Add transformers v5.10 support

.github/workflows/test_openvino_slow.yml

@@ -70,11 +70,6 @@ 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: |

optimum/exporters/openvino/convert.py

@@ -1410,7 +1410,9 @@ 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,
@@ -1472,7 +1474,9 @@ 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/model_patcher.py

@@ -4881,12 +4881,6 @@ def forward_with_precomputed_mask(*args, **kwargs):
 
         self.orig_forward = forward_with_precomputed_mask
 
-    def __enter__(self):
-        super().__enter__()
-
-    def __exit__(self, exc_type, exc_value, traceback):
-        super().__exit__(exc_type, exc_value, traceback)
-
 
 # Forward method of the language model of Gemma3n, needs to be patched to pass 'per_layer_inputs',
 # as original code fails to create per_layer_inputs without the providing of input_ids,
@@ -4969,7 +4963,7 @@ def gemma3n_language_model_forward(
     return outputs
 
 
-# Creates a dict of causal masks with bidirectional attention for vision tokens
+# Creates a dict of causal masks with bidirectional attention for vision tokens,
 # on sliding_attention layers, matching the behavior of transformers
 # create_causal_mask_mapping when use_bidirectional_attention == "vision".
 # Needs to be patched to pass proper 'sliding_mask' for prefill stage.
@@ -5021,7 +5015,9 @@ def _create_gemma4_bidirectional_mask_dict(attention_mask_2d, mm_token_type_ids,
     same_group = (query_groups.unsqueeze(2) == key_groups.unsqueeze(1)) & (key_groups.unsqueeze(1) >= 0)
     same_group = same_group.unsqueeze(1)  # [batch, 1, seq_len, total_len]
 
-    # Undo masking for same-group vision tokens in sliding mask
+    # Un-mask same-group vision tokens in both masks (bidirectional attention within an image).
+    if is_transformers_version(">=", "5.9"):
+        full_mask = full_mask.masked_fill(same_group, 0.0)
     sliding_mask = sliding_mask.masked_fill(same_group, 0.0)
 
     return {
@@ -5229,6 +5225,10 @@ def gemma4_text_attention_forward(
 ) -> tuple:
     from transformers.models.gemma4.modeling_gemma4 import apply_rotary_pos_emb as apply_rotary_pos_emb_gemma4
 
+    # since transformers >= v5.6 (PR #45788) `shared_kv_states` dict and is passed and `kv_shared_layer_index` removed
+    shared_kv_states = kwargs.pop("shared_kv_states", None)
+    legacy_shared_kv_states = hasattr(self, "kv_shared_layer_index")
+
     input_shape = hidden_states.shape[:-1]
     hidden_shape = (*input_shape, -1, self.head_dim)
 
@@ -5239,8 +5239,12 @@ def gemma4_text_attention_forward(
     query_states = apply_rotary_pos_emb_gemma4(query_states, cos, sin, unsqueeze_dim=2)
     query_states = query_states.transpose(1, 2)
 
-    if self.is_kv_shared_layer and past_key_values is not None:
-        key_states, value_states = past_key_values.shared_layers[self.kv_shared_layer_index]
+    if self.is_kv_shared_layer and (not legacy_shared_kv_states or past_key_values is not None):
+        if legacy_shared_kv_states:
+            key_states, value_states = past_key_values.shared_layers[self.kv_shared_layer_index]
+        else:
+            key_states, value_states = shared_kv_states[self.layer_type]
+
         key_states = key_states.to(query_states.device)
         value_states = value_states.to(query_states.device)
     else:
@@ -5255,18 +5259,26 @@ def gemma4_text_attention_forward(
         value_states = value_states.transpose(1, 2)
 
     if past_key_values is not None:
-        cache_kwargs = {
-            "sin": sin,
-            "cos": cos,
-            "cache_position": cache_position,
-            "sliding_window": self.sliding_window,
-        }
-        if not self.is_kv_shared_layer:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
-        if self.store_full_length_kv:
-            if not hasattr(past_key_values, "shared_layers"):
-                past_key_values.shared_layers = {}
-            past_key_values.shared_layers[self.layer_idx] = key_states, value_states
+        if legacy_shared_kv_states:
+            cache_kwargs = {
+                "sin": sin,
+                "cos": cos,
+                "cache_position": cache_position,
+                "sliding_window": self.sliding_window,
+            }
+            if not self.is_kv_shared_layer:
+                key_states, value_states = past_key_values.update(
+                    key_states, value_states, self.layer_idx, cache_kwargs
+                )
+            if self.store_full_length_kv:
+                if not hasattr(past_key_values, "shared_layers"):
+                    past_key_values.shared_layers = {}
+                past_key_values.shared_layers[self.layer_idx] = key_states, value_states
+        else:
+            if not self.is_kv_shared_layer:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if self.store_full_length_kv and shared_kv_states is not None:
+                shared_kv_states[self.layer_type] = key_states, value_states
 
     attention_interface = gemma4_eager_attention_forward_patched
 

optimum/intel/openvino/modeling_visual_language.py

@@ -3160,9 +3160,12 @@ def rot_pos_emb(self, grid_thw):
             wpos_ids = wpos_ids.flatten()
             pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
         pos_ids = torch.cat(pos_ids, dim=0)
-        max_grid_size = grid_thw[:, 1:].max()
-        rotary_pos_emb_full = self._rotary_pos_emb(max_grid_size)
-        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        if is_transformers_version(">=", "5.9"):
+            rotary_pos_emb = self._rotary_pos_emb(pos_ids)
+        else:
+            max_grid_size = grid_thw[:, 1:].max()
+            rotary_pos_emb_full = self._rotary_pos_emb(max_grid_size)
+            rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 
     def get_multimodal_embeddings(
@@ -4286,9 +4289,13 @@ def rot_pos_emb(self, grid_thw):
             )
             pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
         pos_ids = torch.cat(pos_ids, dim=0)
-        max_grid_size = grid_thw[:, 1:].max()
-        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
-        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        if is_transformers_version(">=", "5.9"):
+            rotary_pos_emb = self.rotary_pos_emb(pos_ids)
+        else:
+            max_grid_size = grid_thw[:, 1:].max()
+            rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+            rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+
         return rotary_pos_emb
 
     def get_vision_embeddings(self, pixel_values, grid_thw, **kwargs):

optimum/intel/openvino/utils.py

@@ -68,6 +68,9 @@
 TEXTUAL_INVERSION_EMBEDDING_KEYS = [
     "self.text_model.embeddings.token_embedding.weight",
     "self.model.text_model.embeddings.token_embedding.weight",
+    # `CLIPTextTransformer` was removed since transformers v5.6
+    "self.embeddings.token_embedding.weight",
+    "self.model.embeddings.token_embedding.weight",
 ]
 
 OV_TO_NP_TYPE = {

setup.py

@@ -30,7 +30,7 @@
     "torch>=2.1",
     "safetensors<0.8.0",
     "optimum~=2.3.0",
-    "transformers>=4.51,<5.6",
+    "transformers>=4.51,<5.11",
     "setuptools",
     "huggingface-hub>=0.23.2,<1.22",
     "nncf>=2.19.0",

tests/openvino/test_modeling.py

@@ -64,6 +64,7 @@
     SEED,
     TENSOR_ALIAS_TO_TYPE,
     TEST_IMAGE_URL,
+    is_model_type_transformers_compatible,
 )
 
 from optimum.intel import (
@@ -1041,7 +1042,7 @@ class OVModelForFeatureExtractionIntegrationTest(unittest.TestCase):
         "sentence-transformers-bert",
         "qwen3",
     )
-    if is_transformers_version("<", "5.4") or is_transformers_version(">=", "5.6"):
+    if is_model_type_transformers_compatible("qwen3_vl"):
         SUPPORTED_ARCHITECTURES += ("qwen3_vl_embedding",)
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES)

tests/openvino/test_quantization.py

@@ -1295,7 +1295,11 @@ class OVWeightCompressionTest(unittest.TestCase):
             {
                 "unet": {"names": ["__module.time_embedding.linear_1/aten::linear/MatMul"]},
                 "text_encoder": {
-                    "names": ["__module.text_model.encoder.layers.0.self_attn.q_proj/aten::linear/MatMul"]
+                    "names": [
+                        "__module.text_model.encoder.layers.0.self_attn.q_proj/aten::linear/MatMul"
+                        if is_transformers_version("<", "5.6")
+                        else "__module.encoder.layers.0.self_attn.q_proj/aten::linear/MatMul"
+                    ]
                 },
             },
         ),

tests/openvino/utils_tests.py

@@ -611,13 +611,13 @@ def _resolve_cached_model_paths(model_names: dict) -> dict:
     "gemma4": {
         "lm_model": 54,
         "text_embeddings_model": 1,
-        "vision_embeddings_model": 10,
+        "vision_embeddings_model": 10 if is_transformers_version("<", "5.10") else 11,
         "text_embeddings_per_layer_model": 1,
     },
     "gemma4_moe": {
         "lm_model": 48,
         "text_embeddings_model": 1,
-        "vision_embeddings_model": 10,
+        "vision_embeddings_model": 10 if is_transformers_version("<", "5.10") else 11,
         "text_embeddings_per_layer_model": 0,
     },
     "smollm3": {"model": 30},