# MLA 的核心投影逻辑（伪代码）
# latent_kv: [batch, seq, rank]  ← 训练/推理时缓存这个
# q: [batch, seq, heads, head_dim]
# 推理时：q = q_proj(latent_kv) 恢复到 [batch, seq, heads, head_dim]
# k, v = kv_proj(latent_kv) 恢复到 [batch, seq, heads, head_dim]

# 使用 HuggingFace CLI 下载（如网络受限可配镜像）
huggingface-cli download deepseek-ai/DeepSeek-V3 --repo-type model --local-dir ./models/DeepSeek-V3

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "deepseek-ai/DeepSeek-V3"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="bfloat16",  # 推荐用 BF16
    device_map="auto",
)

messages = [{"role": "user", "content": "解释一下 MoE 架构的核心思想"}]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
outputs = model.generate(input_ids, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

vllm serve deepseek-ai/DeepSeek-V3 \
    --dtype bf16 \
    --tensor-parallel-size 8 \
    --max-model-len 131072

from vllm import LLM, SamplingParams
llm = LLM(model="deepseek-ai/DeepSeek-V3", tensor_parallel_size=8)
params = SamplingParams(temperature=0.7, max_tokens=512)
outputs = llm.generate(["MoE 架构的核心思想是什么？"], params)
print(outputs[0].outputs[0].text)

# 需要先转换为 GGUF 格式（社区工具）
llama-cli -m ./DeepSeek-V3-Q4_K_M.gguf \
    -n 512 \
    -p "MoE架构的核心思想是" \
    --temp 0.7

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.deepseek.com"  # 或自行部署的推理服务端点
)

response = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "用一句话解释 MoE"}],
    max_tokens=128,
    temperature=0.7,
)
print(response.choices[0].message.content)

# 安装依赖
pip install peft transformers datasets bitsandbytes accelerate

from peft import LoraConfig, get_peft_model
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from datasets import load_dataset

model = AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-V3",
    load_in_4bit=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3")

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",
)
model = get_peft_model(model, lora_config)

# 训练（以 QA 数据集为例）
dataset = load_dataset("json", data_files="your_dataset.jsonl")
trainer = TrainingArguments(
    output_dir="./deepseek-v3-finetuned",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    optim="paged_adamw_8bit",
    logging_steps=10,
)
trainer.train()

# 在推理时获取专家激活统计（需要修改 forward 或使用 hook）
# 路由结果通常保存在模型的 router_logits 中

def analyze_expert_routing(model, tokenizer, prompts):
    """统计各专家在给定 prompt 上的激活频率"""
    expert_counts = {}
    # 需要在 model.forward 中注册 hook 捕获 router 输出
    # 此处给出概念示例，详见官方分析脚本
    return expert_counts

# LangChain 接入
from langchain.chat_models import ChatOpenAI
from langchain.schema import HumanMessage

llm = ChatOpenAI(
    model="deepseek-chat",
    openai_api_base="https://api.deepseek.com",  # 自部署时替换
    openai_api_key="YOUR_KEY",
)
messages = [HumanMessage(content="给出一个 Python 单例模式的例子")]
print(llm(messages))

# LlamaIndex 接入（RAG 场景）
from llama_index.llms import OpenAILike
from llama_index import VectorStoreIndex, SimpleDirectoryReader

llm = OpenAILike(model="deepseek-chat", api_base="...", api_key="...")
index = VectorStoreIndex.from_documents(SimpleDirectoryReader("./docs").load_data())
query_engine = index.as_query_engine(llm=llm)
response = query_engine.query("DeepSeek-V3 的训练成本是多少？")
print(response)

from fastapi import FastAPI
from vllm import LLM, SamplingParams
import openai

app = FastAPI(title="DeepSeek-V3 Inference API")

# 初始化 vLLM
llm = LLM(model="deepseek-ai/DeepSeek-V3", tensor_parallel_size=8)
sampling_params = SamplingParams(temperature=0.7, max_tokens=512)

@app.post("/v1/chat/completions")
async def chat_completions(request: openai.ChatCompletionRequest):
    outputs = llm.generate([request.messages[0]["content"]], sampling_params)
    return {
        "choices": [{
            "message": {"role": "assistant", "content": outputs[0].outputs[0].text}
        }]
    }