Upload gen_samples.py with huggingface_hub

gen_samples.py

@@ -53,6 +53,7 @@ def main():
             "--no-iterative-prompt",
             "--chunk-length", "0",
             "--device", "cuda",
+            "--output", out_path,
         ]
 
         print(f"  Generating semantic tokens...")
@@ -63,38 +64,11 @@ def main():
         if result.stderr:
             print(f"  CLI stderr (last 500): {result.stderr[-500:]}")
 
-        # Find generated .pt files
-        pt_files = [f for f in os.listdir(semantic_dir) if f.endswith('.pt')]
-        if not pt_files:
-            print(f"  ❌ No .pt files generated")
-            continue
-
-        print(f"  Generated {len(pt_files)} semantic files")
-
-        # Step 2: Decode semantic tokens to audio using codec
-        import torchaudio
-        import soundfile as sf
-        from fish_speech.models.text2semantic.inference import load_codec_model
-
-        codec = load_codec_model(f"{local_dir}/codec.pth", "cuda", torch.bfloat16)
-
-        for pt_file in pt_files:
-            codes = torch.load(os.path.join(semantic_dir, pt_file), map_location="cuda", weights_only=True)
-            print(f"  Decoding {pt_file}, codes shape: {codes.shape}")
-
-            with torch.no_grad():
-                with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
-                    audio = codec.decode(codes.unsqueeze(0))
-
-            np_audio = audio.squeeze().cpu().float().numpy()
-            sr = getattr(codec, 'sample_rate', 44100)
-            sf.write(out_path, np_audio, sr)
-            dur = len(np_audio) / sr
-            print(f"  ✅ Saved {out_path} ({dur:.1f}s)")
-
-        del codec
-        gc.collect()
-        torch.cuda.empty_cache()
+        if os.path.exists(out_path):
+            sz = os.path.getsize(out_path)
+            print(f"  ✅ Saved {out_path} ({sz/1024:.0f}KB)")
+        else:
+            print(f"  ❌ Output not found: {out_path}")
 
     # Upload
     print(f"\n{'='*60}")