Bump transformers to 5.0.0

examples/llm_compression/onnx/tiny_llama/main.py

@@ -66,9 +66,10 @@ def main():
     ov_model = OVModelForCausalLM.from_pretrained(OUTPUT_DIR, from_onnx=True)
 
     messages = [{"role": "user", "content": "What is PyTorch?"}]
-    input_ids = tokenizer.apply_chat_template(
+    batch_feature = tokenizer.apply_chat_template(
         messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-    ).to(device=model.device)
+    )
+    input_ids = batch_feature["input_ids"]
 
     start_t = time.time()
     output = ov_model.generate(input_ids, max_new_tokens=100)

examples/llm_compression/onnx/tiny_llama/requirements.txt

@@ -1,4 +1,4 @@
-transformers==4.53.0
+transformers==5.0.0
 openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0

examples/llm_compression/onnx/tiny_llama_scale_estimation/main.py

@@ -119,9 +119,10 @@ def main():
     # Infer Model.
     ov_model = OVModelForCausalLM.from_pretrained(OUTPUT_DIR, from_onnx=True)
     messages = [{"role": "user", "content": "What is PyTorch?"}]
-    input_ids = tokenizer.apply_chat_template(
+    batch_feature = tokenizer.apply_chat_template(
         messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-    ).to(device=model.device)
+    )
+    input_ids = batch_feature["input_ids"]
 
     start_t = time.time()
     output = ov_model.generate(input_ids, max_new_tokens=100)

examples/llm_compression/onnx/tiny_llama_scale_estimation/requirements.txt

@@ -1,5 +1,5 @@
 torch==2.10.0
-transformers==4.53.0
+transformers==5.0.0
 openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0

examples/llm_compression/openvino/smollm2_360m_adaptive_codebook/main.py

@@ -151,9 +151,10 @@ def generate_answers(
 
     for question in questions:
         messages.append({"role": "user", "content": question})
-        input_ids = tokenizer.apply_chat_template(
+        batch_feature = tokenizer.apply_chat_template(
             messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-        ).to(device=model.device)
+        )
+        input_ids = batch_feature["input_ids"]
         input_len = len(input_ids[0])
 
         output = model.generate(input_ids, max_new_tokens=max_new_tokens, do_sample=False)[0]

examples/llm_compression/openvino/smollm2_360m_adaptive_codebook/requirements.txt

@@ -2,7 +2,7 @@ datasets==5.0.0
 openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0
-transformers==4.53.0
+transformers==5.0.0
 onnx==1.22.0
 torch==2.10.0
 torchvision==0.25.0

examples/llm_compression/openvino/smollm2_360m_codebook/main.py

@@ -48,9 +48,10 @@ def generate_answers(
 
     for question in questions:
         messages.append({"role": "user", "content": question})
-        input_ids = tokenizer.apply_chat_template(
+        batch_feature = tokenizer.apply_chat_template(
             messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-        ).to(device=model.device)
+        )
+        input_ids = batch_feature["input_ids"]
         input_len = len(input_ids[0])
 
         output = model.generate(input_ids, max_new_tokens=max_new_tokens, do_sample=False)[0]

examples/llm_compression/openvino/smollm2_360m_codebook/requirements.txt

@@ -1,7 +1,7 @@
 openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0
-transformers==4.53.0
+transformers==5.0.0
 onnx==1.22.0
 torch==2.10.0
 torchvision==0.25.0

examples/llm_compression/openvino/smollm2_360m_fp8/main.py

@@ -47,9 +47,10 @@ def generate_answers(questions, model, tokenizer, max_new_tokens=50):
 
     for question in questions:
         messages.append({"role": "user", "content": question})
-        input_ids = tokenizer.apply_chat_template(
+        batch_feature = tokenizer.apply_chat_template(
             messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-        ).to(device=model.device)
+        )
+        input_ids = batch_feature["input_ids"]
         input_len = len(input_ids[0])
 
         output = model.generate(input_ids, max_new_tokens=max_new_tokens, do_sample=False)[0]

examples/llm_compression/openvino/smollm2_360m_fp8/requirements.txt

@@ -2,6 +2,6 @@ datasets==5.0.0
 openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0
-transformers==4.53.0
+transformers==5.0.0
 onnx==1.22.0
 torch==2.10.0

examples/llm_compression/openvino/tiny_llama/main.py

@@ -60,9 +60,10 @@ def transform_fn(data, tokenizer):
     model = OVModelForCausalLM.from_pretrained(OUTPUT_DIR)
 
     messages = [{"role": "user", "content": "What is PyTorch?"}]
-    input_ids = tokenizer.apply_chat_template(
+    batch_feature = tokenizer.apply_chat_template(
         messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
-    ).to(device=model.device)
+    )
+    input_ids = batch_feature["input_ids"]
 
     start_t = time.time()
     output = model.generate(input_ids, max_new_tokens=100)

examples/llm_compression/openvino/tiny_llama/requirements.txt

@@ -4,4 +4,4 @@ openvino==2026.2.0
 optimum-intel[openvino]==2.0.0
 optimum==2.2.0
 torch==2.10.0
-transformers==4.53.0
+transformers==5.0.0

examples/llm_compression/openvino/tiny_llama_find_hyperparams/requirements.txt

@@ -1,8 +1,8 @@
-whowhatbench @ git+https://github.com/AlexanderDokuchaev/openvino.genai@light_req#subdirectory=tools/who_what_benchmark
+whowhatbench @ git+https://github.com/AlexanderDokuchaev/openvino.genai@a26c9b20d07209f035f6d74aeae94d6a72a132ab#subdirectory=tools/who_what_benchmark
 numpy==1.26.4
 openvino==2026.2.0
 optimum-intel==2.0.0
 optimum==2.2.0
-transformers==4.53.0
+transformers==5.0.0
 onnx==1.22.0
 torch==2.10.0

examples/llm_compression/openvino/tiny_llama_synthetic_data/requirements.txt

@@ -4,5 +4,5 @@ numpy>=1.23.5,<2
 openvino==2026.2.0
 optimum-intel==2.0.0
 optimum==2.2.0
-transformers==4.53.0
+transformers==5.0.0
 onnx==1.22.0

examples/llm_compression/torch/distillation_qat_with_lora/requirements.txt

@@ -5,5 +5,5 @@ numpy>=1.23.5,<2
 openvino==2026.2.0
 optimum-intel==2.0.0
 optimum==2.2.0
-transformers==4.53.0
-lm_eval==0.4.8
+transformers==5.0.0
+lm_eval[hf]==0.4.12

examples/llm_compression/torch/downstream_qat_with_nls/requirements.txt

@@ -4,6 +4,6 @@ numpy>=1.23.5,<2
 openvino==2026.2.0
 optimum-intel==2.0.0
 optimum==2.2.0
-transformers==4.53.0
-lm_eval==0.4.8
+transformers==5.0.0
+lm_eval[hf]==0.4.12
 torchao==0.17.0

examples/llm_compression/torch_fx/tiny_llama/main.py

@@ -69,9 +69,10 @@ def main() -> str:
             },
             {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
         ]
-        input_ids = tokenizer.apply_chat_template(
+        batch_feature = tokenizer.apply_chat_template(
             messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
         )
+        input_ids = batch_feature["input_ids"]
 
         print("Warmup...")
         output = compressed_model_hf.generate(input_ids)

examples/llm_compression/torch_fx/tiny_llama/modelling.py

@@ -17,7 +17,6 @@
 from transformers import GenerationMixin
 from transformers import PretrainedConfig
 from transformers import PreTrainedModel
-from transformers.cache_utils import StaticCacheConfig
 from transformers.integrations.executorch import TorchExportableModuleWithStaticCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.llama.configuration_llama import LlamaConfig
@@ -97,8 +96,9 @@ def convert_and_export_with_cache(model: PreTrainedModel) -> tuple[ExportedProgr
     example_cache_position = torch.arange(0, 8, dtype=torch.long)
     model_config = None
     gen_config = None
+    model.generation_config.use_cache = True
     model.generation_config.cache_implementation = "static"
-    model.generation_config.cache_config = StaticCacheConfig(batch_size=1, max_cache_len=512)
+    model.generation_config.cache_config = {"batch_size": 1, "max_cache_len": 512}
     model.generation_config.max_new_tokens = 100
     gen_config = model.generation_config
     model_config = model.config

examples/llm_compression/torch_fx/tiny_llama/requirements.txt

@@ -1,4 +1,4 @@
-transformers==4.53.0
+transformers==5.0.0
 datasets==5.0.0
 openvino==2026.2.0
 optimum==2.2.0

tests/openvino/requirements.txt

@@ -15,6 +15,6 @@ addict>=2.4.0
 timm==0.9.2
 efficientnet_pytorch==0.7.1
 datasets
-transformers==4.53.0
+transformers==5.0.0
 optimum-intel==2.0.0
 optimum==2.2.0

tests/post_training/pipelines/fx_modelling.py

@@ -16,7 +16,6 @@
 from transformers import GenerationMixin
 from transformers import PretrainedConfig
 from transformers import PreTrainedModel
-from transformers.cache_utils import StaticCacheConfig
 from transformers.integrations.executorch import TorchExportableModuleWithStaticCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
@@ -84,8 +83,9 @@ def convert_and_export_with_cache(model: PreTrainedModel):
 
     example_input_ids = torch.ones(1, 8, dtype=torch.long)
     example_cache_position = torch.arange(0, 8, dtype=torch.long)
+    model.generation_config.use_cache = True
     model.generation_config.cache_implementation = "static"
-    model.generation_config.cache_config = StaticCacheConfig(batch_size=1, max_cache_len=512)
+    model.generation_config.cache_config = {"batch_size": 1, "max_cache_len": 512}
     model.generation_config.max_new_tokens = 100
     gen_config = model.generation_config
     model_config = model.config

tests/post_training/pipelines/gpt.py

@@ -24,14 +24,37 @@
 from tests.post_training.pipelines.base import PTQTestPipeline
 
 
+# TODO(AlexandrDokuchaev): Remove this wrapper when the issue with torch.jit.trace and transformers>=5.0 is fixed
+class CausalLMTracingWrapper(torch.nn.Module):
+    """
+    Wraps a Hugging Face causal language model so it can be exported via ``torch.jit.trace``
+    (used internally by ``openvino.convert_model``).
+
+    Since ``transformers>=5.0`` a causal LM forward pass returns a ``DynamicCache`` in its output
+    and builds the attention mask with a ``torch.diff``-based packed-sequence check when no
+    attention mask is provided. Neither construct is supported by ``torch.jit.trace`` / the
+    OpenVINO PyTorch frontend.
+    """
+
+    def __init__(self, model: torch.nn.Module) -> None:
+        super().__init__()
+        self.model = model
+        self.model.config.use_cache = False
+
+    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor | None = None) -> torch.Tensor:
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        return self.model(input_ids=input_ids, attention_mask=attention_mask)
+
+
 class GPT(PTQTestPipeline):
     """Pipeline for causal language models from Hugging Face repository"""
 
     def prepare_model(self) -> None:
         if self.backend in PT_BACKENDS:
             self.model_hf = transformers.AutoModelForCausalLM.from_pretrained(self.model_id)
-            self.model = self.model_hf
-            self.model.config.torchscript = True  # Set to export by convert_model via torch.jit.trace
+            self.model = CausalLMTracingWrapper(self.model_hf)
+
             self.dummy_tensor = self.model_hf.dummy_inputs["input_ids"]
 
         elif self.backend in OV_BACKENDS + [BackendType.FP32]:

tests/post_training/requirements.txt

@@ -22,8 +22,8 @@ optimum-onnx @ git+https://github.com/AlexanderDokuchaev/optimum-onnx.git@b57739
 scikit-learn>=1.2.2,<=1.5.0
 soundfile==0.12.1
 tensorboard==2.20.0
-tensorflow-io==0.37.0
+tensorflow-io==0.37.1
 timm==0.9.2
 accelerate==1.9.0
-transformers==4.53.0
-whowhatbench @ git+https://github.com/AlexanderDokuchaev/openvino.genai@light_req#subdirectory=tools/who_what_benchmark
+transformers==5.0.0
+whowhatbench @ git+https://github.com/AlexanderDokuchaev/openvino.genai@a26c9b20d07209f035f6d74aeae94d6a72a132ab#subdirectory=tools/who_what_benchmark

tests/torch/function_hook/sparsify_activations/helpers.py

@@ -15,7 +15,6 @@
 import openvino as ov
 import torch
 import torch.nn as nn
-import transformers.models
 
 from nncf import IgnoredScope
 from nncf.experimental.torch.sparsify_activations import TargetScope
@@ -36,21 +35,6 @@ def forward(self, input_ids: torch.Tensor):
         return y0, y1
 
 
-def dummy_llama_model():
-    config = transformers.models.llama.configuration_llama.LlamaConfig(
-        vocab_size=32,
-        hidden_size=8,
-        intermediate_size=14,
-        num_attention_heads=2,
-        num_key_value_heads=1,
-        num_hidden_layers=2,
-        use_cache=False,
-        return_dict=False,
-    )
-    model = transformers.AutoModelForCausalLM.from_config(config, attn_implementation="eager")
-    return model
-
-
 def count_sparsifier_patterns_in_ov(model: ov.Model) -> int:
     """
     Counts the number of activation sparsification pattern "Abs -> LessEqual -> Select"

tests/torch/function_hook/sparsify_activations/test_algo.py

@@ -32,7 +32,6 @@
 from tests.cross_fw.shared.paths import TEST_ROOT
 from tests.torch.function_hook.sparsify_activations.helpers import ThreeLinearModel
 from tests.torch.function_hook.sparsify_activations.helpers import count_sparsifier_patterns_in_ov
-from tests.torch.function_hook.sparsify_activations.helpers import dummy_llama_model
 from tests.torch.helpers import set_torch_seed
 from tests.torch.utils import compare_with_reference_file
 from tests.torch.utils import to_comparable_nx_graph
@@ -100,24 +99,6 @@ def __str__(self) -> str:
         ref_num_batches_tracked=3,
         ref_num_patterns_in_ov=2,
     ),
-    SparsifyActivationsAlgorithmTestDesc(
-        name="dummy_llama",
-        model_getter=dummy_llama_model,
-        dataset_getter=lambda device: nncf.Dataset(torch.randint(0, 30, (3, 2, 8)).to(device)),
-        target_sparsity_by_scope={
-            TargetScope(patterns=[".*gate_proj.*"]): 0.2,
-            TargetScope(patterns=[".*up_proj.*"]): 0.3,
-            TargetScope(patterns=[".*down_proj.*"]): 0.4,
-        },
-        ignored_scope=None,
-        ref_sparsifier_target_sparsity={
-            (f"pre_hooks.model/mlp/{name}/linear/{layer_id}__0.0"): sparsity
-            for name, sparsity in [("gate_proj", 0.2), ("up_proj", 0.3), ("down_proj", 0.4)]
-            for layer_id in [0, 1]
-        },
-        ref_num_batches_tracked=3,
-        ref_num_patterns_in_ov=6,
-    ),
 ]
 
 

tests/torch/requirements.txt

@@ -14,9 +14,9 @@ torchvision
 
 addict>=2.4.0
 efficientnet_pytorch==0.7.1
-transformers==4.53.0
+transformers==5.0.0
 
-sentence-transformers==4.1.0
+sentence-transformers==5.6.0
 optimum-intel==2.0.0
 optimum==2.2.0
 accelerate==1.9.0