​​大模型开发从入门到部署：3个月避开87%的新手雷区​​

import torch
# 创建张量并启用GPU加速
tensor_a = torch.tensor([[1, 2], [3, 4]], device='cuda')  # 显存存储
tensor_b = torch.randn(2, 2).cuda()  # 随机张量移至GPU
# 常见计算
result = torch.matmul(tensor_a, tensor_b)  # 矩阵乘法
result = tensor_a @ tensor_b.T  # 等价写法（转置后乘）

x = torch.tensor(3.0, requires_grad=True)
y = x**2 + 2*x + 1
y.backward()  # 自动计算梯度
print(x.grad)  # 输出：8.0 (dy/dx = 2x+2)

import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
# 初始化多进程组
dist.init_process_group(backend='nccl')
model = DDP(model.cuda(), device_ids=[local_rank])
# 数据并行训练
for batch in data_loader:
    outputs = model(batch)
    loss = outputs.loss
    loss.backward()

# 安装PyTorch（CUDA 12.1版本）
pip install torch torchvision torchaudio --index-url https://pypi.tuna.tsinghua.edu.cn/simple

from torch.utils.data import Dataset, DataLoader
class TextDataset(Dataset):
    def __init__(self, texts, labels):
        self.texts = texts
        self.labels = labels
        
    def __len__(self):
        return len(self.texts)
    
    def __getitem__(self, idx):
        return {"input": self.texts[idx], "label": self.labels[idx]}
# 实例化并分批次加载
dataset = TextDataset(["text1", "text2"], [0, 1])
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

import torch.optim as optim
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
optimizer = optim.AdamW(model.parameters(), lr=5e-5)
for epoch in range(3):
    for batch in dataloader:
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()
# 保存模型
torch.save(model.state_dict(), "model.pt")

# 加载微调后的模型
model.load_state_dict(torch.load("model.pt"))
model.ｅｖａｌ()  # 切换为评估模式
# 执行预测
input_text = "这个产品体验非常好"
inputs = tokenizer(input_text, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
prediction = torch.argmax(logits, dim=1).item()  # 输出分类结果

from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceEmbeddings
# 构建向量数据库
embeddings = HuggingFaceEmbeddings(model_name="BAAI/bge-base-zh")
vector_db = Chroma.from_documents(docs, embeddings)
# 检索增强生成
retriever = vector_db.as_retriever()
results = retriever.invoke("量子计算原理")

from langchain.agents import Tool, initialize_agent
tools = [
    Tool(name="WebSearch", func=duckduckgo_search.run, description="搜索引擎"),
    Tool(name="Calculator", func=math_calculator, description="数学计算")
]
agent = initialize_agent(tools, llm, agent_type="structured_chat")
agent.run("特斯拉当前股价是多少？2023年营收增长率是多少？")