TraceML brings live PyTorch training visibility

This is the kind of ML infra utility teams often build badly in-house, so a lightweight open-source version has real value. TraceML’s pitch is strong because it focuses on the question practitioners actually ask mid-run: why is training slower or less stable than it should be?

• –The core UX is excellent: `with trace_step(model):` is a much easier sell than asking researchers to stop and open a full profiler
• –Step-level breakdowns for dataloader, forward, backward, optimizer, and memory hit a practical debugging sweet spot for day-to-day training work
• –Median-vs-worst-rank and skew views are especially useful for catching DDP stragglers before they become a bigger cluster efficiency problem
• –The project is still early and explicitly not a replacement for PyTorch Profiler or Nsight, with multi-node and FSDP support still on the roadmap