Fix a bug where compile_config gets deleted

[remi-or]

Currently, to_dict method of GenerationConfig always deletes the compile_config attribute. This is because when we call save_pretrained on a GenerationConfig object, we don't want to save the CompileConfig object.
But this create an issue where compile_config is deleted even when we load the model. For instance, this snippet:

import torch
from transformers import AutoModelForCausalLM, CompileConfig, GenerationConfig

compile_config = CompileConfig(fullgraph=True)
generation_config = GenerationConfig(compile_config=compile_config)
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", dtype=torch.bfloat16, generation_config=generation_config)

print(f"After model loading:\n{model.generation_config = }")
model.generation_config.save_pretrained("test.json")

reloaded_generation_config = GenerationConfig.from_pretrained("test.json")
print(f"After reloading:\n{reloaded_generation_config = }")

Running on main produces:

After model loading: model.generation_config = GenerationConfig {}
After reloading from file: reloaded_generation_config = GenerationConfig {}

Running after this PR:

After model loading: model.generation_config = GenerationConfig {
  "compile_config": {
    "backend": "inductor",
    "dynamic": null,
    "fullgraph": true,
    "mode": "reduce-overhead",
    "options": null
  }
}
After reloading from file: reloaded_generation_config = GenerationConfig {}

Thus we get the desired behavior of not saving the compile_config without deleting it as soon as we load the model.

[remi-or]

The fails in tests_tokenization and tests_torch seem unrelated. test_processors fails with the same error as on main.

[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.

[ArthurZucker]

keys_to_pop is a bit better