more deletions :D

tests/generation/test_utils.py

@@ -1546,7 +1546,7 @@ def test_past_key_values_format(self):
     @pytest.mark.generate
     @parameterized.expand([(1,), (2,)])
     def test_generate_from_inputs_embeds(self, num_beams):
-      """Tests that we can generate from `inputs_embeds` instead of `input_ids` in LLMs, VLMs, etc"""
+        """Tests that we can generate from `inputs_embeds` instead of `input_ids` in LLMs, VLMs, etc"""
         # When supported, tests that the decoder model can generate from `inputs_embeds` instead of `input_ids`
         # if fails, you should probably update the `prepare_inputs_for_generation` function
         for model_class in self.all_generative_model_classes:

tests/models/idefics/test_modeling_idefics.py

@@ -770,13 +770,6 @@ def test_contrastive_generate_low_memory(self):
     def test_custom_4d_attention_mask(self):
         pass
 
-    @unittest.skip(
-        reason="IDEFICS has specific requirements for working with inputs embeds like passing also the ids and pixels"
-    )
-    @parameterized.expand([(1,), (2,)])
-    def test_generate_from_inputs_embeds_decoder_only(self, num_beams):
-        pass
-
     @unittest.skip(reason="IDEFICS cannot compile due to dynamic control flow when checking inputs")
     def test_generate_compile_fullgraph(self):
         pass

tests/models/idefics2/test_modeling_idefics2.py

@@ -20,7 +20,6 @@
 import unittest
 from io import BytesIO
 
-import pytest
 import requests
 
 from transformers import (
@@ -412,50 +411,6 @@ def test_prompt_lookup_decoding_matches_greedy_search(self):
     def test_flash_attn_2_fp32_ln(self):
         pass
 
-    @pytest.mark.generate
-    def test_generate_from_inputs_embeds_decoder_only(self):
-        # overwrite because IDEFICS needs ids and embeds at the input to be not None
-        for model_class in self.all_generative_model_classes:
-            config, inputs_dict = self.prepare_config_and_inputs_for_generate()
-
-            # Ignore:
-            # a) eos (to always output 20 tokens) and pad (so we don't try to infer the attn mask from the input_ids,
-            #   which would cause a mismatch),
-            config.pad_token_id = config.eos_token_id = -1
-            config.is_decoder = True
-            model = model_class(config).to(torch_device).eval()
-            input_ids = inputs_dict.pop("input_ids")
-
-            # Traditional way of generating text
-            outputs_from_ids = model.generate(
-                input_ids, max_new_tokens=5, return_dict_in_generate=True, output_scores=True
-            )
-            self.assertEqual(outputs_from_ids.sequences.shape, (input_ids.shape[0], input_ids.shape[1] + 5))
-
-            # Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output)
-            inputs_embeds = model.get_input_embeddings()(input_ids)
-            outputs_from_embeds = model.generate(
-                input_ids,
-                inputs_embeds=inputs_embeds,
-                max_new_tokens=5,
-                return_dict_in_generate=True,
-                output_scores=True,
-            )
-            self.assertListEqual(outputs_from_ids.sequences.tolist(), outputs_from_embeds.sequences.tolist())
-
-            # But if we pass different inputs_embeds, we should get different outputs (the output text may be the
-            # same, but the logits will almost surely be different)
-            random_embeds = torch.rand_like(inputs_embeds)
-            outputs_from_rand_embeds = model.generate(
-                input_ids,
-                inputs_embeds=random_embeds,
-                max_new_tokens=5,
-                return_dict_in_generate=True,
-                output_scores=True,
-            )
-            for i in range(len(outputs_from_rand_embeds.scores)):
-                self.assertFalse(torch.allclose(outputs_from_embeds.scores[i], outputs_from_rand_embeds.scores[i]))
-
     # We need to override as we need to prepare such that the image token is the last token
     def test_resize_tokens_embeddings(self):
         (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()

tests/models/idefics3/test_modeling_idefics3.py

@@ -19,7 +19,6 @@
 import unittest
 from io import BytesIO
 
-import pytest
 import requests
 
 from transformers import (
@@ -359,50 +358,6 @@ def test_prompt_lookup_decoding_matches_greedy_search(self):
     def test_flash_attn_2_fp32_ln(self):
         pass
 
-    @pytest.mark.generate
-    def test_generate_from_inputs_embeds_decoder_only(self):
-        # overwrite because IDEFICS needs ids and embeds at the input to be not None
-        for model_class in self.all_generative_model_classes:
-            config, inputs_dict = self.prepare_config_and_inputs_for_generate()
-
-            # Ignore:
-            # a) eos (to always output 20 tokens) and pad (so we don't try to infer the attn mask from the input_ids,
-            #   which would cause a mismatch),
-            config.pad_token_id = config.eos_token_id = -1
-            config.is_decoder = True
-            model = model_class(config).to(torch_device).eval()
-            input_ids = inputs_dict.pop("input_ids")
-
-            # Traditional way of generating text
-            outputs_from_ids = model.generate(
-                input_ids, max_new_tokens=5, return_dict_in_generate=True, output_scores=True
-            )
-            self.assertEqual(outputs_from_ids.sequences.shape, (input_ids.shape[0], input_ids.shape[1] + 5))
-
-            # Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output)
-            inputs_embeds = model.get_input_embeddings()(input_ids)
-            outputs_from_embeds = model.generate(
-                input_ids,
-                inputs_embeds=inputs_embeds,
-                max_new_tokens=5,
-                return_dict_in_generate=True,
-                output_scores=True,
-            )
-            self.assertListEqual(outputs_from_ids.sequences.tolist(), outputs_from_embeds.sequences.tolist())
-
-            # But if we pass different inputs_embeds, we should get different outputs (the output text may be the
-            # same, but the logits will almost surely be different)
-            random_embeds = torch.rand_like(inputs_embeds)
-            outputs_from_rand_embeds = model.generate(
-                input_ids,
-                inputs_embeds=random_embeds,
-                max_new_tokens=5,
-                return_dict_in_generate=True,
-                output_scores=True,
-            )
-            for i in range(len(outputs_from_rand_embeds.scores)):
-                self.assertFalse(torch.allclose(outputs_from_embeds.scores[i], outputs_from_rand_embeds.scores[i]))
-
     # We need to override as we need to prepare such that the image token is the last token
     def test_resize_tokens_embeddings(self):
         (original_config, inputs_dict) = self.model_tester.prepare_config_and_inputs_for_common()

tests/models/mamba2/test_modeling_mamba2.py

@@ -205,7 +205,7 @@ def test_generate_without_input_ids(self):
 
     @unittest.skip(reason="To fix, Mamba 2 cache slicing test case is an edge case")
     @parameterized.expand([(1,), (2,)])
-    def test_generate_from_inputs_embeds_decoder_only(self, num_beams):
+    def test_generate_from_inputs_embeds(self, num_beams):
         pass
 
     @unittest.skip(reason="To fix, Mamba 2 cache slicing test case is an edge case")

tests/models/moshi/test_modeling_moshi.py

@@ -655,59 +655,6 @@ def test_initialization(self):
                             msg=f"Parameter {name} of model {model_class} seems not properly initialized",
                         )
 
-    @pytest.mark.generate
-    @parameterized.expand([(1,), (2,)])
-    def test_generate_from_inputs_embeds_decoder_only(self, num_beams):
-        for model_class in self.all_generative_model_classes:
-            config, input_ids, _, inputs_dict = self._get_input_ids_and_config()
-
-            model = model_class(config).to(torch_device).eval()
-            generation_kwargs = {
-                "return_dict_in_generate": True,
-                "output_scores": True,
-                "num_beams": num_beams,
-                "do_sample": False,
-            }
-
-            # Traditional way of generating text
-            outputs_from_ids = model.generate(input_ids, max_new_tokens=5, **generation_kwargs, **inputs_dict)
-            self.assertEqual(outputs_from_ids.sequences.shape, (input_ids.shape[0], input_ids.shape[1] + 5))
-
-            # Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output)
-            inputs_embeds = model.get_input_embeddings()(input_ids)
-            outputs_from_embeds = model.generate(
-                input_ids,
-                inputs_embeds=inputs_embeds,
-                max_new_tokens=5,
-                **generation_kwargs,
-                **inputs_dict,
-            )
-
-            # But if we pass different inputs_embeds, we should get different outputs (the output text may be the
-            # same, but the logits will almost surely be different)
-            random_embeds = torch.rand_like(inputs_embeds)
-            outputs_from_rand_embeds = model.generate(
-                input_ids,
-                inputs_embeds=random_embeds,
-                max_new_tokens=5,
-                **generation_kwargs,
-                **inputs_dict,
-            )
-            for i in range(len(outputs_from_rand_embeds.scores)):
-                self.assertFalse(torch.allclose(outputs_from_embeds.scores[i], outputs_from_rand_embeds.scores[i]))
-
-            # input_ids is not a required input -- if we don't pass it, the newly generated tokens will be the same
-            outputs_from_embeds_wo_ids = model.generate(
-                inputs_embeds=inputs_embeds,
-                max_new_tokens=5,
-                **generation_kwargs,
-                **inputs_dict,
-            )
-            self.assertListEqual(
-                outputs_from_embeds.sequences[:, inputs_embeds.shape[1] :].tolist(),
-                outputs_from_embeds_wo_ids.sequences.tolist(),
-            )
-
     @unittest.skip(reason="Continuing from past key values is not straightforward as we're dealing with 3 inputs")
     def test_generate_continue_from_past_key_values(self):
         pass