- Replace direct vram-boss imports with model-boss HFManagedLoader
- Remove manual GPU coordination (REDIS_URL, vram config)
- SemanticAttributeDetector now accepts pre-loaded model/processor
- VRAM is auto-estimated from HuggingFace model config
- Update from CLIP to SigLIP2 (google/siglip2-so400m-patch14-384)
- Add warmup on startup for faster first request
- Bump version to 0.2.0
Uses lilith-model-boss>=2.3.0 for zero-config GPU coordination.
Consumer code no longer needs to know about vram-boss or VRAM sizes.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
