fix: don't disable autograd, guidance sampling needs it

app.py

@@ -213,9 +213,14 @@ def build_gap_mask(audio_len, gap_ms, position_ms, sample_rate, device):
 
 
 @spaces.GPU
-@torch.inference_mode()
 def process_fn(input_audio_path, checkpoint_choice, gap_length_ms, gap_position_ms, steps, guidance_strength):
-    """Fill a gap in the input audio using the selected diffusion model."""
+    """Fill a gap in the input audio using the selected diffusion model.
+
+    Not wrapped in torch.inference_mode(): reconstruction guidance (xi > 0,
+    the default) needs torch.autograd.grad() inside the sampler, which
+    inference_mode tensors can never support. The sampler applies its own
+    torch.no_grad() around the branches that don't need gradients.
+    """
     ckpt_cfg = CHECKPOINTS[checkpoint_choice]
     sample_rate = ckpt_cfg["sample_rate"]
     audio_len = ckpt_cfg["audio_len"]