[FEAT] Parse prompt and total token throughput from GuideLLM

[VincentG1234]

Summary

Expose GuideLLM prompt_tokens_per_second (input/prefill throughput) and tokens_per_second (prompt + output) in trial detailed_metrics, so they can be used in optimization.log_metrics, objectives, and the Optuna dashboard alongside existing output throughput metrics.

Why

Users want to track prompt/input throughput in Optuna user attrs (log_metrics) and optionally optimize or compare prefill vs decode performance. GuideLLM has emitted these metrics since the v0.4+ stats refactor; auto-tuning only parsed the original five benchmark scalars.

What changed

• auto_tune_vllm/benchmarks/providers.py — parse prompt_tokens_per_second and tokens_per_second from GuideLLM JSON (total category, same loop as existing metrics).
• auto_tune_vllm/core/config.py — add both base metrics to ObjectiveConfig.valid_metrics (objectives, expressions, log_metrics validation).
• docs/configuration.md — document new metrics and log_metrics example.
• examples/README_optimization_guide.md — metrics table.
• examples/study_config.yaml — commented log_metrics example.

How tested

• ruff check .
• pytest -v tests/ (full suite)
• Manual E2E (maintainer): auto-tune-vllm optimize with log_metrics including prompt_tokens_per_second_median

Verified parsing against guidellm/benchmarks_mp.json (GuideLLM >= 0.5.4 export).

Risks / limitations

• Trials fail if GuideLLM results omit prompt_tokens_per_second (e.g. very old JSON exports). Project already requires guidellm>=0.5.4 in pyproject.toml.
• tokens_per_second was already present in many JSON files but was not parsed until this PR.

Links

• Complements [FEAT] add log_metrics as Optuna trial user attrs #22 (log_metrics → Optuna user attrs)

[Copilot]

Pull request overview

Exposes GuideLLM prompt (prefill) throughput and combined prompt+decode throughput by parsing prompt_tokens_per_second and tokens_per_second into trial detailed_metrics, and wiring these base metrics into objective/log-metrics validation and documentation.

Changes:

• Parse prompt_tokens_per_second and tokens_per_second from GuideLLM benchmark JSON and include them in the required metrics set.
• Extend objective/log-metrics validation to accept the two new base metrics (and their percentile-suffixed identifiers).
• Document the new metrics and add a commented log_metrics example in sample configs/docs.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
auto_tune_vllm/benchmarks/providers.py	Adds parsing/required-metric handling for prompt_tokens_per_second and tokens_per_second from GuideLLM results.
auto_tune_vllm/core/config.py	Adds the new base metrics to ObjectiveConfig.valid_metrics so they can be used in objectives/expressions and log_metrics.
docs/configuration.md	Documents the new metrics and adds a log_metrics example including prompt_tokens_per_second_median.
examples/README_optimization_guide.md	Extends the metrics table to include prompt and combined throughput metrics.
examples/study_config.yaml	Adds a commented log_metrics example entry for prompt_tokens_per_second_median.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.