Add support transformers v4.54 v4.55 (#1406)

* Add support transformers v4.54 v4.55

* fix big code

* fix gpt bigcode

* style

* fix sdpa ernie4_5

* fix mpt

* style

* fix persimmon

* fix VLMs

* decilm

* internlm2

* fix gpt2

* fix stateful seq2seq

* add decilm

* fix speecht5

* fix whisper

* bigcode

* update setup

* trigger test

* fix smolvlm

* fix vision-encoder-decoder

* fix llama4

* style

* add tests transformers version

* update tiny qwen2_5 model

* revert gpt oss

* add max transformers version

* remove tests for incompatible models

* fix deepseek  v3

* fixe sdpa grouped query attention for llama4

* upgrade ernie min transformers version

* remove ernie4_5 export

* fix llava tests triggered only when supported

* update min version

* update min transformers version for deepseek v3

* fix expected sdpa num for tests

* fix transformers version tests

* style

* fix transformers version tests

* upgrade diffusers version

* update diffusers version

* fix workflow diffusers version

* remove sdpa patch for llama4

* add fix transformers version tests

* add fix transformers version tests

* set expected int8 for transformers v4.45

* remove config

* add missing max transformers version

* filter supported model type in tests

* style

* filter supported model type in quantization tests

* update test name to models

* fix quant op

* style

* add test

* update workflows

* style

* upgrade diffusers version for v4.45

* add fixes from optimum and optimum-onnx

* update num sdpa test

* move model patcher

* fix bigcode

* add fix sam patcher

* fix beam search

* remove deepseekv3 new config

Author: echarlaix

.github/workflows/test_openvino.yml

@@ -37,7 +37,7 @@ jobs:
             "*diffusion*",
             "*quantization*",
           ]
-        transformers-version: ["4.36.0", "latest"]
+        transformers-version: ["4.45.0", "latest"]
 
     runs-on: ubuntu-22.04
 
@@ -58,7 +58,7 @@ jobs:
       - if: ${{ matrix.transformers-version != 'latest' }}
         name: Install specific dependencies and versions required for older transformers
         run: |
-          uv pip install transformers==${{ matrix.transformers-version }} accelerate==0.* peft==0.13.* diffusers==0.30.* transformers_stream_generator
+          uv pip install transformers==${{ matrix.transformers-version }} accelerate==0.* peft==0.13.* diffusers==0.32.* transformers_stream_generator
 
       - if: ${{ matrix.transformers-version == 'latest' && matrix.test-pattern == '*decoder*'}}
         name: Install auto-gptq, autoawq
@@ -77,4 +77,4 @@ jobs:
       - if: ${{ matrix.test-pattern == '*modeling*' }}
         name: Test with Pytest - Nightly OpenVINO
         run: |
-          pytest tests/openvino/test_modeling_basic.py --durations=0
+          pytest tests/openvino/test_modeling_basic.py --durations=0

.github/workflows/test_openvino_slow.yml

@@ -33,12 +33,8 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-22.04", "windows-2022"]
-        transformers-version: ["4.36.0", "latest"]
+        transformers-version: ["4.45.0", "latest"]
         include:
-          - transformers-version: "4.40.0"
-            os: "ubuntu-22.04"
-          - transformers-version: "4.45.0"
-            os: "ubuntu-22.04"
           - transformers-version: "main"
             os: "ubuntu-22.04"
 
@@ -65,7 +61,7 @@ jobs:
       - 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==0.30.* transformers_stream_generator
+          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
@@ -86,4 +82,4 @@ jobs:
 
       - name: Test with Pytest (slow)
         run: |
-          pytest tests/openvino -m "run_slow" --durations=0
+          pytest tests/openvino -m "run_slow" --durations=0

optimum/exporters/openvino/convert.py

@@ -21,6 +21,7 @@
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
 
+from packaging.version import Version
 from transformers.generation import GenerationMixin
 from transformers.models.speecht5.modeling_speecht5 import SpeechT5HifiGan
 from transformers.utils import is_tf_available, is_torch_available
@@ -187,6 +188,27 @@ def export(
     if "diffusers" in str(model.__class__) and not is_diffusers_available():
         raise ImportError("The package `diffusers` is required to export diffusion models to OpenVINO.")
 
+    min_version = getattr(config, "MIN_TRANSFORMERS_VERSION", None)
+    max_version = getattr(config, "MAX_TRANSFORMERS_VERSION", None)
+
+    if min_version is not None:
+        if isinstance(min_version, Version):
+            min_version = min_version.base_version
+        if is_transformers_version("<", min_version):
+            raise ValueError(
+                f"The current version of Transformers does not allow for the export of the model. Minimum required is "
+                f"{config.MIN_TRANSFORMERS_VERSION}, got: {_transformers_version}"
+            )
+
+    if max_version is not None:
+        if isinstance(max_version, Version):
+            max_version = max_version.base_version
+        if is_transformers_version(">=", max_version):
+            raise ValueError(
+                f"The current version of Transformers does not allow for the export of the model. Maximum required is "
+                f"{config.MAX_TRANSFORMERS_VERSION}, got: {_transformers_version}"
+            )
+
     if stateful:
         # This will be checked anyway after the model conversion, but checking it earlier will save time for a user if not suitable version is used
         stateful = ensure_stateful_is_available()
@@ -633,7 +655,11 @@ 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", False):
+    if (
+        stateful
+        and is_encoder_decoder
+        and not getattr(model, "_supports_cache_class", is_transformers_version(">=", "4.54"))
+    ):
         stateful = False
     # TODO: support onnx_config.py in the model repo
     if custom_architecture and custom_export_configs is None: