_inputs, do_sample=True, max_new_tokens=20)
print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
# 正如我们所看到的，它的回复遵循了适当的暴徒风格 😎