Fix remaining issues in beam score calculation (#27808)

* Fix issues in add and is_done for BeamHypotheses

* make newly added arguments optional for better compatibility

* Directly use cur_len as generated_len, add note for retrocompatibility

* update test expectation

* make cur_len represents the length of the entire sequence including the decoder prompt

* remove redundant if/else in testing

tests/generation/test_framework_agnostic.py

@@ -633,11 +633,7 @@ class GenerationIntegrationTestsMixin:
             "do_sample": False,
             "num_beams": 3,
         }
-        if is_pt:
-            expectation = 20
-        else:
-            # TODO (joao): fix me
-            expectation = 13
+        expectation = 13
 
         tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
         text = """Hello, my dog is cute and"""

tests/models/vision_encoder_decoder/test_modeling_vision_encoder_decoder.py

@@ -800,7 +800,7 @@ class ViT2GPT2ModelIntegrationTest(unittest.TestCase):
 
         preds, scores = generate_step(pixel_values)
 
-        EXPECTED_SCORES = np.array([-0.64145195])
+        EXPECTED_SCORES = np.array([-0.5956343])
         max_diff = np.amax(np.abs(scores - EXPECTED_SCORES))
         self.assertLessEqual(max_diff, 1e-4)