Support transformers 4.42

[helena-intel]

OpenVINO tests pass for me locally with 4.42. Would be great if we can support this version.

[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.

[helena-intel]

For OpenVINO tests for the regular test there is only a non-related error:

Repository Not Found for url: https://huggingface.co/api/models/optimum-internal-testing/tiny-random-phi-private/tree/main?recursive=True&expand=False.

With the slow tests there is this error:

FAILED tests/openvino/test_modeling.py::OVModelForCausalLMIntegrationTest::test_beam_search_13_llama - AssertionError: False is not true : generation config : GenerationConfig {
  "do_sample": true,
  "max_new_tokens": 10,
  "min_new_tokens": 10,
  "num_beams": 4,
  "top_k": 1
}
, transformers output tensor([[    1, 20628,   338,   263,  7575,  2462,   322,   306,   626,  5520,
         20211, 18061, 13144, 29944, 20372, 29249, 18428, 29572, 31252,  1404],
        [    2,     2,     2,     2,     2,     2,     1,   910,   338,   592,
         27264, 14848,  2618, 20610, 24351, 28333, 17245, 12944, 19234, 21336]]), ov_model_stateful output tensor([[    1, 20628,   338,   263,  7575,  2462,   322,   306,   626,  5520,
         20211, 18061,  1404,  2376, 12454,  1404,  2376, 12454,  1404,  5661],
        [    2,     2,     2,     2,     2,     2,     1,   910,   338,   592,
         27264, 14848,  2618, 20610, 24351, 28333, 17245, 12944, 19234, 21336]])
== 1 failed, 190 passed, 421 deselected, 3066 warnings in 2018.31s (0:33:38) ===```

[IlyasMoutawwakil]

I think we should update the version in the CI as well

optimum-intel/.github/workflows/test_openvino.yml

Line 24 in 60532db

transformers-version: ["4.36.0", "4.41.*"]

[eaidova]

this is probably will require some changes related to generative models with disableing wrapping cache to cache class and getting dtype

    result = model.generate(
  File "/home/ea/work/my_optimum_intel/optimum_env/lib/python3.8/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/ea/work/my_optimum_intel/optimum_env/lib/python3.8/site-packages/optimum/intel/openvino/modeling_decoder.py", line 659, in generate
    result = super().generate(
  File "/home/ea/work/my_optimum_intel/optimum_env/lib/python3.8/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/ea/work/my_optimum_intel/optimum_env/lib/python3.8/site-packages/transformers/generation/utils.py", line 1744, in generate
    model_kwargs["past_key_values"] = self._get_cache(
  File "/home/ea/work/my_optimum_intel/optimum_env/lib/python3.8/site-packages/transformers/generation/utils.py", line 1434, in _get_cache
    cache_dtype = self.dtype
AttributeError: 'OVModelForCausalLM' object has no attribute 'dtype'

[helena-intel]

I think we should update the version in the CI as well

Thanks! I was under the impression the latest version would always be installed. I update the workflow file.

this is probably will require some changes related to generative models with disableing wrapping cache to cache class and getting dtype

:-( I'll make this PR a draft. The regular tests worked out of the box for me so I hoped this would be an easy update.

[eaidova]

@helena-intel how did you tested that? There is also forcing transformers version on optimum side that may affect optimum-intel package selection https://github.com/huggingface/optimum/blob/d0a84a94183222a3931adcfca7f234a7086821db/setup.py#L18

do you install transformers version after optimum-intel install?

[helena-intel]

Closing PR as per Ekaterina's comment this needs more work to be supported.