Elastic Models CLI

CLI tool for benchmarking and testing elastic AI model inference.

Installation

pip install thestage-elastic-models-cli

Commands

Client Inference

# Test single inference requests
elastic-models-client client llm --prompt "Hello" --url <endpoint> --model <name>
elastic-models-client client diffusion --prompt "A cat" --url <endpoint> --model <name>
elastic-models-client client vlm --prompt "Describe image" --image <path> --url <endpoint> --model <name>
elastic-models-client client stt --audio <path> --url <endpoint> --model <name>

Benchmarking

# Run load tests using Locust
elastic-models-client benchmark llm --url <endpoint> --model <name>
elastic-models-client benchmark diffusion --url <endpoint> --model <name>
elastic-models-client benchmark vlm --url <endpoint> --model <name>
elastic-models-client benchmark stt --url <endpoint> --model <name>

# Options: --concurrency, --num-requests, --output-dir, --authorization

Requirements

• Python: >=3.10
• Dependencies: qlip_serve_client, locust, Pillow, requests, aiohttp
• NumPy: <2.0 (Triton compatibility)

⚠️ Important Caveats

1. Triton Dependency: NumPy must be <2.0 for Triton server compatibility
2. Authorization: Use --authorization for authenticated endpoints
3. Ready Endpoint: Server readiness checks may fail with Nginx/Salad setups
4. Metadata: Requires model metadata JSON file or auto-download from server

Quick Example

# Test single inference
elastic-models-client client llm \
  --prompt "Write a haiku about AI" \
  --url https://api.example.com/v2/models \
  --model meta-llama/Llama-3.1-8B

# Benchmark LLM with 4 concurrent users, 100 requests
elastic-models-client benchmark llm \
  --url https://api.example.com/v2/models \
  --model meta-llama/Llama-3.1-8B \
  --concurrency 4 \
  --num-requests 100 \
  --output-dir ./results

Output

• Benchmarks: CSV stats, HTML reports, JSONL logs (optional)
• Client: JSON response with inference results