Cache: init empty cache when use_cache (#34274)

* fix

* fix tests

* fix copies

* add docs

* Revert "add docs"

This reverts commit 32d35634f12ba02781d2ebdee0c8dcfbe992a7b9.

* qwen move deltas

* mllama can potentiall fullgraph compile

* enable mllama compile and fix tests

* remove mllama fixes

tests/generation/test_utils.py

@@ -1531,6 +1531,14 @@ class GenerationTesterMixin:
             embed_dim = getattr(text_config, "d_model", text_config.hidden_size)
             per_head_embed_dim = embed_dim // num_attention_heads
 
+            # some models have diffent num-head for query vs key/value so we need to assign correct value
+            # BUT only after `per_head_embed_dim` is set
+            num_attention_heads = (
+                text_config.num_key_value_heads
+                if getattr(text_config, "num_key_value_heads", None) is not None
+                else num_attention_heads
+            )
+
             past_kv = outputs["past_key_values"]
             self.assertEqual(len(past_kv), num_hidden_layers)
 

tests/models/qwen2_vl/test_modeling_qwen2_vl.py

@@ -333,6 +333,10 @@ class Qwen2VLModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCas
     def test_generate_from_inputs_embeds_with_static_cache(self):
         pass
 
+    @unittest.skip(reason="Can't compile fullgraph due to dynamic control flow in `prepare_inputs_for_generate`")
+    def test_generate_compile_fullgraph(self):
+        pass
+
 
 @require_torch
 class Qwen2VLIntegrationTest(unittest.TestCase):

tests/test_modeling_common.py

@@ -2343,7 +2343,8 @@ class ModelTesterMixin:
                             recursive_check(tuple_iterable_value, dict_iterable_value)
                     elif tuple_object is None:
                         return
-                    else:
+                    # model might return non-tensors objects (e.g. Cache class)
+                    elif isinstance(tuple_object, torch.Tensor):
                         self.assertTrue(
                             torch.allclose(
                                 set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5