quiz bu lëkkale ak tegtal · Dëgër Niveau
Hugging Face Gaawaay ngir tàggat GPU yu bari
Toxal benn skript PyTorch ci biir aparey yi ak Accelerate bi ngay saytu sharding done yi, wàññi metrik yi, jaxasoo ci njub ak kaaraange poñ saytu.
Laaj 1 de 8
What does Accelerate primarily help a PyTorch training script do?
Xoolaatal laaj bu nekk8 laaj
Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.
- What does Accelerate primarily help a PyTorch training script do?
- Defines the model architecture for the task
- Selects an objective function from the labels
- Run across supported hardware and distributed configurations with shared code
- Generates evaluation data from training examples
- Why pass common training objects through Accelerator preparation?
- Automatically tunes the optimizer learning rate
- To prepare them for the configured device and strategy
- Selects a checkpoint from validation metrics
- Rewrites the model training objective
- What can happen if workers do not participate consistently in distributed collectives?
- The loss becomes a calibrated probability
- The process always exits cleanly after one rank stops
- Distributed communication no longer occurs
- The run can hang while processes wait for one another
- Why may local evaluation metrics need reduction across processes?
- Each process may see only part of the evaluation data
- Every process automatically has a globally reduced metric
- Metric values are stored directly in model weights
- Reduction converts labels to distributed indices
- What should be verified about batch size when moving from one GPU to several?
- Assume it stays equal to the single-GPU batch in every setup
- Determine whether the configured size is per process or global
- Assume prepare always makes batch size equal to label count
- Treat batch size as the number of GPUs
- Why guard shared checkpoint writes in a multi-process job?
- Prevent workers from training at the same time
- Avoid concurrent workers racing to write the same shared path
- Ensure every rank reads different labels
- Synchronize gradients before a checkpoint is written
- Does using multiple GPUs guarantee proportional speedup?
- Speedup must equal the number of GPUs if the script is unchanged
- The model size is the only factor that determines scaling
- Data movement and synchronization can limit measured speedup
- Parallel training cannot run on multiple devices
- Which verification checks a saved multi-device artifact for practical reuse?
- Load it in the intended inference environment
- Keep it accessible only inside the worker process
- Assume it includes the full dataset
- Skip validation because the job completed