TraceML

Your training is slow. Where do you look first?

TraceML turns a completed Hugging Face run into:

  • A diagnosis
  • Supporting evidence
  • A next step
See how it works ↓ Run it in Colab ↗
Run summary

hf_baseline

ResNet-50 · Imagenette · Hugging Face Trainer

200 steps · T4
VerdictINPUT-BOUND
Input Wait took 30% of Step Time.
Step timing465.1 ms
Input wait139.3 ms
H2D4.9 ms
Compute311.0 ms
Residual9.9 ms
GPU utilization74.4%
Peak reserved3.16 GB
Next stepFix input
Integrations

Works with your training stack.

PyTorch Hugging Face PyTorch Lightning Ray Train RF-DETR MONAI Slurm

Workflow

Use the Trainer you already have.

No changes to your existing code. Run the same script with traceml run instead of python. A summary is generated when training ends.

Terminal
pip install traceml-ai
traceml run hf_train.py
Install

Install TraceML in your training environment.

Run

Run your existing script with traceml run instead of python.

Diagnose

Read the limiting phase, supporting timing evidence, and suggested next test.

Compare

Change one setting, rerun, and keep it only when the measured summary improves.

Start broad, then go deep. If compute is limiting, move to PyTorch Profiler or Nsight for operator-level detail.