[AutoDeploy] move remaining model patches into export patch system

[lucaslie]

🚀 The feature, motivation and pitch

Move the remaining model patches into the ExportPatch system.

Right now we have certain patches that are not yet integrated into the export patch system, e.g.,

TensorRT-LLM/tensorrt_llm/_torch/auto_deploy/models/patches/decilm.py

Line 15 in bddf183

# TODO: figure out how this can be incorporated into the export patch system

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[lucaslie]

more context:

i haven't had time to look deeper into why this is, but i've noticed we call several patches' get_model_From_config_patched :

  File "/code/tensorrt_llm/tensorrt_llm/_torch/auto_deploy/models/hf.py", line 166, in _build_model
    model = self.automodel_from_config(model_config, trust_remote_code=True)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/code/tensorrt_llm/tensorrt_llm/_torch/auto_deploy/models/patches/phi.py", line 141, in get_model_from_config_patched
    model = _from_config_original(config, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/code/tensorrt_llm/tensorrt_llm/_torch/auto_deploy/models/patches/gptoss.py", line 89, in get_model_from_config_patched
    model = _from_config_original(config, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/code/tensorrt_llm/tensorrt_llm/_torch/auto_deploy/models/patches/deepseek.py", line 175, in get_model_from_config_patched
    model = _from_config_original(config, **kwargs)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/williamz/.local/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 456, in from_config
    return model_class._from_config(config, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

is that intended? (i'm assuming not, but maybe i'm missing something)

from @2ez4bz (original thread: https://nvidia.slack.com/archives/C08T55LHSG4/p1757029001967039)