Fix attn_implementation documentation

[fxmarty]

As reported in #26572 (comment), attn_implementation is wrongfully documented under PretrainedConfig, and is not under AutoModel.from_config & PreTrainedModel.from_pretrained as it should be.

[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.

[amyeroberts]

@fxmarty I don't think this is quite right - if someone can pass in

model = AutoModel.from_config(config, attn_implementation="foo")

they should also be able to pass it in for the config constructor and then pass that to the model

config = ModelConfig(attn_implementation="foo")
model = AutoModel.from_config(config)

[fxmarty]

@amyeroberts This is not what was suggested in #26572 (comment) by @patrickvonplaten.

Happy to change that, but I think it should be in an other PR. This PR simply reflects in the documentation the current expected usage by users:

model = AutoModel.from_config(cfg, attn_implementation="eager")
model = LlamaModel.from_pretrained("xxx", attn_implementation="eager")

while the following is illegal/not supported:

from transformers import AutoModelForCausalLM, AutoConfig, LlamaForCausalLM

cfg = AutoConfig.from_pretrained("fxmarty/tiny-llama-fast-tokenizer")
cfg.attn_implementation = "eager"

model = LlamaForCausalLM(cfg)

they should also be able to pass it in for the config constructor

Ultimately I agree with you, PreTrainedModel.__init__ should in the future have a way to obey to a specified attn_implementation (currently not exposed to users, only config._attn_implementation = "eager" works).

[amyeroberts]

@fxmarty OK, thanks for explaining and linking to the relevant comment! If it's something that causes a lot of confusion we can circle back on enabling it through the config creation

[amyeroberts]

Thanks for following up on this and making the docstrings consistent!