Ray Train / Tune Quiz
Multiple Choice Questions
- What does Ray Train not author automatically?
- A) Underlying worker coordination
- B) Framework process-group setup
- C) All model/data-partition/state-save-and-restore logic
- D) Ray resource requests
Show Answer
Answer: C
Prepare model/data-loader integration and the real model, optimizer, and checkpoint logic.
- What is the reviewed Train V2 default in 2.58.0?
- A) V2 when the environment variable is unset
- B) Only V1 can run
- C) The TorchTrainer import was removed
- D) Ray extras automatically install PyTorch
Show Answer
Answer: A
Distinguish runs explicitly selecting the older implementation. Frameworks are separate dependencies.
- What happens when legacy trainer_resources is set in V2?
- A) It always reserves more controller CPU
- B) A deprecation error is raised
- C) GPU count increases
- D) Tune trial count changes
Show Answer
Answer: B
Controller, training-worker, and Tune-driver resource scopes differ.
- What does Checkpoint.from_directory do?
- A) Automatically captures every model/optimizer/RNG state
- B) References files in a user-prepared checkpoint directory
- C) Deploys the model
- D) Anonymizes the dataset
Show Answer
Answer: B
Author the recovery payload and load the checkpoint returned by get_checkpoint.
- What is the 2.58.0 V2 report participation rule?
- A) Only rank 0 calls it
- B) Every worker reaches the barrier the same number of times
- C) Every metric is automatically averaged
- D) It cannot be called without a checkpoint
Show Answer
Answer: B
Other workers report checkpoint=None even when only rank 0 saves files.
- Does max_failures=0 disable every retry?
- A) Yes
- B) No; controller and preemption retries have separate settings
- C) It always means infinite retries
- D) It only disables Karpenter retries
Show Answer
Answer: B
Reviewed defaults for controller_failure_limit and max_preemption_failures are both -1.
- What is the current V2 Train/Tune integration pattern?
- A) Pass the V2 Trainer instance directly to Tuner
- B) Construct and fit a Trainer inside a function trainable; connect callbacks as needed
- C) The libraries cannot be combined
- D) A separate Kubernetes cluster is always needed
Show Answer
Answer: B
Direct V2 Trainer input raised TuneError in the native check. Integration needs wiring and resource planning.
- What does TuneReportCallback forward?
- A) An automatically averaged worker metric
- B) A second checkpoint upload
- C) The first worker metric dictionary and existing checkpoint path
- D) It always works outside Tune sessions
Show Answer
Answer: C
Construct it in a Tune session. Perform metric aggregation separately.
Short Answer Questions
- Why budget trial drivers and Train workers together?
Show Answer
Drivers can occupy resources required by their nested workers or placement groups. Check concurrency, worker bundles, cluster bounds, and per-node feasibility together.
- What does a successful scalar Tune example not establish?
Show Answer
It does not prove model accuracy, PyTorch/DDP or GPU performance, multi-node checkpoint recovery, or EKS autoscaling. It checks APIs and collection of two scalar-trial results.