Update README — mirror is now bf16 (~15.7 GB, half the fp32 footprint)

README.md

@@ -69,8 +69,8 @@ Generation knobs exposed: prompt, negative prompt, duration, steps, CFG scale, A
 ## Format
 
 - **All weights are `safetensors`.** No `.pt` / `.ckpt` / `.bin` in this mirror.
-- Mirror is **fp32 verbatim** — files were copied from upstream without re-saving. Runtime fp16 cast happens in the inference path (`model_half=True` on CUDA), so on-disk size is larger than the runtime VRAM footprint.
-- Approximate disk sizes per subdir: small variants ~2.2 GB each, medium variants ~8.7 GB each, SAME-S ~0.41 GB, SAME-L ~3.2 GB. Total mirror footprint ≈ 30 GB.
+- Mirror is **bf16** — re-saved via `safetensors.torch.save_model` (preserves shared RotaryEmbedding buffers that bare `save_file` would corrupt). Bytewise this halves disk size vs the fp32 upstream. The MAESTRO runner upcasts to fp32 transiently during `load_state_dict` then casts to fp16 (`model_half=True`) for inference — runtime VRAM is unchanged from the fp32 mirror, but disk + I/O + initial safetensors-read CPU spike are all halved.
+- Approximate disk sizes per subdir: small variants ~1.14 GB each, medium variants ~4.61 GB each, SAME-S ~0.22 GB, SAME-L ~1.70 GB. Total mirror footprint ≈ 15.7 GB.
 
 ## Usage