Instantiating a model from pipeline() ignores model_kwargs parameter

[aphedges]

Environment info

• transformers version: 4.8.1
• Platform: Linux-3.10.0-1160.31.1.el7.x86_64-x86_64-with-centos-7.9.2009-Core
• Python version: 3.7.9
• PyTorch version (GPU?): 1.8.1+cu102 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help

This should be a one-line fix, so I will be submitting a PR shortly.

Information

Model I am using: gpt2 (not model-specific issue, though)

The problem arises when using:

• the official example scripts: (give details below)
• my own modified scripts: (give details below)

The tasks I am working on is:

• an official GLUE/SQUaD task: (give the name)
• my own task or dataset: (give details below)

To reproduce

Steps to reproduce the behavior:

1. Ensure the model cache is already populated with the correct model by running the following code:

from transformers import AutoModelForCausalLM

_ = AutoModelForCausalLM.from_pretrained("gpt2", cache_dir="model_cache")

2. Put the following code in test.py:

from transformers import pipeline

_ = pipeline("text-generation", model="gpt2", model_kwargs={"cache_dir": "model_cache"})

3. Run time TRANSFORMERS_OFFLINE=1 python test.py to force the cache to be hit
4. See that the following exception is returned:

Cannot find the requested files in the cached path and outgoing traffic has been disabled. To enable model look-ups and downloads online, set 'local_files_only' to False.
Traceback (most recent call last):
  File "test.py", line 3, in <module>
    _ = pipeline("text-generation", model="gpt2", model_kwargs={"cache_dir": "model_cache"})
  File "venv/lib/python3.7/site-packages/transformers/pipelines/__init__.py", line 409, in pipeline
    model, model_classes=model_classes, config=config, framework=framework, revision=revision, task=task
  File "venv/lib/python3.7/site-packages/transformers/pipelines/base.py", line 136, in infer_framework_load_model
    model = model_class.from_pretrained(model, **kwargs)
  File "venv/lib/python3.7/site-packages/transformers/utils/dummy_tf_objects.py", line 991, in from_pretrained
    requires_backends(cls, ["tf"])
  File "venv/lib/python3.7/site-packages/transformers/file_utils.py", line 612, in requires_backends
    raise ImportError("".join([BACKENDS_MAPPING[backend][1].format(name) for backend in backends]))
ImportError:
TFGPT2LMHeadModel requires the TensorFlow library but it was not found in your environment. Checkout the instructions on the
installation page: https://www.tensorflow.org/install and follow the ones that match your environment.

(I edited the stack trace to remove the parts of the path outside the virtual environment.)

Expected behavior

There should be no output because the model should be loaded from the cache without issues.

[LysandreJik]

You could do it like so:

from transformers import AutoModelForCausalLM, pipeline, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("gpt2", cache_dir="model_cache")
tokenizer = AutoTokenizer.from_pretrained("gpt2", cache_dir="model_cache")

_ = pipeline("text-generation", model=model, tokenizer=tokenizer)

[aphedges]

While that's a good temporary workaround (I'm currently using a different one), I was hoping for a longer term solution so pipeline() works as the docs say:

model_kwargs – Additional dictionary of keyword arguments passed along to the model’s from_pretrained(..., **model_kwargs) function.

model_kwags actually used to work properly, at least when the framework parameter was set, but #12025 broke it. #12449 should fix it, although it doesn't address the issue that #12025 broke this behavior without any tests failing.

[LysandreJik]

Great, thank you for bringing me up to speed! Indeed, I hadn't realized this was a regression - your PR looks great, let's discuss over there.