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。
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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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