Files
foxhunt/crates
jgrusewski 2d426b5a4f perf(ppo): eliminate GPU roundtrip in metrics + fix MlDevice type
- Replaced compute_metrics_tensors (CPU→GPU→CPU roundtrip: 3 HtoD +
  8 kernels + 4 DtoH) with compute_metrics_cpu (pure f64 two-pass,
  data already on host). Zero GPU transfers for metrics.
- Fixed Device→MlDevice type, cuda_stream() extraction pattern
- Removed dead sample_action method (referenced non-existent Tensor type)
- Fixed GpuVarStore::new() to pass stream arg

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-18 08:31:06 +01:00
..