Documentation: Hierarchy of CLI arguments

[anandhu-eng]

I was trying to run the endpoints cli in verbose mode. Since argparse enforces argument ordering, the following command works::

inference-endpoint -v -v benchmark  online  --concurrency 2  --endpoint http://localhost:1143   --model Qwen/Qwen2.5-1.5B   --dataset tests/datasets/dummy_1k.pkl   --num-samples 60     --output benchmark_results.json   --report-path benchmark_report --workers 2 --load-pattern=poisson --target-qps 2

But this one fails:

inference-endpoint benchmark  online -v -v --concurrency 2  --endpoint http://localhost:1143   --model Qwen/Qwen2.5-1.5B   --dataset tests/datasets/dummy_1k.pkl   --num-samples 60     --output benchmark_results.json   --report-path benchmark_report --workers 2 --load-pattern=poisson --target-qps 2

This happened because -v is a global option and must appear before the command and subcommand.

To make this clearer to users, would it be helpful to expose the argument hierarchy in the documentation? For example(or maybe something smaller):

inference-endpoint
│
├── Global Options
│   ├── --version
│   └── --verbose, -v
│
├── benchmark
│   │
│   ├── offline
│   │   ├── Shared Benchmark Args
│   │   │   ├── --endpoint, -e
│   │   │   ├── --model
│   │   │   ├── --dataset, -d
│   │   │   ├── --api-key
│   │   │   ├── --workers
│   │   │   ├── --duration
│   │   │   ├── --num-samples
│   │   │   ├── --streaming
│   │   │   ├── --mode
│   │   │   ├── --min-output-tokens
│   │   │   ├── --max-output-tokens
│   │   │   └── --report-dir
│   │   │
│   │   └── Auxiliary Args
│   │       ├── --output, -o
│   │       └── --timeout
│   │
│   ├── online
│   │   ├── Shared Benchmark Args  (same list as offline)
│   │   ├── Online-Specific Args
│   │   │   ├── --load-pattern
│   │   │   ├── --target-qps        (required for poisson)
│   │   │   └── --concurrency       (required for concurrency)
│   │   └── Auxiliary Args
│   │       ├── --output, -o
│   │       └── --timeout
│   │
│   └── from-config
│       ├── --config, -c   (YAML file - required)
│       └── Auxiliary Args
│           ├── --output, -o
│           └── --timeout
│
├── eval
│   ├── --dataset
│   ├── --endpoint, -e
│   ├── --api-key
│   ├── --output, -o
│   └── --judge
│
├── probe
│   ├── --endpoint, -e
│   ├── --api-key
│   ├── --model
│   ├── --requests
│   └── --prompt
│
├── info
│   (no arguments)
│
├── validate
│   └── --config, -c
│
└── init
    ├── --template   (offline|online|eval|submission)
    └── --output, -o

[arekay-nv]

Thanks @anandhu-eng for opening this issue. For now we are focusing on the yaml config file to specify arguments. Once we have the structure finalized and cover all the options users would need, we can plug them into the CLI parser. Most args should work via CLI, but it is easier to add/test new args via yaml.
Once we have that, we can add documentation like the one you have above.

[nv-alicheng]

We should consider using https://typer.tiangolo.com/ to simplify the CLI parsing / setup code.

[arekay-nv]

Addressed by @viraatc in #193 .