fix: transpose VAE latent from [B,C,T] to [B,T,C] before generator

VAE encoder returns channels-first [B, latent_dim, T]; the generator
expects time-first [B, T, latent_dim] (same convention as decode which
already does .transpose(1,2)). Fixes normalize() size mismatch.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

nodes/selva_lora_trainer.py

@@ -338,7 +338,8 @@ class SelvaLoraTrainer:
                 # encode_audio is @inference_mode — .clone() exits inference mode
                 audio_b = audio.unsqueeze(0).to(device)
                 dist = vae_utils.encode_audio(audio_b)
-                x1   = dist.mode().clone().cpu()
+                # VAE outputs [B, latent_dim, T]; generator expects [B, T, latent_dim]
+                x1   = dist.mode().clone().transpose(1, 2).cpu()
 
                 # Text → CLIP features (reuse already-loaded CLIP from inference model)
                 text_clip = feature_utils_orig.encode_text_clip([prompt]).cpu()

train_lora.py

@@ -139,7 +139,8 @@ def extract_audio_latent(audio: torch.Tensor, feature_utils, device, dtype) -> t
     """
     audio_b = audio.unsqueeze(0).to(device, dtype)  # [1, L]
     dist = feature_utils.encode_audio(audio_b)
-    return dist.mode().clone().cpu()                 # [1, seq_len, latent_dim]
+    # VAE outputs [B, latent_dim, T]; generator expects [B, T, latent_dim]
+    return dist.mode().clone().transpose(1, 2).cpu()  # [1, seq_len, latent_dim]
 
 
 # ---------------------------------------------------------------------------