基于MindSpore NLP的LLM推理 - - 实操

选择线性同余生成器（LCG）解码策略法,重新一下代码
import mindspore
from mindspore.communication import init
from mindnlp.transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "LLM-Research/Meta-Llama-3-8B-Instruct"

init()
tokenizer = AutoTokenizer.from_pretrained(model_id, mirror='modelscope')
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    ms_dtype=mindspore.float16,
    mirror='modelscope',
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="ms"
)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=100,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

import mindspore
from mindspore.communication import init
from mindnlp.transformers import AutoTokenizer, AutoModelForCausalLM
import numpy as np

# LCG 参数设置
a = 1664525
c = 1013904223
m = 2**32

# 初始化种子
seed = 1

def lcg_generator(a, c, m, seed):
    while True:
        seed = (a * seed + c) % m
        yield seed

# 创建LCG生成器
lcg_gen = lcg_generator(a, c, m, seed)

model_id = "LLM-Research/Meta-Llama-3-8B-Instruct"

init()
tokenizer = AutoTokenizer.from_pretrained(model_id, mirror='modelscope')
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    ms_dtype=mindspore.float16,
    mirror='modelscope',
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="ms"
)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

# 获取初始输入长度
input_len = input_ids.shape[-1]

# 初始化输出
outputs = input_ids

# 使用LCG解码策略生成文本
for _ in range(100):  # 生成100个新token
    # 获取当前输入的logits
    logits = model(outputs).logits[:, -1, :]
    
    # 使用LCG生成伪随机数作为采样依据
    random_seed = next(lcg_gen)
    np.random.seed(random_seed)
    
    # 根据logits和随机数进行采样
    probs = mindspore.ops.Softmax()(logits)
    probs = probs.asnumpy()
    next_token = np.random.choice(len(probs[0]), p=probs[0])
    
    # 添加生成的token到输出中
    next_token = mindspore.Tensor([next_token], dtype=mindspore.int64).reshape(1, 1)
    outputs = mindspore.ops.Concat(axis=-1)((outputs, next_token))
    
    # 判断是否遇到终止符
    if next_token in terminators:
        break

response = outputs[0][input_len:]
print(tokenizer.decode(response, skip_special_tokens=True))