引導式測驗 · 中等 等級
用於設備上 PyTorch 測驗的 ExecuTorch
遵循 ExecuTorch 匯出和降低,了解後端委託和分區,並驗證目標裝置上的相容性。
問題 1 的 8
Which deployment task is ExecuTorch intended to support?
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。
- Which deployment task is ExecuTorch intended to support?
- Training PyTorch models on edge devices during inference
- Exporting arbitrary Python code without input or operator constraints
- Exporting PyTorch programs for execution with supported edge runtimes
- Converting TensorFlow SavedModels into cloud containers
- What does a backend delegate do?
- Runs every model operation even when the backend lacks support
- Runs supported operations through a backend implementation
- Converts a model into a portable artifact without executing operations
- Sets application input validation rules
- What can happen when a model contains operations a delegate does not support?
- Every unsupported operator automatically runs on the delegate
- All unsupported operators always fall back to the portable runtime
- Unsupported operators are approximated without checking outputs
- Some operations may remain on a runtime path, or lowering may fail
- Why can partitioning a graph across backend and runtime add overhead?
- Data movement or synchronization can occur at partition boundaries
- Partitioning guarantees that tensors stay on one device
- Partition boundaries eliminate copies by definition
- Every partition incurs a fixed penalty regardless of workload
- What should be compared after export?
- Compare only the exported artifact name
- Compare only how many nodes the lowered graph contains
- Compare runtime outputs with eager PyTorch on representative inputs
- Compare output values only after changing model preprocessing
- Why test the target device rather than relying only on export success?
- Export proves identical behavior on every device
- Backend coverage and performance depend on the target and build
- The runtime has no platform-specific libraries
- Every delegate supports all exported operators
- Which deployment measurements are useful for an on-device model?
- Training loss and validation accuracy only
- Load time, warm latency, memory, binary size, and energy
- Backend compile time and training throughput only
- Model parameter count and source-code line count
- What should be recorded with an exported model artifact?
- Source version, export settings, backend, and runtime pairing
- Only a portable Python checkpoint without export metadata
- Only example output values without the source version
- The exact private inputs submitted by app users