exp3: video VAE bf16 decode

app.py

@@ -203,6 +203,12 @@ def load_models() -> str | None:
         else:
             pipe.transformer.set_attention_backend(ATTENTION)
 
+        # The video VAE ships fp32 (`_keep_in_fp32_modules` over every module); decode it in bf16 instead —
+        # half the decode FLOPs and traffic on Blackwell. The audio VAE keeps its fp32 modules: a bf16
+        # audio VAE decodes the soundtrack roughly 20 dB too quiet (see the module docstring).
+        pipe.vae.to(torch.bfloat16)
+        print(f"[gen] video VAE cast to {next(pipe.vae.parameters()).dtype}", flush=True)
+
         _install_forward_timing(pipe)
 
         if PLACEMENT == "offload":