# 創(chuàng)建虛擬環(huán)境
python -m venv deepseek_env
# 激活虛擬環(huán)境（Linux/Mac）
source deepseek_env/bin/activate
# 激活虛擬環(huán)境（Windows）
deepseek_env\Scripts\activate

# 安裝 PyTorch，根據(jù) CUDA 版本選擇合適的安裝命令
# 若使用 CUDA 11.8
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
# 若不使用 GPU
pip install torch torchvision torchaudio

# 安裝 Transformers 庫(kù)
pip install transformers

# 安裝其他可能需要的庫(kù)
pip install sentencepiece accelerate

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "deepseek-ai/deepseek-llm-7b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# 保存模型和分詞器到本地
model.save_pretrained("./local_deepseek_model")
tokenizer.save_pretrained("./local_deepseek_model")

git lfs install
git clone https://huggingface.co/deepseek-ai/deepseek-llm-7b

from transformers import AutoTokenizer, AutoModelForCausalLM

# 加載本地模型和分詞器
model_path = "./local_deepseek_model"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

# 輸入文本
input_text = "今天天氣怎么樣？"
input_ids = tokenizer.encode(input_text, return_tensors="pt")

# 生成輸出
output = model.generate(input_ids, max_length=100, num_return_sequences=1)
output_text = tokenizer.decode(output[0], skip_special_tokens=True)

print("輸入：", input_text)
print("輸出：", output_text)

from fastapi import FastAPI
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

app = FastAPI()

# 加載本地模型和分詞器
model_path = "./local_deepseek_model"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

if torch.cuda.is_available():
    model = model.cuda()

@app.post("/generate")
async def generate_text(input_text: str):
    input_ids = tokenizer.encode(input_text, return_tensors="pt")
    if torch.cuda.is_available():
        input_ids = input_ids.cuda()

    output = model.generate(input_ids, max_length=100, num_return_sequences=1)
    output_text = tokenizer.decode(output[0], skip_special_tokens=True)

    return {"input": input_text, "output": output_text}

uvicorn main:app --host 0.0.0.0 --port 8000

curl -X POST "http://localhost:8000/generate" -H "Content-Type: application/json" -d '{"input_text": "今天天氣怎么樣？"}'

from transformers import AutoTokenizer, AutoModelForCausalLM
from optimum.onnxruntime import ORTQuantizer, ORTModelForCausalLM
from optimum.onnxruntime.configuration import AutoQuantizationConfig

model_path = "./local_deepseek_model"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

# 量化配置
qconfig = AutoQuantizationConfig.avx512_vnni(is_static=False, per_channel=True)
quantizer = ORTQuantizer.from_pretrained(model)

# 量化模型
quantized_model_path = "./local_deepseek_model_quantized"
quantizer.quantize(save_dir=quantized_model_path, quantization_config=qconfig)

import torch
import torch.distributed as dist
import torch.multiprocessing as mp
from transformers import AutoTokenizer, AutoModelForCausalLM

def setup(rank, world_size):
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '12355'

    # initialize the process group
    dist.init_process_group("nccl", rank=rank, world_size=world_size)

def cleanup():
    dist.destroy_process_group()

def inference(rank, world_size):
    setup(rank, world_size)
    model_path = "./local_deepseek_model"
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    model = AutoModelForCausalLM.from_pretrained(model_path)
    model = model.to(rank)
    model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[rank])

    # 輸入文本
    input_text = "今天天氣怎么樣？"
    input_ids = tokenizer.encode(input_text, return_tensors="pt").to(rank)

    # 生成輸出
    output = model.module.generate(input_ids, max_length=100, num_return_sequences=1)
    output_text = tokenizer.decode(output[0], skip_special_tokens=True)

    print(f"Rank {rank}: 輸入：{input_text}, 輸出：{output_text}")

    cleanup()

if __name__ == "__main__":
    world_size = torch.cuda.device_count()
    mp.spawn(inference, args=(world_size,), nprocs=world_size, join=True)