NOTES

Mermaid 图集

flowchart LR GGUF[GGUF 模型] --> Inspect[结构检查] Inspect --> Tok[Tokenizer 导出] Inspect --> Payload[Tensor payload 导出] Tok --> Bundle[PS bundle] Bundle --> Image[Q25TOKRT image] Payload --> Golden[Host gol

Qwen2.5/06_Mermaid_Diagrams.md 2026-07-13 约 4 分钟阅读

1. 项目总流程

flowchart LR
    GGUF[GGUF 模型] --> Inspect[结构检查]
    Inspect --> Tok[Tokenizer 导出]
    Inspect --> Payload[Tensor payload 导出]
    Tok --> Bundle[PS bundle]
    Bundle --> Image[Q25TOKRT image]
    Payload --> Golden[Host golden]
    Payload --> Fixture[HLS fixture]
    Fixture --> HLS[CSim / cosim / csynth]
    Image --> PS[PS/A53 UART]
    PS --> Board[ZCU106 transcript]
    Board --> Archive[Validator / archive]

流程或关系图。

2. transformer 算子链

flowchart LR
    Token[token id] --> Emb[Embedding]
    Emb --> Norm1[RMSNorm]
    Norm1 --> QKV[Q / K / V projection]
    QKV --> RoPE[RoPE on Q/K]
    RoPE --> Score[masked score + softmax]
    Score --> Context[value mix]
    Context --> OutProj[attention output projection]
    OutProj --> Add1[attention residual]
    Add1 --> Norm2[RMSNorm]
    Norm2 --> Gate[FFN gate]
    Norm2 --> Up[FFN up]
    Gate --> SwiGLU[SiLU(gate) * up]
    Up --> SwiGLU
    SwiGLU --> Down[FFN down]
    Down --> Add2[MLP residual]
    Add2 --> Final[final norm]
    Final --> LM[lm_head + top-k]

流程或关系图。

3. PS UART 命令分发

flowchart TD
    UART[UART line] --> Dispatch{命令}
    Dispatch -->|TEXT_PROMPT| Encode[BPE image 或 byte fallback]
    Dispatch -->|TEXT_GENERATE| Suffix[固定后缀编码 + detokenize]
    Dispatch -->|NL_PROMPT_IDS| Store[保存 token ids]
    Dispatch -->|NL_GENERATE| Demo[递增 token 演示]
    Dispatch -->|MODEL_LOAD / TOKEN_STEP / TEXT_STEP| Stub[P4 command stub]
    Dispatch -->|INFO / STATUS / PING| Meta[状态或 PONG]
    Encode --> Response[UART response]
    Suffix --> Response
    Store --> Response
    Demo --> Response
    Stub --> Response
    Meta --> Response

流程或关系图。

4. HLS 的外部接口

flowchart LR
    PS[PS scheduler 或 HLS testbench] -->|DDR pointer + AXI-Lite scalar| Kernel[HLS kernel]
    Kernel -->|m_axi read| Weight[quantized weights in DDR]
    Kernel -->|m_axi read| Act[activation / fixture in DDR]
    Kernel -->|m_axi write| Result[vector / scores / top-k in DDR]
    Kernel -->|ap_done / status| PS

流程或关系图。