Fix warning messages about `config.json` when the base `model_id` is local.

[elementary-particle]

It should be possible for the user to specify a local directory as the base model in the library. However, currently the library only checks for remote presence of config.json, and fails to check the actual config.json when using a local repo. This PR adds check for a local model_id and fixes the behavior.

[BenjaminBossan]

Thanks for the PR, what you describe sounds reasonable. Do you have a small example where this change would apply? Ideally, we can use that to create a unit test.

Also, could you please remove the \ for line breaks?

[elementary-particle]

Sure, a simple example is to create a LoRA adapter for a local base model and saving it. for example, create a PeftModel for a local snapshot of mistralai/Mistral-7B-v0.1, and saving the adapter issues a warning saying

warnings.warn(
    f"Could not find a config file in {model_id} - will assume that the vocabulary was not modified."
)

And the configuration is not checked therefore.

[elementary-particle]

from transformers import AutoModelForCausalLM
from peft import LoraConfig, PeftModel
from peft import prepare_model_for_kbit_training, get_peft_model

local_dir = 'path/to/model'

base_model = AutoModelForCausalLM.from_pretrained(local_dir)
peft_config = LoraConfig(
    lora_alpha=16,
    lora_dropout=0.1,
    r=64,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj","gate_proj"]
)
peft_model = get_peft_model(base_model, peft_config)
peft_model.save_pretrained("test")

[elementary-particle]

Is this sufficient? @BenjaminBossan

[BenjaminBossan]

Nice, thanks for providing the example. I used it to create a test:

class TestLocalModel:
    def test_local_model_saving_no_warning(self, recwarn, tmp_path):
        model_id = "facebook/opt-125m"
        model = AutoModelForCausalLM.from_pretrained(model_id)
        local_dir = tmp_path / model_id
        model.save_pretrained(local_dir)
        del model

        base_model = AutoModelForCausalLM.from_pretrained(local_dir)
        peft_config = LoraConfig()
        peft_model = get_peft_model(base_model, peft_config)
        peft_model.save_pretrained(local_dir)

        for warning in recwarn.list:
            assert "Could not find a config file" not in warning.message.args[0]

We could for instance put it into tests/test_hub_features.py, WDYT? Running it locally on main, it currently fails, but on your branch, it should pass.

[elementary-particle]

Certainly, I followed the syntax in testing_common.py and created a test unit for the issue. Maybe some further checks? @BenjaminBossan

[BenjaminBossan]

Certainly, I followed the syntax in testing_common.py and created a test unit for the issue.

The way you added the test, it's not executed. You would have to add corresponding methods that call this method in test_decoder_models.py, test_encoder_decoder_models.py, etc. But this is overkill, we don't really need to check this with all kind of different model architectures. Instead, as I suggested earlier, just add this test to tests/test_hub_features.py and it should be good. Let's also add a comment that explains why we need this test.

[elementary-particle]

The test case is fixed as advised and comments are added to explain the issue. Please review the changes, thanks.

[BenjaminBossan]

@elementary-particle Thanks for the update. Could you please run make style?

[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.

[BenjaminBossan]

@elementary-particle This PR is almost good to go, just a small merge conflict, could you please check it out?

[elementary-particle]

Thanks for keeping up with this PR. The merge conflict is resolved.