Triton ensemble export

[abidwael]

Exports Triton configs and scripted models as well an an ensemble config.

1. train a ludwig model.

from ludwig.api import LudwigModel
from ludwig.datasets import titanic

training_set, test_set, _ = titanic.load(split=True)
model = LudwigModel(config="./titanic.yaml")

train_stats, preprocessed_data, output_directory = model.train(training_set=training_set,
                                                               test_set=test_set,
                                                               experiment_name="simple_experiment",
                                                               model_name="simple_model",
                                                               skip_save_processed_input=True)

2. To export to models and configs to a Triton-compliant structure

export_triton(model, data_example, output_path, model_name, model_version, device, device_count)

3. Find the exported models under model_repository/

[brightsparc]

@abidwael This is looking good. I suggest you look to leverage the inference_utils which is being split out in this draft pr #2213 (perhaps you could land this first). So that you are not repeating the pre/post processing functions.

There will need to be some additional work to support images see example for model that includes text and image columns.

Would also be nice if we could have a way of providing triton InstanceGroup settings to be passed in, at a minimum cpu_count and gpu_count could be useful properties which we could then use herufistic

pre/post processing == cpu_count
predict = gpu_count if gpu_count > 0 else cpu_count

[github-actions]

Unit Test Results

       6 files  ±0         6 suites  ±0   2h 44m 41s ⏱️ - 3m 47s
2 966 tests ±0  2 914 ✔️  - 3    52 💤 +3  0 ❌ ±0 
8 898 runs  ±0  8 706 ✔️  - 9  192 💤 +9  0 ❌ ±0 

Results for commit e34d411. ± Comparison against base commit f654591.

♻️ This comment has been updated with latest results.

[brightsparc]

This is getting there, I've added some comments.

I'd like to see a bit more flexibility added around the list of paths that we return so that we can support different file types in future, and pass along the content type when uploading to s3. This could be done in another PR if necessary.

We do need this PR to update integration tests and the cli export as well as fix any remaining tests.

[brightsparc]

LGTM; just add file size.