主流大模型实战对比指南：从Llama到DeepSeek的选型与部署

# 1. 下载模型
huggingface-cli login
huggingface-cli download meta-llama/Llama-2-7b-hf --local-dir ./models/llama-2-7b-hf

# 2. 基础测试
python -c "
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained('./models/llama-2-7b-hf')
model = AutoModelForCausalLM.from_pretrained('./models/llama-2-7b-hf', device_map='auto')
inputs = tokenizer('Hello', return_tensors='pt').to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))
"

from peft import LoraConfig, get_peft_model
from transformers import BitsAndBytesConfig

# 4-bit 量化配置
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4"
)

# LoRA 配置
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

conda create -n qwen python=3.10
conda activate qwen

pip install torch==2.0.1+cu118 --index-url https://download.pytorch.org/whl/cu118
pip install transformers==4.32.0
pip install accelerate==0.24.0
pip install bitsandbytes==0.41.0

# 国内访问 HuggingFace 慢，用镜像
export HF_ENDPOINT=https://hf-mirror.com

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen-7B-Chat-Int8",
    device_map="auto",
    trust_remote_code=True
).eval()

tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen-7B-Chat-Int8",
    trust_remote_code=True
)

response, history = model.chat(tokenizer, "Python 中如何用装饰器实现单例？", history=None)
print(response)

# 降低 temperature 提升稳定性
response, _ = model.chat(
    tokenizer,
    prompt,
    history=None,
    temperature=0.7,  # 从 0.9 降低
    top_p=0.8
)

# 1. 下载模型
git clone https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct

# 2. 转换为 GGUF 格式
cd llama.cpp
make
python3 convert-hf-to-gguf.py \
    ../DeepSeek-Coder-V2-Lite-Instruct \
    --outfile deepseek-coder-v2-lite-f16.gguf \
    --outtype f16

# 3. 量化为 INT4
./quantize \
    deepseek-coder-v2-lite-f16.gguf \
    deepseek-coder-v2-lite-q4_k_m.gguf \
    Q4_K_M

# 4. 推理
./main -m deepseek-coder-v2-lite-q4_k_m.gguf \
    -n 512 \
    -p "Write a Python function to merge two sorted lists:" \
    --gpu-layers 35

# 先全 CPU 跑，看实际需求
./main -m model.gguf -p "test" --gpu-layers 0

# 逐步增加，找到平衡点
./main -m model.gguf -p "test" --gpu-layers 20
./main -m model.gguf -p "test" --gpu-layers 35  # 平衡点

from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "ZhipuAI/glm-4-9b-chat"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map="auto",
    load_in_4bit=True
)

def chat(message, history=[]):
    input_text = tokenizer.apply_chat_template(
        history + [{"role": "user", "content": message}],
        tokenize=False,
        add_generation_prompt=True
    )
    inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7
    )
    return tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)

# 流式生成
for response in model.stream_chat(tokenizer, message, history=history):
    print(response, end='', flush=True)

# 启用 KV Cache（默认开启）
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    use_cache=True,  # 关键
    past_key_values=None
)

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.float16,
    attn_implementation="flash_attention_2"  # Flash Attention 2
)

# 综合优化
# 1. 用 vLLM 替代 Transformers
from vllm import LLM
llm = LLM(model="Qwen/Qwen-7B-Chat")

# 2. 量化
# 3. KV Cache（默认开启）
# 4. Flash Attention
# 5. 批处理

export HF_ENDPOINT=https://hf-mirror.com

pip install modelscope
python -c "from modelscope import snapshot_download; snapshot_download('ZhipuAI/glm-4-9b')"