add tritonclient to base requirements

[nv-alaiacano]

The tritonclient library seems to be required in order to construct an Ensemble. I imagine that we could work around this by having a non-Triton runtime as the default setting for Ensembles. But if TritonExecutorRuntime is going to be the default runtime, we should include the tritonclient library in systems.

Sample single-node workflow to reproduce:

import pandas as pd

from merlin.io import Dataset
from merlin.systems.dag.ensemble import Ensemble
from merlin.systems.dag.ops.workflow import TransformWorkflow
from nvtabular import Workflow
from nvtabular.ops import LambdaOp, Rename

# Make a simple temperature scale conversion workflow
c_to_f_workflow = Workflow(
    ["input_celcius"]
    >> LambdaOp(lambda c: c * 9.0 / 5.0 + 32)
    >> Rename(name="temp_fahrenheit")  # type:ignore
)
fit_data = Dataset(pd.DataFrame({"input_celcius": [0, 25, 100]}))
c_to_f_workflow.fit(fit_data)

serving_operators = c_to_f_workflow.input_schema.column_names >> TransformWorkflow(c_to_f_workflow)
ensemble = Ensemble(serving_operators, c_to_f_workflow.input_schema)

The error I get:

ModuleNotFoundError                       Traceback (most recent call last)
Cell In[18], line 6
      3 import pandas as pd
      5 from merlin.io import Dataset
----> 6 from merlin.systems.dag.ensemble import Ensemble
      7 from merlin.systems.dag.ops.workflow import TransformWorkflow
      8 from nvtabular import Workflow

File ~/github/nvidia-merlin/systems/merlin/systems/dag/__init__.py:19
      1 #
      2 # Copyright (c) 2022, NVIDIA CORPORATION.
      3 #
   (...)
     17 # alias submodules here to avoid breaking everything with moving to submodules
     18 # flake8: noqa
---> 19 from .ensemble import Ensemble
     20 from .node import Node

File ~/github/nvidia-merlin/systems/merlin/systems/dag/ensemble.py:26
     24 from merlin.core.protocols import Transformable
     25 from merlin.dag import Graph
---> 26 from merlin.systems.dag.runtimes.triton import TritonExecutorRuntime
     29 class Ensemble:
     30     """
     31     Class that represents an entire ensemble consisting of multiple models that
     32     run sequentially in tritonserver initiated by an inference request.
     33     """

File ~/github/nvidia-merlin/systems/merlin/systems/dag/runtimes/triton/__init__.py:17
      1 #
      2 # Copyright (c) 2022, NVIDIA CORPORATION.
      3 #
   (...)
     15 #
     16 # flake8: noqa
---> 17 from merlin.systems.dag.runtimes.triton.runtime import TritonExecutorRuntime  # noqa

File ~/github/nvidia-merlin/systems/merlin/systems/dag/runtimes/triton/runtime.py:22
     19 from shutil import copyfile
     20 from typing import List, Tuple
---> 22 import tritonclient.grpc.model_config_pb2 as model_config
     23 from google.protobuf import text_format
     25 from merlin.dag import postorder_iter_nodes

ModuleNotFoundError: No module named 'tritonclient'

[karlhigley]

Yeah, I think we should go the other direction and make it possible to use Ensembles without tritonclient

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/systems/review/pr-325

[nv-alaiacano]

Does it seem reasonable to add the requirement to ensure that it works for the upcoming release, then remove that requirement when we remove the dependency on it in the Ensemble code?

[karlhigley]

Sure, I guess that could work 🤷🏻