transformers v5.5

.github/workflows/test_openvino.yml

@@ -38,7 +38,7 @@ jobs:
             "*diffusion*",
             "*quantization*",
           ]
-        transformers-version: ["4.45.0", "4.57.6", "latest"]
+        transformers-version: ["4.57.6", "5.3.0", "latest"]
 
     runs-on: ubuntu-22.04
 

.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: ["4.57.6", "latest"]
         include:
           - transformers-version: "main"
             os: "ubuntu-22.04"

optimum/exporters/openvino/model_patcher.py

@@ -52,8 +52,8 @@
     ModelPatcher,
     gpt_oss_forward,
     override_arguments,
-    sdpa_mask_without_vmap,
 )
+from optimum.exporters.onnx.model_patcher import sdpa_mask_without_vmap as sdpa_mask_without_vmap_legacy
 from optimum.intel.utils.import_utils import (
     is_diffusers_version,
     is_openvino_version,
@@ -251,19 +251,34 @@ def patch_cos_sin_cached_fp32(model):
                 )
 
 
+def sdpa_mask_without_vmap(**kwargs) -> Optional[torch.Tensor]:
+    kwargs.pop("use_vmap", None)
+    if is_transformers_version("<", "5"):
+        return sdpa_mask_without_vmap_legacy(**kwargs)
+    elif is_transformers_version(">=", "5.4") and is_transformers_version("<=", "5.7"):
+        q_length = kwargs.pop("q_length", None)
+        q_offset = kwargs.pop("q_offset", 0)
+        cache_position = torch.arange(q_offset, q_offset + q_length, device=q_length.device)
+        return sdpa_mask(q_length=cache_position, use_vmap=False, **kwargs)
+    else:
+        return sdpa_mask(use_vmap=False, **kwargs)
+
+
 # Adapted from https://github.com/huggingface/transformers/blob/v4.53.0/src/transformers/masking_utils.py#L433
 # Specifically for OpenVINO, we use torch.finfo(torch.float16).min instead of torch.finfo(dtype).min
-def eager_mask_without_vmap(*args, **kwargs) -> Optional[torch.Tensor]:
+def eager_mask_without_vmap(**kwargs) -> Optional[torch.Tensor]:
     kwargs.pop("allow_is_causal_skip", None)
-    dtype = kwargs.get("dtype", torch.float32)
-    mask = sdpa_mask_without_vmap(*args, allow_is_causal_skip=False, **kwargs)
-    # we use torch.finfo(torch.float16).min instead torch.finfo(dtype).min to avoid an overflow but not
-    # sure this is the right way to handle this, we are basically pretending that -65,504 is -inf
-    mask = torch.where(
-        mask,
-        torch.tensor(0.0, device=mask.device, dtype=dtype),
-        torch.tensor(torch.finfo(torch.float16).min, device=mask.device, dtype=dtype),
-    )
+    kwargs.pop("allow_torch_fix", None)
+    dtype = kwargs.pop("dtype", torch.float32)
+    mask = sdpa_mask_without_vmap(allow_is_causal_skip=False, allow_torch_fix=False, **kwargs)
+    if mask is not None:
+        # we use torch.finfo(torch.float16).min instead torch.finfo(dtype).min to avoid an overflow but not
+        # sure this is the right way to handle this, we are basically pretending that -65,504 is -inf
+        mask = torch.where(
+            mask,
+            torch.tensor(0.0, device=mask.device, dtype=dtype),
+            torch.tensor(torch.finfo(torch.float16).min, device=mask.device, dtype=dtype),
+        )
     return mask
 
 

optimum/intel/openvino/modeling_decoder.py

@@ -668,7 +668,8 @@ def _update_model_kwargs_for_generation(
             outputs=outputs, model_kwargs=model_kwargs, is_encoder_decoder=is_encoder_decoder, **kwargs
         )
 
-        if "position_ids" in model_kwargs:
+        # _prepare_position_ids_for_generation will infer position ids since transformers v5.2
+        if "position_ids" in model_kwargs and not hasattr(self, "_prepare_position_ids_for_generation"):
             position_ids = model_kwargs["position_ids"]
             new_position_id = position_ids[..., -1:].clone()
             new_position_id += 1

optimum/intel/openvino/modeling_seq2seq.py

@@ -875,6 +875,8 @@ def forward(
 
         # Add the attention_mask inputs when needed
         if "attention_mask" in self.input_names:
+            if attention_mask is None:
+                attention_mask = torch.ones_like(inputs[self.main_input_name])
             inputs["attention_mask"] = attention_mask
 
         # Run inference

setup.py

@@ -28,8 +28,8 @@
 
 INSTALL_REQUIRE = [
     "torch>=2.1",
-    "optimum-onnx@git+https://github.com/huggingface/optimum-onnx.git@transformers-v5",
-    "transformers>=4.45,<5.1",
+    "optimum-onnx@git+https://github.com/huggingface/optimum-onnx.git@transformers-v5.5",
+    "transformers>=4.57,<5.5",
     "setuptools",
     "huggingface-hub>=0.23.2,<2.0",
     "nncf>=2.19.0",

tests/openvino/test_decoder.py

@@ -104,20 +104,26 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
 
     SUPPORTED_ARCHITECTURES += SUPPORTED_SSM_ARCHITECTURES
 
-    if is_transformers_version(">=", "4.48.0"):
+    # config loading failing coming from type mismatch coming from transformers v5.4
+    if is_transformers_version(">=", "4.48.0") and is_transformers_version("!=", "5.4"):
         SUPPORTED_ARCHITECTURES += ("cohere2",)
 
     if is_transformers_version(">=", "4.46.0"):
-        SUPPORTED_ARCHITECTURES += ("glm", "mistral-nemo", "phimoe")
+        SUPPORTED_ARCHITECTURES += ("glm", "mistral-nemo")
 
         if is_transformers_version("<", "4.54.0"):
             SUPPORTED_ARCHITECTURES += ("deepseek",)
 
+        # config loading failing coming from type mismatch coming from transformers v5.4
+        if is_transformers_version("!=", "5.4"):
+            SUPPORTED_ARCHITECTURES += ("phimoe",)
+
         # gptq and awq install disabled for windows test environment
         if platform.system() != "Windows" and is_transformers_version("<", "4.56.0"):
             SUPPORTED_ARCHITECTURES += ("opt_gptq", "mixtral_awq")
 
-    if is_transformers_version(">", "4.47"):
+    # config loading failing coming from type mismatch coming from transformers v5.4
+    if is_transformers_version(">", "4.47") and is_transformers_version("!=", "5.4"):
         SUPPORTED_ARCHITECTURES += ("olmo2",)
 
     if is_transformers_version(">=", "4.50"):

tests/openvino/test_modeling.py

@@ -97,7 +97,7 @@
     TemporaryDirectory,
 )
 from optimum.intel.pipelines import pipeline as optimum_pipeline
-from optimum.intel.utils.import_utils import _langchain_hf_available, is_transformers_version
+from optimum.intel.utils.import_utils import _langchain_hf_available, is_datasets_version, is_transformers_version
 from optimum.intel.utils.modeling_utils import _find_files_matching_pattern
 from optimum.utils import (
     DIFFUSION_MODEL_TEXT_ENCODER_2_SUBFOLDER,
@@ -931,8 +931,8 @@ def test_pipeline(self, model_arch):
     @pytest.mark.run_slow
     @slow
     @pytest.mark.skipif(
-        is_transformers_version(">=", "5.3"),
-        reason="requires transformers < v5.3 since question-answering pipeline is deprecated in v5.3",
+        is_transformers_version(">=", "5.3") or is_datasets_version("<", "4"),
+        reason="requires datasets >= 4 or transformers < v5.3 since question-answering pipeline is deprecated in v5.3",
     )
     def test_metric(self):
         model_id = "distilbert-base-cased-distilled-squad"
@@ -1113,7 +1113,6 @@ class OVModelForMaskedLMIntegrationTest(unittest.TestCase):
         "albert",
         "bert",
         "camembert",
-        "convbert",
         "deberta",
         "deberta-v2",
         "distilbert",
@@ -1131,13 +1130,16 @@ class OVModelForMaskedLMIntegrationTest(unittest.TestCase):
     )
 
     # accuracy issue, need additional investigation
-    if is_transformers_version("<", "4.51.0"):
+    if is_transformers_version("<", "4.51"):
         SUPPORTED_ARCHITECTURES += ("nystromformer",)
 
     # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
     if is_transformers_version("<", "5"):
         SUPPORTED_ARCHITECTURES += ("data2vec-text", "flaubert", "xlm")
 
+    if is_transformers_version("!=", "5.2"):
+        SUPPORTED_ARCHITECTURES += ("convbert",)
+
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_to_transformers(self, model_arch):
         model_id = MODEL_NAMES[model_arch]

tests/openvino/test_seq2seq.py

@@ -143,8 +143,6 @@ class OVModelForSeq2SeqLMIntegrationTest(OVSeq2SeqTestMixin):
         "blenderbot",
         "blenderbot-small",
         "longt5",
-        "m2m_100",
-        "mbart",
         "pegasus",
         "t5",
     )
@@ -153,20 +151,18 @@ class OVModelForSeq2SeqLMIntegrationTest(OVSeq2SeqTestMixin):
     TASK = "text2text-generation"
     GENERATION_LENGTH = 100
     SPEEDUP_CACHE = 1.1
-    UNSUPPORTED_ARCHITECTURES = set()
-    if not (is_openvino_version(">=", "2025.3.0") and is_openvino_version("<", "2026.1")) and is_transformers_version(
-        "<", "5"
-    ):
-        # There are known issues with marian model on OpenVINO 2025.3.x and 2025.4.x
-        SUPPORTED_ARCHITECTURES += ("marian",)
-    else:
-        UNSUPPORTED_ARCHITECTURES.add("marian")
-
-    # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-    if is_transformers_version("<", "5"):
-        SUPPORTED_ARCHITECTURES += ("mt5",)
-    else:
-        UNSUPPORTED_ARCHITECTURES.add("mt5")
+    _is_model_supported = {
+        # config loading failing coming from type mismatch coming from transformers v5.4
+        "m2m_100": is_transformers_version("!=", "5.4"),
+        "mbart": is_transformers_version("!=", "5.4"),
+        # known issues with marian on OpenVINO 2025.3.x and 2025.4.x
+        "marian": not (is_openvino_version(">=", "2025.3.0") and is_openvino_version("<", "2026.1"))
+        and is_transformers_version("<", "5"),
+        # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly (mt5)
+        "mt5": is_transformers_version("<", "5"),
+    }
+    SUPPORTED_ARCHITECTURES += tuple(arch for arch, supported in _is_model_supported.items() if supported)
+    UNSUPPORTED_ARCHITECTURES = {arch for arch, supported in _is_model_supported.items() if not supported}
 
     SUPPORT_STATEFUL = ("t5", "mt5", "longt5")
     if is_transformers_version(">=", "4.52.0"):
@@ -445,10 +441,17 @@ def test_pipeline(self, model_arch):
 
 
 class OVModelForVision2SeqIntegrationTest(OVSeq2SeqTestMixin):
-    SUPPORTED_ARCHITECTURES = ["vision-encoder-decoder", "trocr", "donut"]
+    SUPPORTED_ARCHITECTURES = ["vision-encoder-decoder", "trocr"]
     # GOT-OCR2 models shouldn't be exported using the task image-to-text (currently equivalent to exporting the model using image-text-to-text) and will be deprecated v1.29
     # TODO: move pix2struct tests from OVModelForPix2StructIntegrationTest
-    UNSUPPORTED_ARCHITECTURES = {"got_ocr2", "pix2struct"}
+
+    # config loading failing coming from type mismatch coming from transformers v5.4
+    _is_model_supported = {"donut": is_transformers_version("!=", "5.4")}
+    SUPPORTED_ARCHITECTURES += [arch for arch, supported in _is_model_supported.items() if supported]
+    UNSUPPORTED_ARCHITECTURES = {"got_ocr2", "pix2struct"} | {
+        arch for arch, supported in _is_model_supported.items() if not supported
+    }
+
     TASK = "image-to-text"
     OVMODEL_CLASS = OVModelForVision2Seq
     AUTOMODEL_CLASS = transformers_auto_class
@@ -568,33 +571,17 @@ class OVModelForVisualCausalLMIntegrationTest(OVSeq2SeqTestMixin):
     if is_transformers_version(">=", "4.46.0"):
         SUPPORTED_ARCHITECTURES += ["maira2"]
 
-        # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-        if is_transformers_version("<", "5"):
-            SUPPORTED_ARCHITECTURES += ["idefics3"]
-
     if is_transformers_version(">=", "4.49.0"):
         SUPPORTED_ARCHITECTURES += ["qwen2_5_vl"]
         SUPPORT_VIDEO.append("qwen2_5_vl")
 
-        # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-        if is_transformers_version("<", "5"):
-            SUPPORTED_ARCHITECTURES += ["got_ocr2"]
-
         if is_transformers_version("<", "4.54.0"):
             # remote code models differs after transformers v4.54
             SUPPORTED_ARCHITECTURES += ["phi4mm"]
             SUPPORT_AUDIO.append("phi4mm")
 
     if is_transformers_version(">=", "4.50"):
         SUPPORTED_ARCHITECTURES += ["gemma3"]
-        # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-        if is_transformers_version("<", "5"):
-            SUPPORTED_ARCHITECTURES += ["smolvlm"]
-
-    # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-    if is_transformers_version(">=", "4.51") and is_transformers_version("<", "5"):
-        # SUPPORTED_ARCHITECTURES += ["llama4", "phi4_multimodal"]
-        SUPPORTED_ARCHITECTURES += ["llama4"]
 
     if is_transformers_version("<", "4.52"):
         SUPPORTED_ARCHITECTURES += ["minicpmo"]
@@ -611,10 +598,15 @@ class OVModelForVisualCausalLMIntegrationTest(OVSeq2SeqTestMixin):
         SUPPORTED_ARCHITECTURES += ["internvl_chat", "minicpmv"]
 
     # TODO: add fix for v5 and update MAX_TRANSFORMERS_VERSION accordingly
-    if is_transformers_version("<", "5"):
-        SUPPORTED_ARCHITECTURES += ("llava_next_video",)
-    else:
-        UNSUPPORTED_ARCHITECTURES.update({"got_ocr2", "idefics3", "llama4", "llava_next_video", "smolvlm"})
+    _is_model_supported = {
+        "idefics3": is_transformers_version(">=", "4.46.0") and is_transformers_version("<", "5"),
+        "got_ocr2": is_transformers_version(">=", "4.49.0") and is_transformers_version("<", "5"),
+        "smolvlm": is_transformers_version(">=", "4.50") and is_transformers_version("<", "5"),
+        "llama4": is_transformers_version(">=", "4.51") and is_transformers_version("<", "5"),
+        "llava_next_video": is_transformers_version("<", "5"),
+    }
+    SUPPORTED_ARCHITECTURES += [arch for arch, supported in _is_model_supported.items() if supported]
+    UNSUPPORTED_ARCHITECTURES.update(arch for arch, supported in _is_model_supported.items() if not supported)
     REMOTE_CODE_MODELS = ["internvl_chat", "minicpmv", "minicpmo", "llava-qwen2", "phi3_v", "maira2", "phi4mm"]
     IMAGE = Image.open(
         requests.get(