100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 390
After updating the service, engineers need convincing operational verification: Pin container and library versions and record seeds, data and hardware settings. This intervention addresses: A custom model changes predictions after a new container image is pulled despite the same source dataset and random seed. Which post-change check most directly proves the intended result?
- Verify stable training steps and expected validation loss without OOM.
- Reproduce a short run in the pinned image and compare metrics and artifact manifests.
- Compare best-checkpoint performance with the final epoch on holdout data.
- Measure throughput, validation accuracy and synchronization delay versus one GPU.
Explanation
Correct Answer
B:This is the appropriate post-change test: Reproduce a short run in the pinned image and compare metrics and artifact manifests. The change under test is: Pin container and library versions and record seeds, data and hardware settings. The original condition was: A custom model changes predictions after a new container image is pulled despite the same source dataset and random seed. The expected behavior is: Repeated training becomes more explainable with controlled sources of variation.
Incorrect Answers
A:This check relates to another system: A large transformer job on GPU fails when sequence length doubles and memory allocation spikes. It would not verify whether the current problem was fixed: A custom model changes predictions after a new container image is pulled despite the same source dataset and random seed. The appropriate acceptance test is: Reproduce a short run in the pinned image and compare metrics and artifact manifests.
C:This check relates to another system: A deep sequence model's training error falls every epoch while validation error climbs after a specific checkpoint. It would not verify whether the current problem was fixed: A custom model changes predictions after a new container image is pulled despite the same source dataset and random seed. The appropriate acceptance test is: Reproduce a short run in the pinned image and compare metrics and artifact manifests.
D:This check relates to another system: An image classifier fits on one GPU but takes days; several GPU devices and a sharded dataset are available. It would not verify whether the current problem was fixed: A custom model changes predictions after a new container image is pulled despite the same source dataset and random seed. The appropriate acceptance test is: Reproduce a short run in the pinned image and compare metrics and artifact manifests.
Question 391
An ML platform team is resolving 2 independent workstreams. Workstream A: A small gradient-boosted tree model trains in minutes on a CPU but a proposal adds costly GPUs without a measured bottleneck. Workstream B: A model fits on each accelerator but has billions of training examples that one GPU would process too slowly. Select exactly TWO actions, one directly addressing each separate workstream.
- Choose a supported topology and high-bandwidth worker placement for the distributed job.
- Select a right-sized CPU machine for the baseline instead of defaulting to a GPU.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Use data parallelism with synchronized gradient aggregation across workers.
- Evaluate a compatible edge inference accelerator or optimized CPU deployment.
Explanation
Correct Answers
B:Select a right-sized CPU machine for the baseline instead of defaulting to a GPU. Benchmark total cost and elapsed training time on CPU and GPU alternatives. Other model families and larger data may still benefit from accelerators.
D:Use data parallelism with synchronized gradient aggregation across workers. Track per-device sample throughput, gradient synchronization and validation accuracy. Global batch size and network synchronization affect convergence.
Incorrect Answers
A:Choosing a supported topology and high-bandwidth worker placement for the distributed job is relevant when a multi-GPU training job scales badly because all-reduce synchronization saturates a low-bandwidth interconnect; it does not establish the result required by this workstream.
C:The approach “Partition the model across accelerators using a supported tensor or pipeline model-parallel approach” is relevant when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices; it does not establish the result required by this workstream.
E:Evaluating a compatible edge inference accelerator or optimized CPU deployment is relevant when a factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction; it does not establish the result required by this workstream.
Question 392
An engineer is choosing an appropriate design for this workload: A convolutional network takes hours of dense tensor operations on a general-purpose CPU machine. Which action best addresses the actual technical requirement?
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
- Use supported mixed-precision training with suitable loss scaling and precision guards.
- Use a compatible NVIDIA GPU accelerator and supported framework runtime.
- Instrument the input pipeline and resize or repair the actual data-loading bottleneck.
Explanation
Correct Answer
C:Use a compatible NVIDIA GPU accelerator and supported framework runtime. Compare step time and GPU utilization at representative batches. Small batches and excessive host preprocessing can waste GPU capacity.
Incorrect Answers
A:That measure applies when a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient. The required intervention here is different: Use a compatible NVIDIA GPU accelerator and supported framework runtime. The need is a convolutional network takes hours of dense tensor operations on a general-purpose CPU machine.
B:That measure applies when a deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The required intervention here is different: Use a compatible NVIDIA GPU accelerator and supported framework runtime. The need is a convolutional network takes hours of dense tensor operations on a general-purpose CPU machine.
D:That measure applies when an organization spends heavily on accelerated training but GPU memory and compute usage remain near idle during most jobs. The required intervention here is different: Use a compatible NVIDIA GPU accelerator and supported framework runtime. The need is a convolutional network takes hours of dense tensor operations on a general-purpose CPU machine.
Question 393
An ML platform team is resolving 2 independent workstreams. Workstream A: A JAX training workload uses large dense tensor operations and can be compiled for Google's TPU runtime. Workstream B: A long experiment can restart from checkpoints and the team wants lower compute spend despite occasional capacity loss. Select exactly TWO actions, one directly addressing each separate workstream.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Train on an appropriately sized TPU with XLA compilation and device placement.
- Use data parallelism with synchronized gradient aggregation across workers.
- Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan.
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
Explanation
Correct Answers
B:Train on an appropriately sized TPU with XLA compilation and device placement. Measure steady-state throughput after compilation and data pipeline warm-up. Unsupported custom ops and small irregular batches may reduce performance.
D:Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan. Compare actual completed-job costs and restart counts with on-demand alternatives. Capacity is not guaranteed and savings can disappear under repeated restarts.
Incorrect Answers
A:The approach “Partition the model across accelerators using a supported tensor or pipeline model-parallel approach” is relevant when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices; it does not establish the result required by this workstream.
C:Using data parallelism with synchronized gradient aggregation across workers is relevant when a model fits on each accelerator but has billions of training examples that one GPU would process too slowly; it does not establish the result required by this workstream.
E:The approach “Stage suitable input shards onto fast local SSD or other managed high-throughput storage” is relevant when a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient; it does not establish the result required by this workstream.
Question 394
An ML engineering team must resolve this production problem: A transformer runs out of GPU memory during attention even though compute utilization is not high. Which action best addresses the actual technical requirement?
- Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches.
- Select a right-sized CPU machine for the baseline instead of defaulting to a GPU.
- Instrument the input pipeline and resize or repair the actual data-loading bottleneck.
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
Explanation
Correct Answer
A:Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches. Confirm peak memory headroom and effective training throughput. More GPU FLOPS without memory capacity does not solve allocation failures.
Incorrect Answers
B:That measure applies when a small gradient-boosted tree model trains in minutes on a CPU but a proposal adds costly GPUs without a measured bottleneck. The required intervention here is different: Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches. The need is a transformer runs out of GPU memory during attention even though compute utilization is not high.
C:That measure applies when an organization spends heavily on accelerated training but GPU memory and compute usage remain near idle during most jobs. The required intervention here is different: Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches. The need is a transformer runs out of GPU memory during attention even though compute utilization is not high.
D:That measure applies when a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient. The required intervention here is different: Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches. The need is a transformer runs out of GPU memory during attention even though compute utilization is not high.
Question 395
An ML platform team is resolving 2 independent workstreams. Workstream A: A model fits on each accelerator but has billions of training examples that one GPU would process too slowly. Workstream B: A multi-GPU training job scales badly because all-reduce synchronization saturates a low-bandwidth interconnect. Select exactly TWO actions, one directly addressing each separate workstream.
- Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Instrument the input pipeline and resize or repair the actual data-loading bottleneck.
- Choose a supported topology and high-bandwidth worker placement for the distributed job.
- Use data parallelism with synchronized gradient aggregation across workers.
Explanation
Correct Answers
D:Choose a supported topology and high-bandwidth worker placement for the distributed job. Compare weak and strong scaling efficiency across worker counts. Extra accelerators can make jobs slower if interconnect remains the bottleneck.
E:Use data parallelism with synchronized gradient aggregation across workers. Track per-device sample throughput, gradient synchronization and validation accuracy. Global batch size and network synchronization affect convergence.
Incorrect Answers
A:Using eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan is relevant when a long experiment can restart from checkpoints and the team wants lower compute spend despite occasional capacity loss; it does not establish the result required by this workstream.
B:The approach “Partition the model across accelerators using a supported tensor or pipeline model-parallel approach” is relevant when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices; it does not establish the result required by this workstream.
C:The approach “Instrument the input pipeline and resize or repair the actual data-loading bottleneck” is relevant when an organization spends heavily on accelerated training but GPU memory and compute usage remain near idle during most jobs; it does not establish the result required by this workstream.
Question 396
A Google Cloud change request describes the following issue: One giant model does not fit within any single accelerator, but the cluster has multiple connected devices. Which action best addresses the actual technical requirement?
- Use data parallelism with synchronized gradient aggregation across workers.
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
- Choose a supported topology and high-bandwidth worker placement for the distributed job.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
Explanation
Correct Answer
D:Partition the model across accelerators using a supported tensor or pipeline model-parallel approach. Check memory per device and communication versus compute time. Partition boundaries and network links introduce communication bottlenecks.
Incorrect Answers
A:That measure applies when a model fits on each accelerator but has billions of training examples that one GPU would process too slowly. The required intervention here is different: Partition the model across accelerators using a supported tensor or pipeline model-parallel approach. The need is one giant model does not fit within any single accelerator, but the cluster has multiple connected devices.
B:That measure applies when a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient. The required intervention here is different: Partition the model across accelerators using a supported tensor or pipeline model-parallel approach. The need is one giant model does not fit within any single accelerator, but the cluster has multiple connected devices.
C:That measure applies when a multi-GPU training job scales badly because all-reduce synchronization saturates a low-bandwidth interconnect. The required intervention here is different: Partition the model across accelerators using a supported tensor or pipeline model-parallel approach. The need is one giant model does not fit within any single accelerator, but the cluster has multiple connected devices.
Question 397
An ML platform team is resolving 2 independent workstreams. Workstream A: A long experiment can restart from checkpoints and the team wants lower compute spend despite occasional capacity loss. Workstream B: An organization spends heavily on accelerated training but GPU memory and compute usage remain near idle during most jobs. Select exactly TWO actions, one directly addressing each separate workstream.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Instrument the input pipeline and resize or repair the actual data-loading bottleneck.
- Choose a supported topology and high-bandwidth worker placement for the distributed job.
- Use data parallelism with synchronized gradient aggregation across workers.
- Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan.
Explanation
Correct Answers
B:Instrument the input pipeline and resize or repair the actual data-loading bottleneck. Measure images or tokens per second before and after the data-path change. A high accelerator allocation does not guarantee high model throughput.
E:Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan. Compare actual completed-job costs and restart counts with on-demand alternatives. Capacity is not guaranteed and savings can disappear under repeated restarts.
Incorrect Answers
A:The approach “Partition the model across accelerators using a supported tensor or pipeline model-parallel approach” is relevant when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices; it does not establish the result required by this workstream.
C:Choosing a supported topology and high-bandwidth worker placement for the distributed job is relevant when a multi-GPU training job scales badly because all-reduce synchronization saturates a low-bandwidth interconnect; it does not establish the result required by this workstream.
D:Using data parallelism with synchronized gradient aggregation across workers is relevant when a model fits on each accelerator but has billions of training examples that one GPU would process too slowly; it does not establish the result required by this workstream.
Question 398
An engineer is choosing an appropriate design for this workload: A GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient. Which action best addresses the actual technical requirement?
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Use eligible Spot or preemptible accelerator capacity with durable checkpoints and fallback plan.
- Choose adequate accelerator memory or adopt supported activation checkpointing and microbatches.
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
Explanation
Correct Answer
D:Stage suitable input shards onto fast local SSD or other managed high-throughput storage. Validate read throughput and restoration from durable Cloud Storage originals. Ephemeral local SSD contents can be lost on instance termination.
Incorrect Answers
A:That measure applies when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices. The required intervention here is different: Stage suitable input shards onto fast local SSD or other managed high-throughput storage. The need is a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient.
B:That measure applies when a long experiment can restart from checkpoints and the team wants lower compute spend despite occasional capacity loss. The required intervention here is different: Stage suitable input shards onto fast local SSD or other managed high-throughput storage. The need is a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient.
C:That measure applies when a transformer runs out of GPU memory during attention even though compute utilization is not high. The required intervention here is different: Stage suitable input shards onto fast local SSD or other managed high-throughput storage. The need is a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient.
Question 399
An ML platform team is resolving 2 independent workstreams. Workstream A: A multi-GPU training job scales badly because all-reduce synchronization saturates a low-bandwidth interconnect. Workstream B: A small gradient-boosted tree model trains in minutes on a CPU but a proposal adds costly GPUs without a measured bottleneck. Select exactly TWO actions, one directly addressing each separate workstream.
- Evaluate a compatible edge inference accelerator or optimized CPU deployment.
- Choose a supported topology and high-bandwidth worker placement for the distributed job.
- Select a right-sized CPU machine for the baseline instead of defaulting to a GPU.
- Partition the model across accelerators using a supported tensor or pipeline model-parallel approach.
- Use data parallelism with synchronized gradient aggregation across workers.
Explanation
Correct Answers
B:Choose a supported topology and high-bandwidth worker placement for the distributed job. Compare weak and strong scaling efficiency across worker counts. Extra accelerators can make jobs slower if interconnect remains the bottleneck.
C:Select a right-sized CPU machine for the baseline instead of defaulting to a GPU. Benchmark total cost and elapsed training time on CPU and GPU alternatives. Other model families and larger data may still benefit from accelerators.
Incorrect Answers
A:Evaluating a compatible edge inference accelerator or optimized CPU deployment is relevant when a factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction; it does not establish the result required by this workstream.
D:The approach “Partition the model across accelerators using a supported tensor or pipeline model-parallel approach” is relevant when one giant model does not fit within any single accelerator, but the cluster has multiple connected devices; it does not establish the result required by this workstream.
E:Using data parallelism with synchronized gradient aggregation across workers is relevant when a model fits on each accelerator but has billions of training examples that one GPU would process too slowly; it does not establish the result required by this workstream.
Question 400
An ML engineering team must resolve this production problem: A factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction. Which action best addresses the actual technical requirement?
- Use data parallelism with synchronized gradient aggregation across workers.
- Stage suitable input shards onto fast local SSD or other managed high-throughput storage.
- Use supported mixed-precision training with suitable loss scaling and precision guards.
- Evaluate a compatible edge inference accelerator or optimized CPU deployment.
Explanation
Correct Answer
D:Evaluate a compatible edge inference accelerator or optimized CPU deployment. Benchmark offline inference latency and watts on the actual target hardware. Training TPUs or large data-center GPUs are unsuitable for small edge devices.
Incorrect Answers
A:That measure applies when a model fits on each accelerator but has billions of training examples that one GPU would process too slowly. The required intervention here is different: Evaluate a compatible edge inference accelerator or optimized CPU deployment. The need is a factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction.
B:That measure applies when a GPU fleet stalls when repeatedly reading many tiny remote files although model computation is efficient. The required intervention here is different: Evaluate a compatible edge inference accelerator or optimized CPU deployment. The need is a factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction.
C:That measure applies when a deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The required intervention here is different: Evaluate a compatible edge inference accelerator or optimized CPU deployment. The need is a factory sensor must make decisions offline with a strict power budget and cannot rely on cloud-hosted online prediction.
Professional Machine Learning Engineer Practice Tests