cutout-sam2-tiny: SAM 2.1 hiera tiny, Copyright (c) Meta Platforms, Inc. and
affiliates. Code: facebookresearch/sam2 at commit
2b90b9f5ceec907a1c18123530e92e794ad901a4. Weights: facebook/sam2.1-hiera-tiny
at revision de431c4043854a71d8101e17995dfe596bf101a5 (sam2.1_hiera_tiny.pt).
Both Apache License 2.0; see LICENSE.

Changed for Core AI, in export_sam2.py:
- The image predictor's own steps are traced as two entry points: `encode`
  (normalisation, image encoder, no-memory embedding) and `decode` (prompt
  encoder for one point, mask decoder, multimask output).
- Weights and arithmetic are float16, except the prompt encoder, which stays
  float32; every input and output is float32.
- The 256x256 mask logits are returned unclamped (the PyTorch predictor clamps
  them to +/-32; the host thresholds them).
The trained weights are otherwise the authors', unmodified.