[VincentG1234]

Summary

Add optional benchmark.prewarm for guidellm_trace_replay: a short concurrent GuideLLM run before the replay benchmark so vLLM kernel autotuners (CUDA graphs, torch.compile, etc.) can settle before timed trace requests.

Why

Trace replay measures inter-arrival timing faithfully, but the first requests after vLLM startup can pay one-time kernel compilation and graph-capture costs. That skews early replay metrics unless the server is warmed up separately. benchmark.warmup does not apply to the replay profile, so this adds a dedicated pre-run phase.

What changed

• auto_tune_vllm/benchmarks/config.py — benchmark.prewarm field with validation (duration, concurrency; trace replay only)
• auto_tune_vllm/benchmarks/profiles.py — derive prompt/output token mean+stdev from local trace files; render_prewarm_args() for synthetic concurrent prewarm data
• auto_tune_vllm/benchmarks/trace_replay.py — run prewarm subprocess before replay; fail the trial on prewarm error/timeout/cancellation
• auto_tune_vllm/benchmarks/providers.py — optional log_path on _launch_subprocess() for separate prewarm logs
• tests/benchmarks/test_guidellm_trace_replay.py — config validation, token-stat derivation, CLI args, failure/timeout behavior
• docs/configuration.md — Prewarm section under trace replay
• examples/study_config_trace_replay.yaml — commented prewarm example

How tested

• ruff check .
• pytest -v tests/ (108 passed)
• Manual E2E (maintainer): auto-tune-vllm optimize with examples/study_config_trace_replay.yaml and benchmark.prewarm enabled

Risks / limitations

• Prewarm adds wall time per trial (duration seconds + subprocess overhead).
• Token statistics are auto-derived from the local trace file; for hf:// datasets or missing files, falls back to benchmark.prompt_tokens / benchmark.output_tokens with a warning.

Links

• Follow-up to [FEAT] Add GuideLLM trace replay benchmark profile #38