diff --git a/tests/models/gemma3n/test_modeling_gemma3n.py b/tests/models/gemma3n/test_modeling_gemma3n.py index b5beff5fa2de..dbe7de0d3357 100644 --- a/tests/models/gemma3n/test_modeling_gemma3n.py +++ b/tests/models/gemma3n/test_modeling_gemma3n.py @@ -31,7 +31,6 @@ Gemma3nAudioConfig, Gemma3nAudioFeatureExtractor, Gemma3nConfig, - GenerationConfig, StaticCache, is_torch_available, ) @@ -748,7 +747,7 @@ def setUp(self): audio_ds = load_dataset( "etechgrid/28.5k_wavfiles_dataset", "default", data_files="wav_dataset/103-1240-0000.wav" ) - self.audio_file_path = audio_ds["train"][0]["audio"]["path"] + self.audio_file_path = audio_ds["train"][0]["audio"].metadata.path cleanup(torch_device, gc_collect=True) def tearDown(self): @@ -996,15 +995,13 @@ def test_generation_beyond_sliding_window_with_generation_config(self): input_size = inputs.input_ids.shape[-1] self.assertTrue(input_size > model.config.get_text_config().sliding_window) - out = model.generate(**inputs, generation_config=GenerationConfig(max_new_tokens=20, do_sample=False))[ - :, input_size: - ] + out = model.generate(**inputs, max_new_tokens=20, do_sample=False)[:, input_size:] output_text = tokenizer.batch_decode(out) EXPECTED_COMPLETIONS = Expectations({ # FIXME: This test is VERY flaky on ROCm ("cuda", None): [" and I am glad to be here. This is a nice place. This is a nice place.", ", green, yellow, purple, orange, pink, brown, black, white.\n\nHere are"], ("rocm", (9, 4)): [' and I think it makes this place special. This is a nice place. This is a nice place', ', green, yellow, purple, orange, pink, brown, black, white.\n\nHere are'], - ("xpu", None): [" and I think it is very nice. I think it is nice. This is a nice place.", ", green, yellow, purple, orange, pink, brown, black, white.\n\nHere are"], + ("xpu", None): [" and I think it's a nice place to visit. This is a nice place. This is", ", green, yellow, orange, purple, pink, brown, black, white.\n\nHere'"], }).get_expectation() # fmt: skip self.assertEqual(output_text, EXPECTED_COMPLETIONS)