Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium.

Summary

The post introduces system-level profiling of inference performance on AWS Trainium using TorchNeuron. It explains how to configure torch.profiler with the Neuron backend by passing a NeuronConfig and using NeuronProfiler to capture traces. The walkthrough profiles three implementations of RMSNorm: naive PyTorch operations, the fused torch.rms_norm, and naive operations compiled with torch.compile. The key_averages() table shows that the naive implementation spends 2,355 µs on the Neuron device across six operator dispatches, while both fused and compiled versions reduce device time to about 400 µs with a single dispatch. Neuron Explorer visually confirms the lifecycle of each operation from framework dispatch through runtime scheduling to hardware execution, revealing the gaps that fusion eliminates.

Takeaways

1. 1

   The naive RMSNorm implementation with individual PyTorch operations dispatches six separate ATen operations, each crossing the framework-to-hardware boundary independently, resulting in 2,355 µs of Neuron device time.

2. 2

   Both torch.rms_norm and torch.compile fuse the RMSNorm computation into a single kernel dispatch, reducing Neuron device time to approximately 400 µs by eliminating gaps between hardware executions and avoiding materialization of intermediate tensors.

3. 3

   Neuron Explorer’s dependency chain viewer allows following a single operation from framework dispatch through runtime scheduling to hardware execution, confirming the lifecycle and gaps that fusion removes.

FAQ