[VLMs] support passing embeds along with pixels (#38467)

* VLMs can work with embeds now

* update more models

* fix tests

* fix copies

* fixup

* fix

* style

* unskip tests

* fix copies

* fix tests

* style

* omni modality models

* qwen models had extra indentation

* fix some other tests

* fix copies

* fix test last time

* unrelated changes revert

* we can't rely only on embeds

* delete file

* de-flake mistral3

* fix qwen models

* fix style

* fix tests

* fix copies

* deflake the test

* modular reverted by fixes, fix again

* flaky test, overwritten

* fix copies

* style

tests/models/instructblipvideo/test_modeling_instructblipvideo.py

@@ -20,7 +20,6 @@ import unittest
 import numpy as np
 import pytest
 from huggingface_hub import hf_hub_download
-from parameterized import parameterized
 
 from transformers import (
     CONFIG_MAPPING,
@@ -535,12 +534,6 @@ class InstructBlipVideoForConditionalGenerationDecoderOnlyTest(
     def test_model_common_attributes(self):
         pass
 
-    @unittest.skip(
-        "InstructBLIPVideo cannot generate only from input ids, and requires pixel values in all cases to be present"
-    )
-    def test_generate_from_inputs_embeds_with_static_cache(self):
-        pass
-
     def test_forward_signature(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -669,13 +662,6 @@ class InstructBlipVideoForConditionalGenerationDecoderOnlyTest(
             # They should result in very similar logits
             torch.testing.assert_close(next_logits_wo_padding, next_logits_with_padding, rtol=1e-5, atol=1e-5)
 
-    @unittest.skip(
-        "InstructBLIPVideo cannot generate only from input ids, and requires pixel values in all cases to be present"
-    )
-    @parameterized.expand([("greedy", 1), ("beam search", 2)])
-    def test_generate_from_inputs_embeds(self, _, num_beams):
-        pass
-
     @require_torch_sdpa
     def test_sdpa_can_dispatch_composite_models(self):
         """