Reproduce the recorded RTX 3080 examples

These two small, single-GPU experiments produced the published RTX 3080 reports. Keep their YAML bytes unchanged when reproducing that historical setup. For new H100 experiments with your own models, use the H100 suite.

File	Purpose
report-showcase.yaml	Baseline, six valid grid settings, two deliberately invalid settings, two workloads, four repeats, parameter associations, and finalist validation when drift is observed.
baseline-control.yaml	Compare a concurrency-capable baseline with a deliberately restricted candidate; the measured winner is never forced.

The showcase's max-num-seqs: 0 cases intentionally fail server validation. Its zero drift threshold deliberately exercises finalist validation whenever the repeat halves differ. It is a demonstration policy, not a production recommendation. Timing noise and the winning configuration can differ by GPU.

Prepare the historical environment

Use Linux, one visible NVIDIA GPU, Python 3.12.3, and a driver compatible with the captured PyTorch CUDA 13.0 runtime. Local evidence uses Ubuntu WSL and RTX 3080. Other hardware produces new measurements, not the recorded results. Do not run other inference jobs alongside these measurements.

From the repository root, create a fresh virtual environment and install the captured runtime package versions and the checkout you intend to evaluate:

python3.12 -m venv /var/tmp/vllm-optimizer-demo-venv
source /var/tmp/vllm-optimizer-demo-venv/bin/activate
python -m pip install -r docs/experiments/runtime-requirements.txt
python -m pip install --no-deps -e .
python docs/experiments/prepare_model.py
python -m pip check

The model is facebook/opt-125m, revision 27dcfa74d334bc871f3234de431e71c6eeba5dd6. The preparation helper verifies model and tokenizer checksums, including the weight file. The YAML records fixed absolute model/output paths.

The two VLLM_USE_* flags capture local WSL compatibility settings. They belong to this historical recipe; the separate H100 templates use native runtime defaults.

Run unchanged

sha256sum docs/experiments/*.yaml
vllm-opt validate -c docs/experiments/report-showcase.yaml
vllm-opt validate -c docs/experiments/baseline-control.yaml
vllm-opt -c docs/experiments/report-showcase.yaml
vllm-opt -c docs/experiments/baseline-control.yaml

The showcase intentionally includes two invalid configurations. The CLI can return 0 when other trials complete, so inspect the persisted run status and trial counts. A nonzero exit needs investigation; the two planned failures alone do not imply it. Each run prints its timestamped directory under /var/tmp/vllm-optimizer-demo/runs. Copy the results off the rented machine before shutting it down.

Preserve the model, workloads, seeds, serving settings, and optimizer source revision when reproducing historical behavior. Current code ranks new runs by arithmetic mean; these archived examples used median repeat scores. Reusing their YAML with current code does not reproduce that historical scoring rule. Offline report regeneration preserves the legacy rule for the stored artifacts. New H100 scenarios can use different models and GPU allocations.

Publish the evidence

See recorded GPU results. Retain the original YAML, result.json, results.csv, self-contained report.html, normalized trial results, manifests, raw benchmark output, and logs. Record the checkout commit, YAML/model checksums, package versions, GPU model/count, driver, OS, and date. Remove credentials, GPU UUIDs, and personal paths from the public copy; retain the untouched original privately and disclose any redaction.

H100 results belong under docs/results/h100/<date>/<experiment>/<run-id>/. Add a row to its index only after actual execution; include failures rather than publishing only winning trials. Compare corresponding workloads and accepted executions rather than treating these short runs as general hardware performance rankings.