test : transformers v5.10

.github/workflows/test_openvino.yml

@@ -39,7 +39,7 @@ jobs:
             "*quantization*",
             "*transformation*",
           ]
-        transformers-version: ["4.45.0", "4.57.6", "latest"]
+        transformers-version: ["5.8.0", "latest"]
 
     runs-on: ubuntu-22.04
 
@@ -70,17 +70,7 @@ jobs:
         run: |
           uv pip install transformers==${{ matrix.transformers-version }}
 
-      - if: ${{ matrix.transformers-version == '4.57.6' }}
-        name: Install qwen-asr package
-        run: |
-          uv pip install qwen-asr
-
-      - if: ${{ matrix.transformers-version == '4.45.0' }}
-        name: Install specific dependencies and versions required for older transformers
-        run: |
-          uv pip install transformers==${{ matrix.transformers-version }} accelerate==0.* peft==0.13.* diffusers==0.32.* transformers_stream_generator
-
-      - if: ${{ matrix.transformers-version != '4.45.0' && matrix.test-pattern == '*decoder*'}}
+      - if: ${{ matrix.test-pattern == '*decoder*'}}
         name: Install auto-gptq, autoawq
         run: |
           uv pip install auto-gptq "autoawq<0.2.8"

.github/workflows/test_openvino_preview_models.yml

@@ -61,18 +61,8 @@ jobs:
         run: |
           uv pip install --pre -U openvino openvino-tokenizers openvino-genai --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
 
-      - name: Gemma 4 Validation
-        run: |
-          uv pip install transformers==5.5.0
-          pytest tests/openvino/${{ matrix.test-pattern }} -k "gemma4 and not gemma4_unified" --durations=0
-
       - name: Gemma 4 Unified Validation
         if: ${{ matrix.test-pattern != '*quantization*' }}
         run: |
           uv pip install "transformers>=5.10,<5.11"
-          pytest tests/openvino/${{ matrix.test-pattern }} -k gemma4_unified --durations=0
-
-      - name: Qwen3.5 and Qwen3.5-Moe Validation
-        run: |
-          uv pip install "transformers>=5.2,<5.3"
-          pytest tests/openvino/${{ matrix.test-pattern }} -k "qwen3_5 or qwen3_5_moe" --durations=0
+          pytest tests/openvino/${{ matrix.test-pattern }} -k gemma4_unified --durations=0

.github/workflows/test_openvino_slow.yml

@@ -36,7 +36,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-22.04", "windows-2022"]
-        transformers-version: ["4.45.0", "latest"]
+        transformers-version: ["5.8.0", "latest"]
         include:
           - transformers-version: "main"
             os: "ubuntu-22.04"
@@ -69,11 +69,6 @@ jobs:
           path: ~/.cache/huggingface/hub
           key: hf-models
 
-      - if: ${{ matrix.transformers-version != 'latest' && matrix.transformers-version != 'main' }}
-        name: Install specific dependencies and versions required for older transformers
-        run: |
-          uv pip install transformers==${{ matrix.transformers-version }} accelerate==0.* peft==0.13.* diffusers==0.32.* transformers_stream_generator
-
       - if: ${{ matrix.transformers-version == 'main' }}
         name: Install transformers from repository
         run: |

docs/source/openvino/inference.mdx

@@ -48,7 +48,7 @@ As shown in the table below, each task is associated with a class enabling to au
 | `OVModelForCausalLM`                 | `text-generation-with-past`          |
 | `OVModelForSeq2SeqLM`                | `text2text-generation-with-past`     |
 | `OVModelForSpeechSeq2Seq`            | `automatic-speech-recognition`       |
-| `OVModelForVision2Seq`               | `image-to-text`                      |
+| `OVModelForImageTextToText`          | `image-to-text`                      |
 | `OVModelForTextToSpeechSeq2Seq`      | `text-to-audio`                      |
 
 ### Diffusers models

docs/source/openvino/reference.mdx

@@ -102,9 +102,9 @@ The following classes are available for the following computer vision tasks.
 
 The following classes are available for the following multimodal tasks.
 
-### OVModelForVision2Seq
+### OVModelForImageTextToText
 
-[[autodoc]] openvino.modeling_seq2seq.OVModelForVision2Seq
+[[autodoc]] openvino.modeling_seq2seq.OVModelForImageTextToText
     - forward
 
 ### OVModelForPix2Struct

optimum/exporters/openvino/__main__.py

@@ -24,7 +24,7 @@
 
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
 from requests.exceptions import ConnectionError as RequestsConnectionError
-from transformers import AutoConfig, AutoTokenizer, PreTrainedTokenizerBase, ProcessorMixin
+from transformers import AutoConfig, AutoTokenizer, Mxfp4Config, PreTrainedTokenizerBase, ProcessorMixin
 from transformers.utils import is_torch_available
 
 from openvino import Core, Type, save_model
@@ -53,9 +53,6 @@
 )
 
 
-if is_transformers_version(">=", "4.55"):
-    from transformers import Mxfp4Config
-
 FORCE_ATTN_MODEL_CLASSES = {"phi3_v": "eager", "gemma2": "sdpa", "llama4": "sdpa"}
 
 if TYPE_CHECKING:
@@ -360,7 +357,7 @@ def main_export(
             loading_kwargs["config"] = update_config_for_eagle3(config)
 
         # mxfp4 quantized model will be dequantized to bf16
-        if quant_method == "mxfp4" and is_transformers_version(">=", "4.55"):
+        if quant_method == "mxfp4":
             dtype = torch.bfloat16
             loading_kwargs["quantization_config"] = Mxfp4Config(dequantize=True)
 

optimum/exporters/openvino/convert.py

@@ -656,11 +656,7 @@ def export_from_model(
         ensure_export_task_support_stateful(task) or ensure_model_type_support_stateful(model_type)
     )
 
-    if (
-        stateful
-        and is_encoder_decoder
-        and not getattr(model, "_supports_cache_class", is_transformers_version(">=", "4.54"))
-    ):
+    if stateful and is_encoder_decoder and not getattr(model, "_supports_cache_class", True):
         stateful = False
     # TODO: support ov_config.py in the model repo
     if custom_architecture and custom_export_configs is None:
@@ -1245,7 +1241,9 @@ def get_sd3_models_for_export(pipeline, exporter, int_dtype, float_dtype):
     text_encoder = getattr(pipeline, "text_encoder", None)
     if text_encoder is not None:
         text_encoder.config.output_hidden_states = True
-        text_encoder.text_model.config.output_hidden_states = True
+        # `CLIPTextTransformer` removed since transformers v5.6
+        if hasattr(text_encoder, "text_model"):
+            text_encoder.text_model.config.output_hidden_states = True
         text_encoder_config_constructor = TasksManager.get_exporter_config_constructor(
             model=text_encoder,
             exporter=exporter,
@@ -1307,7 +1305,9 @@ def get_sd3_models_for_export(pipeline, exporter, int_dtype, float_dtype):
     text_encoder_2 = getattr(pipeline, "text_encoder_2", None)
     if text_encoder_2 is not None:
         text_encoder_2.config.output_hidden_states = True
-        text_encoder_2.text_model.config.output_hidden_states = True
+        # `CLIPTextTransformer` removed since transformers v5.6
+        if hasattr(text_encoder_2, "text_model"):
+            text_encoder_2.text_model.config.output_hidden_states = True
         export_config_constructor = TasksManager.get_exporter_config_constructor(
             model=text_encoder_2,
             exporter=exporter,

optimum/exporters/openvino/input_generators.py

@@ -30,7 +30,6 @@
     FalconDummyPastKeyValuesGenerator,
     MistralDummyPastKeyValuesGenerator,
     NormalizedTextConfig,
-    is_transformers_version,
 )
 from optimum.utils.input_generators import DTYPE_MAPPER
 from optimum.utils.normalized_config import NormalizedConfig, NormalizedVisionConfig
@@ -42,46 +41,27 @@ def __init__(self, task: str, normalized_config: NormalizedTextConfig, **kwargs)
         self.multi_query = normalized_config.multi_query
 
     def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
-        if is_transformers_version("<", "4.54"):
-            if self.multi_query:
-                shape = (
-                    self.batch_size,
-                    self.sequence_length,
-                    self.hidden_size // self.num_attention_heads * 2,
-                )
-            else:
-                shape = (
-                    self.batch_size,
-                    self.num_attention_heads,
-                    self.sequence_length,
-                    self.hidden_size // self.num_attention_heads * 2,
-                )
-            pkv = [
-                self.random_float_tensor(shape, framework=framework, dtype=float_dtype) for _ in range(self.num_layers)
-            ]
-
+        if self.multi_query:
+            shape = (
+                self.batch_size,
+                1,
+                self.sequence_length,
+                self.hidden_size // self.num_attention_heads,
+            )
         else:
-            if self.multi_query:
-                shape = (
-                    self.batch_size,
-                    1,
-                    self.sequence_length,
-                    self.hidden_size // self.num_attention_heads,
-                )
-            else:
-                shape = (
-                    self.batch_size,
-                    self.num_attention_heads,
-                    self.sequence_length,
-                    self.hidden_size // self.num_attention_heads,
-                )
-            pkv = [
-                (
-                    self.random_float_tensor(shape, framework=framework, dtype=float_dtype),
-                    self.random_float_tensor(shape, framework=framework, dtype=float_dtype),
-                )
-                for _ in range(self.num_layers)
-            ]
+            shape = (
+                self.batch_size,
+                self.num_attention_heads,
+                self.sequence_length,
+                self.hidden_size // self.num_attention_heads,
+            )
+        pkv = [
+            (
+                self.random_float_tensor(shape, framework=framework, dtype=float_dtype),
+                self.random_float_tensor(shape, framework=framework, dtype=float_dtype),
+            )
+            for _ in range(self.num_layers)
+        ]
 
         return pkv