引導式測驗 · 中等 等級

TensorRT 和推理引擎測驗

應用 TensorRT 引擎優化、精度選擇和形狀配置文件,包括為什麼部署相容性和準確性需要實際驗證。

相關引導路徑張量和推理引擎
問題 1 的 8

使用 TensorRT 構建推論引擎的目標是什麼?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. 使用 TensorRT 構建推論引擎的目標是什麼?
    • Guarantee the fastest possible execution on every system
    • Optimize execution of a supported trained model for the configured target
    • Choose training labels without a dataset
    • Replace every GPU with a CPU
  2. 層融合如何加速推理?
    • 透過將多個操作合併到一個 GPU 核心來減少記憶體流量
    • 透過刪除不需要的圖層
    • 透過在單獨的 GPU 上運行每一層
    • 透過轉換模型在CPU上運行
  3. 在代表性資料 INT8 校準工作流程中,為什麼要檢查模型的值分佈?
    • To guarantee no accuracy change under all future inputs
    • To retrain the model from zero automatically
    • To select quantization scales suited to the observed numerical ranges
    • To remove every compatibility requirement
  4. 為什麼編譯後的 TensorRT 引擎通常是特定於一種 GPU 架構?
    • 因為它被加密為該 GPU 的序號
    • 因為 NVIDIA 按設備授予許可
    • 因為它儲存了引擎內部的訓練數據
    • 因為內核自動調整會根據 GPU 的精確核心和記憶體佈局烘焙最快的內核
  5. TensorRT-LLM 的哪些功能特別有助於高效率服務大型語言模型?
    • 動態(連續)批次和分頁 KV 緩存
    • 自動影像增強
    • 內建網頁寄存
    • 音訊降噪
  6. Why evaluate task accuracy after choosing lower-precision TensorRT execution?
    • Lower precision always preserves every output bit
    • Only the model filename affects predictions
    • Reduced numerical precision can change model outputs or quality
    • Accuracy evaluation is needed only during tokenization
  7. A TensorRT dynamic-shape engine receives input outside its configured profile range. What should deployment validation check?
    • Whether the engine can accept every shape without a profile
    • Only whether the model name is short
    • Whether training can restart from a screenshot
    • Whether the profile supports that requested shape
  8. A TensorRT engine is moved to a different runtime and GPU configuration. What must the team establish?
    • That any serialized engine runs unchanged everywhere
    • That the selected compatibility settings support the new environment
    • That the engine contains all its original training data
    • That every GPU shares one serial number