Upload to HF fails for non-LLM trained

[thelinuxkid]

Describe the bug When a model is trained for categories/classification, the model weights are saved a file called model/model_weights. If the model is trained with type llm instead, the weights are saved to the directory model/model_weights with contents README.md, adapter_config.json, adapter_model.bin.

To Reproduce Steps to reproduce the behavior:

1. Train a model with name MODEL_NAME=bug-reprod-model and config

{
    "input_features": [
        {
            "name": "text",
            "type": "text",
            "encoder": {
                "trainable": true,
                "pretrained_model_name_or_path": "meta-llama/Llama-2-7b-hf",
                "adapter": {
                    "type": "lora"
                }
            }
        }
    ],

    "output_features": [
        {
            "name": "label",
            "type": "category"
        }
    ]
}

2. Attempt to uploaded the trained model for hugging face account HF_ID=bug-reprod-hf-id

ludwig upload hf_hub -r $HF_ID/$MODEL_NAME -m $MODEL_NAME/api_experiment_$MODEL_NAME

You should see an error like this

3. Manually move the weights file to a directory

   pushd $MODEL_NAME/api_experiment_$MODEL_NAME/model && \
   mv model_weights adapter_model.bin && \
   mkdir model_weights && \
   mv adapter_model.bin model_weights && \
   cp ~/save/$MODEL_NAME/{adapter_config.json,README.md} model_weights && \
   popd

   4. The upload to HF should now be successful

   ludwig upload hf_hub -r $HF_ID/$MODEL_NAME -m $MODEL_NAME/api_experiment_$MODEL_NAME

   

Expected behavior The model should upload to HF without having to manually create the directory

Environment (please complete the following information):

• OS: Distributor ID: Ubuntu Description: Ubuntu 20.04.6 LTS Release: 20.04 Codename: focal
• Version: CUDA 11.8
• Pytorch: 2.0.0+cu118
• Python: 3.8.10
• Ludwig: 0.8.1.post1

Additional context Add any other context about the problem here.

@arnavgarg1

[tgaddair]

Thanks for reporting this issue @thelinuxkid! @arnavgarg1 what do you think about modifying the upload utility to allow uploading ECD artifacts? Would this be useful?