PyTorch Partner Collaborations
summary
Shows a concrete integration roadmap: the PyTorch profiler + TensorBoard plugin (pip install) surfaces CPU/GPU traces, flags data loading as 20.8% of step time, reveals 86% nvml utilization but poor SM warp use, and a change (DataLoader workers → 10 / larger batch) cuts an example step from ~260 ms to ~125 ms. TorchServe provides model versioning, autoscaling and canary rollouts with Prometheus/Grafana and Inferentia support. KFServing/Kubeflow pipelines and MLflow integrations add model registry, scripted model save/load, Captain interpretability, and Ax/BoTorch HPO.
AI-written summary · about
chapters (3)
related talks
05mVisualization with TensorBoard
26mThunder: Distribute and Optimize Your PyTorch Models With Zero...
47mGet to Know PyTorch - A Cutting Edge Open Source AI Framework for Deep Learning
43mPyTorch Technical Deep Dive
09mTogether Goes Brrr: Threading Research & Production with Torch Compile
13mLightning Talk: Exploring PiPPY, Tensor Parallel and Torchserve for Large...
36mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
23mEfficient Integration of Kubeflow Tools: An ISS Data Case Study on Anomaly Detection