fix(zerogpu): fix numpy float audio conversion in MiniMax Music 3

app.py

@@ -705,6 +705,7 @@ def generate_zerogpu_music(
         # detailed music description and produces complete vocal songs up to 5 min.
         if repo_id == "MiniMaxAI/MiniMax-Music3":
             from diffusers import ModularPipeline
+            import numpy as np
             if repo_id not in _audio_pipeline_cache:
                 pipe = ModularPipeline.from_pretrained(repo_id)
                 pipe.load_components(dtype=torch.bfloat16)
@@ -712,14 +713,26 @@ def generate_zerogpu_music(
                 _audio_pipeline_cache[repo_id] = pipe
             else:
                 pipe = _audio_pipeline_cache[repo_id]
-            audio = pipe(
+            raw_output = pipe(
                 prompt=prompt.strip(),
                 lyrics=lyrics.strip(),
                 audio_duration=float(dur),
                 generator=torch.Generator("cuda").manual_seed(int(seed)),
                 output="audios",
-            )[0]
-            sf.write(out_audio_path, audio.T.float().cpu().numpy(), pipe.sampling_rate)
+            )
+            audio = raw_output[0]
+            if hasattr(audio, "detach"):
+                audio_np = audio.detach().cpu().float().numpy()
+            elif isinstance(audio, np.ndarray):
+                audio_np = audio.astype(np.float32)
+            else:
+                audio_np = np.asarray(audio, dtype=np.float32)
+
+            if audio_np.ndim > 1 and audio_np.shape[0] < audio_np.shape[1]:
+                audio_np = audio_np.T
+
+            sr = getattr(pipe, "sampling_rate", 44100)
+            sf.write(out_audio_path, audio_np, sr)
             return out_audio_path
 
         # Stable Audio 3 uses its own pipeline and may require HF access approval.