Skip to content

About

Reproducible H100 GPU-seconds/request study for 91 pinned Command A+ vLLM runs, with held-out evaluation, coverage diagnostics, and independent verification.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

H100 Serving Estimator

Estimates H100 GPU-seconds/request from request shape and serving settings, using 91 published runs from one checkpoint, vLLM build, and 2xH100 node.

Results

Held-out partition Rows Model MdAPE Token-rate baseline Coverage
Calibration, 32K 11 7.1% 13.0% 90.9%
Test, 128K 14 6.5% 8.6% 64.3%
OOD, unseen tunedmax 12 14.6% 6.5% 41.7%

The frozen test gate passed: MdAPE was below 25%, with 24.6% improvement over the stronger baseline. Coverage missed nominal 80% on test and OOD; the token-rate baseline won OOD.

Held-out predictions and residuals

Evidence

  • Source: SamMausberg/capp-vllm-bench, results/results.csv, commit 7e195dbc1192092e4765c166953e1bdcd533cb72
  • Git blob: f7add3069a6c3d6fcdadce00bb2907f61566a9c0; mirror SHA-256: 282675fdea50480ea1fa6106caa8b29e523ea90735c7d4b37558d5c2410c364b
  • Corpus: CohereLabs/command-a-plus-05-2026-w4a4, patched vLLM 0.26.0, tensor parallelism 2, two NVLink H100 80GB GPUs, random-token prompts
  • Groups: (config_id, input_len); none cross fit, calibration, test, or OOD
  • Target: 2 * total tokens per request / aggregate token throughput

The 91-row ledger includes every miss. Features exclude duration, throughput, latency, memory, and KV utilization; artifacts contain no currency. servemodel.* remains the complete v1 schema namespace for compatibility.

Reproduce

Python 3.11+; no runtime dependencies.

python -B run.py build --output artifacts/local_rebuild
python -B verify_independent.py artifacts/local_rebuild
python -B tools/generate_figure.py --check
python -B tools/verify_project.py

The independent verifier imports no producer code and rebuilds the split, fit, interval, 91 predictions, metrics, and hashes. Re-signed tamper tests reject leakage and GPU-hour mutation semantically.

Predict with python -B run.py predict or the installed h100-serving-estimator predict command. Dollar fields appear only with a positive --usd-per-h100-hour value.

See METHODS.md and the canonical report.

Scope

This workload-specific observational model is not a portability claim, fleet simulator, or provider bill. OOD is descriptive and excluded from the gate.

About

Reproducible H100 GPU-seconds/request study for 91 pinned Command A+ vLLM runs, with held-out evaluation, coverage diagnostics, and independent verification.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages