Transformers v5

[echarlaix]

with huggingface/optimum#2408

List of architectures for which are deprecating support for transformers < v5
(remote code modeling incompatible with v5)

• codegen2
• exaone
• decilm
• internlm2
• orion
• aquila2
• jais
• baichuan2
• internvl_chat
• minicpmv
  (remote code models)
• aquila
• xverse
• internlm
• exaone4

List of architectures for which we removed support for transformers < v5 for which support can be added later on

• xlm
• data2vec-text
• flaubert
• marian
• granitemoe
• zamba2
• phimoe
• drbx
• eagle3
• mt5
• llava
• idefics3
• gemma3
• smolvlm
• got_ocr2
• llava_next_video
• llama4
• bitnet
• qwen3_next

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[echarlaix]

redundant ("image-text-to-text" -> AutoModelForImageTextToText) so removing https://github.com/huggingface/optimum/blob/114647f51404f9747410afc287a8079f1f16389b/optimum/exporters/tasks.py#L157

[Copilot]

Pull request overview

This PR adds compatibility with transformers v5 across the OpenVINO Intel integration, updating model patchers, test suites, pipeline utilities, and model configs to handle the API changes introduced in transformers v5. It also deprecates support for architectures with remote modeling code incompatible with v5.

Changes:

• Deprecates/gates ~20 architectures (codegen2, exaone, decilm, internlm2, etc.) behind is_transformers_version("<", "5") conditions in test files and model configs
• Updates model patchers and utility code to handle transformers v5 API changes (renamed infer_framework_load_model → load_model, DynamicCache.from_legacy_cache → DynamicCache(...), return value changes, etc.)
• Updates setup.py to widen the transformers version range from <4.58 to <5.1 and pin optimum-onnx to the transformers-v5 branch

Reviewed changes

Copilot reviewed 31 out of 31 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
setup.py	Widens transformers version bound to <5.1, pins optimum-onnx to @transformers-v5 branch
optimum/intel/pipelines/accelerator_utils.py	Handles v5 API change: renamed pipeline function and changed return value
optimum/exporters/openvino/model_patcher.py	Updates cache conversion logic for v5, adds postprocess_past_key_values helper
optimum/exporters/openvino/model_configs.py	Adds MAX_TRANSFORMERS_VERSION = "4.57.6" to unsupported architectures; refactors init_model_configs
optimum/intel/openvino/modeling_base.py	Guards _get_non_default_generation_parameters call behind < "5" check
optimum/intel/openvino/modeling_seq2seq.py	Updates auto_model_class to use version-gated transformers_auto_class
optimum/intel/utils/import_utils.py	Adds is_huggingface_hub_version helper
optimum/intel/utils/modeling_utils.py	Replaces deprecated HfFolder().get_token() with get_token()
tests/openvino/test_decoder.py	Gates v5-incompatible architectures; updates comparison operators
tests/openvino/test_seq2seq.py	Gates v5-incompatible models; replaces deprecated APIs
tests/openvino/test_modeling.py	Updates image preprocessor and tokenizer API usage
tests/openvino/test_quantization.py	Updates expected node counts for v5; gates incompatible architectures
tests/openvino/test_genai.py	Removes v5-incompatible architectures from test scope
tests/openvino/test_diffusion.py	Conditionally includes stable-diffusion-3 based on version
tests/openvino/test_exporters_cli.py	Adjusts expected counts and gates architectures for v5
.github/workflows/*.yml	Adds transformers version matrix and diffusers installation steps

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.