引導式測驗 · 中等 等級
優化 CPU 上的模型推理測驗
透過執行緒、運行時核心、量化、記憶體行為和端到端測量來調整 CPU 推理,無需假設一種設定獲勝。
問題 1 的 8
Why can increasing CPU inference threads make a concurrent service slower?
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。
- Why can increasing CPU inference threads make a concurrent service slower?
- Additional threads always reduce memory use
- Labels are removed when threads compete
- Threads can oversubscribe cores and contend across requests
- Vector instructions stop working when multiple requests arrive
- What can quantization trade against faster or smaller CPU inference?
- Quantization removes the need for representative inputs
- Lower precision can alter prediction quality
- Quantization eliminates preprocessing costs
- Quantization makes model versioning unnecessary
- Which overhead does operator fusion aim to reduce?
- 輸出類別的數量
- Dataset licensing checks
- The count of available CPU cores
- Intermediate operations or data movement between compatible operators
- Why benchmark the exact deployment CPU and container limits?
- Core availability, memory behavior, and instruction support affect runtime
- The CPU model determines the task labels
- Container limits change the model graph
- Hardware has no effect on benchmark results
- What can dominate end-to-end latency even if model kernels are fast?
- The title used to display the model
- Training split identifiers
- Tokenization, decoding, or feature preparation
- The repository branch name
- How can batch size affect CPU serving?
- Larger batches always reduce each request latency
- Larger batches may improve throughput while adding wait and memory cost
- Batch size affects accuracy but not serving resources
- Batching has no relationship to CPU cache behavior
- What should be rechecked after quantizing or fusing a model?
- Only compare the serialized model file size
- Only verify the API route name
- Only check that training loss was logged
- Recheck prediction quality and input/preprocessing compatibility
- Which optimization workflow is most reliable?
- Profile a baseline, change one factor, then compare quality and end-to-end speed
- Maximize every thread setting without benchmarking
- Change inputs and hardware between comparison runs
- Measure only one warm forward pass