Model performance
Policy profiling
Fixed-batch inference and training measurements.
Policies
—in this runPassed
—all requested operationsNeeds attention
—pending| Fetching profiling artifacts… |
Recorded measurements
—
| Run | Median · ms | p95 · ms | GPU | Setup | Commit | Status |
|---|
What is measured. Median latency in milliseconds, after warmup, on one fixed preprocessed batch. Inference computes a fresh action, including chunk generation. Training includes backward and optimizer updates. Memory is peak allocated GPU memory in GiB. Downloads, decoding, preprocessing and optional traces are outside the timing samples. These fixtures measure compute, not policy quality.
Source: huggingface/lerobot · GitHub Actions artifactsDownload JSON ↓