vLLM & AI 推理核心概念全景解析

graph TB
    subgraph GPU["GPU 硬件层"]
        GPUArch[GPU 架构
CUDA/Tensor Core]
        HBM[HBM 显存层次]
        Quant[量化
FP8/INT4/AWQ]
    end

    subgraph Model["模型层"]
        Trans[Transformer]
        PE[位置编码
RoPE/ALiBi]
        Attn[Attention 机制
MHA/GQA/MQA]
        Act[激活函数
SwiGLU/GELU]
        Norm[归一化
RMSNorm/LayerNorm]
        MoE[MoE 混合专家]
        Mamba[SSM / Mamba]
        Tokenizer[Tokenization
BPE/SentencePiece]
    end

    subgraph Infra["推理引擎层"]
        PA[PagedAttention]
        KV[KV Cache 管理]
        Prefill[Prefill 阶段]
        Decode[Decode 阶段]
        CB[Continuous Batching]
        CP[Chunked Prefill]
        PC[Prefix Caching]
        SD[Speculative Decoding]
        FlashAttn[FlashAttention]
    end

    subgraph Dist["分布式层"]
        TP[TP 张量并行]
        PP[PP 流水线并行]
        EP[EP 专家并行]
        DS[Disaggregated Serving]
    end

    subgraph Serving["服务层"]
        GenT[Generation Task]
        PoolT[Pooling Task]
        Embed[Embedding Model]
        Rerank[Rerank Model]
        Reward[Reward Model]
        LMCache[LMCache]
        SGLang[SGLang]
        Gateway[推理网关]
        LoRA[LoRA / QLoRA]
        K8s[K8s 部署]
    end

    subgraph App["应用层"]
        RAG[RAG 检索增强]
        Agent[AI Agent]
        MCP[MCP 协议]
        RLHF[RLHF / DPO]
        Eval[评估指标
MMLU/HumanEval]
        Multi[多模态 VLM]
    end

    GPU --> Model
    Model --> Infra
    Infra --> Dist
    Dist --> Serving
    Serving --> App

    Trans --> Attn
    Trans --> PE
    Trans --> Act
    Trans --> Norm
    Attn --> Mamba
    MoE --> EP

    PA --> KV
    PA --> PC
    KV --> Prefill
    KV --> Decode
    CB --> CP
    CB --> SD

    Embed --> RAG
    Rerank --> RAG
    Reward --> RLHF
    Embed --> Agent
    LMCache --> KV
    SGLang --> PC
    LoRA --> Reward