server-lifecycle

Inference server lifecycle management is automated by orchestrating environment cleanup, server launch, health verification, benchmark execution, results collection, and safe teardown. It applies deterministic workflows across backends like vLLM, SGLang, and TRT-LLM using project-specific arguments.

Consistent benchmarking across inference server runs is achieved by sourcing project-specific arguments from CLAUDE.md or AGENTS.md files. This deterministic workflow ensures repeatable experiments by applying identical configurations during server launch and benchmark execution phases.

TRT-LLM is supported alongside vLLM and SGLang for cross-backend compatibility. The lifecycle automation supports optional custom launch scripts when project complexity warrants them, deriving arguments from CLAUDE.md or AGENTS.md memory indexes for proper execution.

The best way to handle inference server teardown and health checks is using a coordinated lifecycle workflow. It verifies server health after launch, executes benchmarks, aggregates results, and performs safe teardown to ensure clean environment cleanup and prevent resource conflicts.

CLAUDE.md or AGENTS.md files are required to provide project-specific arguments for the server lifecycle automation. These memory indexes supply the deterministic configurations needed for environment setup, server launch, and benchmark execution across various inference backends.