pub fn gpu_batch_input_limit() -> usizeExpand description
VRAM-adaptive GPU batch-input limit. Bigger buffers mean fewer device dispatches per multi-TB scan; each kernel launch is a fixed ~50-300 µs cost regardless of payload, so doubling the input halves dispatch overhead. Capped by host VRAM (input + transition tables + match output must fit) and by a 1 GiB upper bound so the pre-compile time stays bounded.
| VRAM detected | Input length | Adapter examples |
|---|---|---|
| >= 24 GiB | 1 GiB | RTX 4090 / 5090, A100 / H100 |
| 12 - 23 GiB | 512 MiB | RTX 3090, RTX 4080, M-Max |
| 8 - 11 GiB | 256 MiB | RTX 3080, RTX 4070, M-Pro |
| < 8 GiB / Unknown | 128 MiB | iGPU, software, no-GPU CI runner |
Cached on first call; the result is stable for the process lifetime so routing and cache identities stay consistent across every batch.