Fix max_length criteria when using inputs_embeds (#28994)

* fix max_length for inputs_embeds

* make style

* Update src/transformers/generation/utils.py

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Static Cache: load models with MQA or GQA (#28975)

* fix

* fix tests

* fix tests

* Update src/transformers/generation/utils.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* more fixes

* make style

---------

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/generation/utils.py

@@ -441,6 +441,9 @@ class GenerationMixin:
             if isinstance(value, torch.Tensor):
                 batch_size = value.shape[0]
                 break
+
+        if "inputs_embeds" in model_kwargs:
+            return torch.ones((batch_size, 0), dtype=torch.long, device=self.device)
         return torch.ones((batch_size, 1), dtype=torch.long, device=self.device) * bos_token_id
 
     def _prepare_attention_mask_for_generation(
@@ -1421,6 +1424,14 @@ class GenerationMixin:
                 )
             generation_config.max_length = generation_config.max_new_tokens + input_ids_length
 
+        # otherwise the total length [inputs-embeds-len + new-tokens-len] will go beyond indicated `max_length``
+        elif (
+            model_input_name == "inputs_embeds"
+            and inputs_tensor.shape[:-1] != input_ids.shape
+            and not self.config.is_encoder_decoder
+        ):
+            generation_config.max_length -= inputs_tensor.shape[1]
+
         # if we don't pass `past_key_values` and a cache_implementation is specified
         if generation_config.cache_implementation in NEED_SETUP_CACHE_CLASSES_MAPPING and not model_kwargs.get(
             "past_key_values", False

tests/generation/test_utils.py

@@ -1963,7 +1963,7 @@ class GenerationTesterMixin:
             )
             self.assertListEqual(
                 outputs_from_embeds[:, inputs_embeds.shape[1] :].tolist(),
-                outputs_from_embeds_wo_ids[:, 1:].tolist(),
+                outputs_from_embeds_wo_ids.tolist(),
             )
 
     def test_generate_continue_from_past_key_values(self):
@@ -2730,6 +2730,20 @@ class GenerationIntegrationTests(unittest.TestCase, GenerationIntegrationTestsMi
                 **model_kwargs,
             )
 
+    def test_max_length_if_input_embeds(self):
+        # PT-only test: TF doesn't have StoppingCriteria
+        article = "Today a dragon flew over Paris."
+        model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
+        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
+        inputs_embeds = model.get_input_embeddings()(input_ids)
+
+        max_length = 20
+        input_len = input_ids.shape[-1]
+        out_gen = model.generate(input_ids=input_ids, max_length=max_length)
+        out_gen_embeds = model.generate(inputs_embeds=inputs_embeds, max_length=max_length)
+        self.assertEqual(out_gen.shape[-1], input_len + out_gen_embeds.shape[-1])
+
     def test_custom_stopping_criteria_overload_error(self):
         # PT-only test: TF doesn't have StoppingCriteria
         article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""