引导式测验 · 中等 等级
Triton 推理服务器测验
了解 Triton 如何服务模型、批处理请求、安排并发工作以及公开操作信号以调整生产推理系统。
问题 1 的 8
是什么让 Triton 推理服务器“与框架无关”?
预览每个问题8 问题
当您参加测验时,就会出现答案和解释。
- 是什么让 Triton 推理服务器“与框架无关”?
- 单个实例可以同时为来自 PyTorch、TensorFlow、ONNX、TensorRT 和自定义后端的模型提供服务
- 它仅运行 TensorRT 引擎
- 它将每个模型转换为 Java
- 它要求所有模型在一个框架中重新训练
- 动态批处理如何提高性能?
- 通过在每个请求到达时立即处理它,一次处理一个请求
- 通过将同时到达的请求分组为 GPU 的单个批次
- 通过删除到达速度太慢的请求
- 将推理转移到 CPU
- What trade-off can Triton dynamic batching introduce?
- Every request becomes lower latency regardless of traffic
- It raises model accuracy by changing the weights
- Compatible stateless requests may gain throughput while queued requests wait within a configured delay
- It only works when inference runs on a CPU
- Triton“模型集成”(或业务逻辑脚本)可以让您做什么?
- 一起训练多个模型
- 加密模型文件
- 降低模型的准确性
- 将预处理、推理和后处理链接到一个服务器端管道中
- What does Triton concurrent model execution permit?
- Multiple models or instances to execute in parallel on the same system, subject to available resources
- Only one model may run until the server restarts
- A model to change its weights during each request
- Requests to bypass the scheduler and share another model output
- Triton 通常公开哪些协议供客户端发送推理请求?
- 仅FTP
- HTTP/REST 和 gRPC
- 仅电子邮件
- 仅蓝牙
- Which monitoring interface does the Triton guide say it exposes?
- A built-in SQL reporting dashboard
- A browser-only interactive notebook
- Prometheus metrics
- A mandatory cloud billing API
- What should a serving team measure when tuning dynamic batching?
- Only the largest possible batch size
- Only average model parameter count
- Only startup time on an idle server
- Throughput and latency under representative request traffic