[BUG] RuntimeError: PyTorch execute failure: Expected Tensor but got GenericList

Bug description

Running inference of transformers4rec model on triton fails with
RuntimeError: PyTorch execute failure: Expected Tensor but got GenericList.

Similar issues in the past were solved by updating the forward function (see here, for example - https://github.com/triton-inference-server/server/issues/3348). However, this solution is not straightforward for transformers4rec.

Steps/Code to reproduce bug

1. Model ensemble is generated using a traced model and merlin.systems.dag.Ensemble:

traced_model = torch.jit.trace(
            model, model_input_dict, strict=False
        )
trasform_workflow = TransformWorkflow.TransformWorkflow(workflow, max_batch_size=0)
pt_predict_workflow = PredictPyTorch.PredictPyTorch(traced_model, workflow.output_schema, model.output_schema)
pipeline = workflow.input_schema.column_names >> trasform_workflow >> pt_predict_workflow

ensemble = Ensemble(pipeline, workflow.input_schema)
ensemble.export(f"{ensemble_artifacts_repository_path}", name=model_name)

The model ensemble contains a nvtabular transformation workflow and t4rec pytorch prediction workflow, with the following structure: (list of categorical and continuous features is substantially larger than is listed here)

ensemble_artifacts_repository
├── model_name
│   ├── config.pbtxt
│   └── 1
│       ├── model.py
│       └── ensemble
│           ├── metadata.json
│           └── ensemble.pkl
├── 1_predictpytorchtriton
│   ├── config.pbtxt
│   └── 1
│       └── model.pt
└── 0_transformworkflowtriton
  ├── config.pbtxt
  └── 1
      ├── model.py
      └── workflow
          ├── metadata.json
          ├── workflow.pkl
          └── categories
              ├── unique.feat1.parquet
              └── unique.feat2.parquet

2. Triton is called with

curl -X POST   -H "Content-Type: application/json"   -d @instances.json   localhost:8000/v2/models/model_name/infer | jq -c '.[]?'

instances.json example file:

{
    "id": "42",
    "inputs": [
        {
            "name": "id",
            "shape": [
                5
            ],
            "datatype": "BYTES",
            "data": [ "ABCDE12345", "ABCDE12345", "ABCDE12345", "ABCDE12345", "ABCDE12345"]
            ]
        },
        {
            "name": "event_date",
            "shape": [
                5
            ],
            "datatype": "BYTES",
            "data": [
                "2024-02-19", "2024-02-18", "2024-02-18", "2024-03-01", "2024-03-01"
            ]
        },
        {
            "name": "feat1",
            "shape": [
                5
            ],
            "datatype": "BYTES",
            "data": [
                "cat1", "cat2", "cat3", "cat4", "cat5"
            ]
        },
        {
            "name": "feat2",
            "shape": [
                5
            ],
            "datatype": "BYTES",
            "data": [
                "cat1", "cat2", "cat3", "cat4", "cat5"
            ]
        }
    ]
}

3. Feature transformation completes successfully, the above error is raised on 1_predictpytorchtriton .

Expected behavior

Triton would return prediction output.

Environment details

• Transformers4Rec version: 23.8.00
• Platform: GCP Vertex AI
• Python version: 3.10.13
• Huggingface Transformers version: 4.27.1
• PyTorch version (GPU?): 1.13.1, cuda 12.1

Additional context

@niraj06 @evagian

Start by inspecting the generated ensemble artifacts, especially 1_predictpytorchtriton/model.pt and the PredictPyTorch traced-model setup. Compare the model inputs and outputs at the 1_predictpytorchtriton stage where Triton reports GenericList instead of a Tensor. Done means the same ensemble accepts the shown instances.json request and returns prediction output.