nit

audiocraft/models/musicgen.py

@@ -115,7 +115,7 @@ class MusicGen:
                 should we extend the audio each time. Larger values will mean less context is
                 preserved, and shorter value will require extra computations.
         """
-        assert extend_stride <= self.max_duration - 5, "Keep at least 5 seconds of overlap!"
+        assert extend_stride < self.max_duration, "Cannot stride by more than max generation duration."
         self.extend_stride = extend_stride
         self.duration = duration
         self.generation_params = {

tests/models/test_musicgen.py

@@ -13,7 +13,7 @@ from audiocraft.models import MusicGen
 class TestSEANetModel:
     def get_musicgen(self):
         mg = MusicGen.get_pretrained(name='debug', device='cpu')
-        mg.set_generation_params(duration=2.0)
+        mg.set_generation_params(duration=2.0, stride_extend=2.)
         return mg
 
     def test_base(self):
@@ -51,7 +51,7 @@ class TestSEANetModel:
 
     def test_generate_long(self):
         mg = self.get_musicgen()
-        mg.set_generation_params(duration=4.)
+        mg.set_generation_params(duration=4., stride_extend=2.)
         wav = mg.generate(
             ['youpi', 'lapin dort'])
         assert list(wav.shape) == [2, 1, 32000 * 4]