[BUG] when we serve topK model on Triton it only returns scores

[rnyak]

Bug description

when we serve topK model for a session-based model on Triton it only returns scores, but it also needs to serve topK ids together with the scores.

This issue is related to model signatures, the topK model output from model signature is only one, but it should be two.

Steps/Code to reproduce bug

Run this gist first. Check out the ensemble model output. you will see it has only one output noT two.

To check the Triton restponse, then do the following steps:

1. Launch triton on terminal and load the model with tritonserver --model-repository={OUTPUT_DATA_DIR}/<name of ensemble folder>/

2. prepare input data and send a request

from merlin.systems.triton import convert_df_to_triton_input
import tritonclient.grpc as grpcclient

validation_data = pd.read_parquet('/workspace/data/interactions_merged_df.parquet')
inputs = convert_df_to_triton_input(wf.input_schema, validation_data.iloc[:100])

with grpcclient.InferenceServerClient("localhost:8001") as client:
    response = client.infer('executor_model', inputs)

output = response.as_numpy('item_id-list/categorical_output')

Expected behavior

Environment details

• Merlin version:
• Platform:
• Python version:
• PyTorch version (GPU?):
• Tensorflow version (GPU?):

Additional context