increase timeout duration + timing inference

app.py

@@ -18,6 +18,7 @@ except ImportError:
 
 import functools
 import tempfile
+import time
 import numpy as np
 import scipy.signal
 import torch
@@ -225,12 +226,14 @@ model_card = ModelCard(
 )
 
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=180)
 def process_fn(vocals_path, bass_path, drums_path, other_path):
+    t_start = time.time()
     global model_ready
     if not model_ready:
         super_net.to(DEVICE)
         model_ready = True
+    print(f'[timing] model-to-device: {time.time() - t_start:.2f}s')
 
     paths = {'vocals': vocals_path, 'bass': bass_path, 'drums': drums_path, 'other': other_path}
 
@@ -280,6 +283,7 @@ def process_fn(vocals_path, bass_path, drums_path, other_path):
     max_samples = min(max(samples), int(MAX_LENGTH * SR))
     max_samplingrate = max(samplingrate_list)
     max_subtype = max(subtype_list)
+    print(f'[timing] load+resample stems: {time.time() - t_start:.2f}s')
 
     # --- Build input tensor ---
     data = np.zeros((len(STEMS), 1, max_samples, N_CHANNELS), dtype=np.float32)
@@ -295,11 +299,13 @@ def process_fn(vocals_path, bass_path, drums_path, other_path):
     # round up to the next multiple of KERNEL_SIZE_ENCODER as required by the encoder
     new_samples = (1 + max_samples // KERNEL_SIZE_ENCODER) * KERNEL_SIZE_ENCODER
     data = np.pad(data, [(0, 0), (0, 0), (0, new_samples - max_samples), (0, 0)])
+    print(f'[timing] build input tensor: {time.time() - t_start:.2f}s')
 
     # --- Effect normalization ---
     # changed from org. inference #512 because parallelization may not be safe on HF's CPU tier
     for k, stem in enumerate(STEMS):
         data[k][0] = normalize_audio_wave((data[k][0], EFFECTS, stem, features_mean))
+    print(f'[timing] effect normalization: {time.time() - t_start:.2f}s')
 
     # --- Run model inference ---
     test_data = torch.from_numpy(data).to(DEVICE)
@@ -308,6 +314,7 @@ def process_fn(vocals_path, bass_path, drums_path, other_path):
         audio_out = test_out[DataType.TIME_SAMPLES].cpu().numpy()
         audio_out = audio_out[..., :max_samples]  # trim encoder padding
         audio_out = audio_out[..., pad_samples:]  # trim the 30s context pad added at load
+    print(f'[timing] model inference: {time.time() - t_start:.2f}s')
 
     # first output stem is the mix
     audio_output = audio_out[0, 0, ...]
@@ -317,10 +324,12 @@ def process_fn(vocals_path, bass_path, drums_path, other_path):
         audio_output = librosa.resample(audio_output, orig_sr=SR,
                                         target_sr=max_samplingrate,
                                         res_type='kaiser_best')
+    print(f'[timing] output resample: {time.time() - t_start:.2f}s')
 
     with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as f:
         out_path = f.name
     save_wav(out_path, max_samplingrate, audio_output.T, subtype=max_subtype)
+    print(f'[timing] total: {time.time() - t_start:.2f}s')
 
     return out_path