[cache] add a test to confirm we can use cache at train time

tests/utils/test_modeling_utils.py

@@ -37,6 +37,7 @@
     AutoModel,
     AutoModelForImageClassification,
     AutoModelForSequenceClassification,
+    DynamicCache,
     LlavaForConditionalGeneration,
     OwlViTForObjectDetection,
     PretrainedConfig,
@@ -1790,6 +1791,43 @@ def test_load_model_with_state_dict_only_low_cpu_mem_usage(self):
         )
         self.assertTrue(check_models_equal(model, model_loaded))
 
+    def test_cache_at_train_time(self):
+        """
+        Some fine-tuning methods require the use of cache, like prefix tuning in PEFT. This test checks that a cache
+        is at train time used if we request it. Related issue: #35648
+        """
+        model = AutoModelForCausalLM.from_pretrained(TINY_MISTRAL)
+        tokenizer = AutoTokenizer.from_pretrained(TINY_MISTRAL)
+        model_inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+
+        # By default it is not training, we have to set it
+        self.assertFalse(model.training)
+        model.train()
+
+        # If we set `use_cache=True` while training, then a cache is returned
+        model_outputs = model(**model_inputs, use_cache=True)
+        self.assertIsInstance(model_outputs.past_key_values, DynamicCache)
+        self.assertTrue(model.training)
+
+        # simulate injecting virtual tokens like in prefix tuning
+        num_virtual_tokens = 3
+        past_key_values = [torch.randn(2, 1, 2, num_virtual_tokens, 8)] * 2
+        past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        model_inputs["attention_mask"] = torch.cat(
+            (
+                model_inputs["attention_mask"],
+                torch.ones(1, num_virtual_tokens).to(model_inputs["attention_mask"].device),
+            ),
+            dim=1,
+        )
+        model_outputs = model(**model_inputs, past_key_values=past_key_values, use_cache=True)
+        self.assertTrue(model.training)
+
+        # We can also disable the cache to skip a few operations, if the training loop doesn't need cache
+        model_outputs = model(**model_inputs, use_cache=False)
+        self.assertIsNone(model_outputs.past_key_values)
+        self.assertTrue(model.training)
+
 
 @slow
 @require_torch