REDDIT · REDDIT// 22d agoPRODUCT UPDATE

TraceML adds zero-code PyTorch runtime visibility

TraceML's new watch mode gives PyTorch users a fast, zero-code terminal view of system and process behavior during training while keeping stdout and stderr visible. It is positioned as a lightweight first pass for slow runs, meant to help you spot bottlenecks before reaching for heavier profiling tools.

// ANALYSIS

Sharp idea: this lowers the friction of “just tell me why this run is slow” without turning profiling into a project.

–Best fit is the first diagnostic pass when training feels off and you want to separate input stalls, compute issues, optimizer overhead, or rank imbalance.
–The zero-code flow, `traceml watch train.py`, is the headline feature because it lets people inspect a live run without instrumenting code first.
–The tradeoff is scope: today it’s aimed at single-GPU and single-node DDP workflows, so larger distributed setups still need deeper tooling.

// TAGS

pytorchtrainingprofilingobservabilitycliopen-sourcellmddp

DISCOVERED

22d ago

2026-03-21

PUBLISHED

22d ago

2026-03-20

RELEVANCE

8/ 10

AUTHOR

traceml-ai