master
feat: Phase 1 memory optimizations — embedding cache, sequential TE, tiled VAE
refactor: extract device.rs
feat: log approximate model size after each weight file is loaded
polish: GGUF loader cleanup, --verbose flag, seed label, channel check, SAFETY comments, failed run logging
fix: elapsed_secs in log, drop DiT before VAE, deduplicate timestep lookup
refactor: extract lora.rs
refactor: introduce trait Pipeline — dispatch via pipeline::for_model, main stays a thin orchestrator