- Python 3.14, torch 2.14.1 (cu130 wheels), NumPy 2; frozen parity
references pass unchanged
- trainer.compile (dynamic-shape torch.compile) and trainer.precision
(bfloat16 autocast; graph sums, loss, optimizer in float32)
- CUDA graph steps use device-tensor learning rates, so one capture per
bucket follows the schedule; checkpoints store portable optimizer state
- trainer=paper_fast: compile + CUDA graphs + bfloat16, 5.9 vs 19.2 ms per
GamGam batch for float32 eager on an A100
- loader workers fork explicitly (Python 3.14 defaults to forkserver);
ONNX export keeps the TorchScript exporter
Correction to 5b73bcb: the pre-change float32 eager step was 21.3 ms per
batch (not ~29 ms; that profile used unshuffled ttH-only batches), so the
default path improved by about 1.2x and the large gains are opt-in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>