Fix Qwen2AudioForConditionalGeneration.forward() and test_flash_attn_kernels_inference_equivalence (#39503)

* Add missing cache_position argument.

* Pass cache_position to language model.

* Overwrite prepare_inputs_for_generation.

* Set model to half precision for Flash Attention test.

* Cast model to bfloat16.

src/transformers/models/qwen2_audio/modeling_qwen2_audio.py

@@ -19,7 +19,6 @@ from dataclasses import dataclass
 from typing import Callable, Optional, Union
 
 import torch
-import torch.utils.checkpoint
 from torch import nn
 
 from ...activations import ACT2FN
@@ -727,6 +726,7 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
     ) -> Union[tuple, Qwen2AudioCausalLMOutputWithPast]:
         r"""
         feature_attention_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
@@ -845,6 +845,7 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
+            cache_position=cache_position,
         )
 
         logits = outputs[0]
@@ -878,5 +879,19 @@ class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMi
             attention_mask=attention_mask,
         )
 
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        # Overwritten -- we should not pass input_features when we are in cached decoding stage
+
+        input_features = kwargs.pop("input_features", None)
+        cache_position = kwargs.get("cache_position")
+
+        model_inputs = super().prepare_inputs_for_generation(*args, **kwargs)
+
+        if cache_position is not None and cache_position[0] == 0:
+            # input_features should only be passed when we are not in cached decoding stage
+            model_inputs["input_features"] = input_features
+
+        return model_inputs
+
 
 __all__ = ["Qwen2AudioForConditionalGeneration", "Qwen2AudioPreTrainedModel", "Qwen2AudioEncoder"]

tests/models/qwen2_audio/test_modeling_qwen2_audio.py

@@ -34,6 +34,7 @@ from transformers.testing_utils import (
     torch_device,
 )
 
+from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
 
@@ -132,14 +133,12 @@ class Qwen2AudioModelTester:
 
 
 @require_torch
-class Qwen2AudioForConditionalGenerationModelTest(ModelTesterMixin, unittest.TestCase):
+class Qwen2AudioForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
     """
     Model tester for `Qwen2AudioForConditionalGeneration`.
     """
 
     all_model_classes = (Qwen2AudioForConditionalGeneration,) if is_torch_available() else ()
-    # Doesn't run generation tests. TODO eustache/joao: some generation tests are broken, the errors seem cache-related
-    all_generative_model_classes = ()
     test_pruning = False
     test_head_masking = False
     _is_composite = True

tests/test_modeling_common.py

@@ -3484,6 +3484,7 @@ class ModelTesterMixin:
             model = model_class(config)
 
             model.to(torch_device)
+            model.to(torch.bfloat16)
             dummy_input = inputs_dict[model.main_input_name][:1]
             if dummy_input.dtype in [torch.float32, torch.float16]:
                 dummy_input = dummy_input.to(torch.bfloat16)