Skip to content

ml_workload_diagnostics.md

Latest commit

 

History

History
74 lines (56 loc) · 3.19 KB

ml_workload_diagnostics.md

File metadata and controls

74 lines (56 loc) · 3.19 KB

Running a workload with Google Cloud ML Diagnostics Enabled

This guide provides an overview on how to enable ML Diagnostics for your MaxText workload.

Overview

Google Cloud ML Diagnostics is an end-to-end managed platform for ML Engineers to optimize and diagnose their AI/ML workloads on Google Cloud. The product allows ML Engineers to collect and visualize all their workload metrics, configs and profiles with one single platform, all within the same UI. The current product offering focuses on workloads running on XLA-based frameworks (JAX, Pytorch XLA, Tensorflow/Keras) on Google Cloud TPUs and GPUs. Current support is for JAX on Google Cloud TPUs only.

Enabling ML Diagnostics on Maxtext Workload

MaxText has integrated the ML Diagnostics SDK in its code. You can enable ML Diagnostics with the managed-mldiagnostics flag. If this is enabled, it will

  • Create a managed MachineLearning run with all the MaxText configs.
  • Upload profiling traces, if the profiling is enabled by profiler="xplane".
  • Upload training metrics, at the defined log_period interval.

Examples

  1. Enable ML Diagnostics to just capture Maxtext metrics and configs

       python3 -m maxtext.trainers.pre_train.train \
      run_name=${USER}-tpu-job \
      base_output_directory="gs://your-output-bucket/" \
      dataset_path="gs://your-dataset-bucket/" \
      steps=100 \
      log_period=10 \
      managed_mldiagnostics=True

  2. Enable ML Diagnostics to capture Maxtext metrics, configs and singlehost profiles (on the first TPU device)

       python3 -m maxtext.trainers.pre_train.train \
      run_name=${USER}-tpu-job \
      base_output_directory="gs://your-output-bucket/" \
      dataset_path="gs://your-dataset-bucket/" \
      steps=100 \
      log_period=10 \
      profiler=xplane \
      managed_mldiagnostics=True

  3. Enable ML Diagnostics to capture Maxtext metrics, configs and multihost profiles (on all TPU devices)

       python3 -m maxtext.trainers.pre_train.train \
      run_name=${USER}-tpu-job \
      base_output_directory="gs://your-output-bucket/" \
      dataset_path="gs://your-dataset-bucket/" \
      steps=100 \
      log_period=10 \
      profiler=xplane \
      upload_all_profiler_results=True \
      managed_mldiagnostics=True

Users can deploy the workload across all supported environments, including the standard XPK workload types (xpk workload create or xpk workload create-pathways) or by running the workload directly on a standalone TPU VM.