Generate: Add new decoding strategy "DoLa" in .generate() (#29619)

Co-authored-by: Joao Gante <joao@huggingface.co>

tests/models/mistral/test_modeling_mistral.py

@@ -555,6 +555,30 @@ class MistralIntegrationTest(unittest.TestCase):
         text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
         self.assertEqual(EXPECTED_TEXT_COMPLETION[self.cuda_compute_capability_major_version], text)
 
+    @slow
+    def test_model_7b_dola_generation(self):
+        # ground truth text generated with dola_layers="low", repetition_penalty=1.2
+        EXPECTED_TEXT_COMPLETION = (
+            """My favourite condiment is 100% ketchup. I love it on everything, and I’m not ash"""
+        )
+        prompt = "My favourite condiment is "
+        tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", use_fast=False)
+        model = MistralForCausalLM.from_pretrained(
+            "mistralai/Mistral-7B-v0.1", device_map="auto", torch_dtype=torch.float16
+        )
+        input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)
+
+        # greedy generation outputs
+        generated_ids = model.generate(
+            input_ids, max_new_tokens=20, temperature=0, dola_layers="low", repetition_penalty=1.2
+        )
+        text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
+        self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
+
+        del model
+        backend_empty_cache(torch_device)
+        gc.collect()
+
     @require_bitsandbytes
     @slow
     @require_flash_attn