MLPerf Inference v6.0 — GPT-OSS-120B on 8xB300

Quick start

export WORK_DIR=/home/ubuntu/mlperf

curl -fsSL https://raw.githubusercontent.com/LambdaLabsML/mlperf_v60_inf_automated_b300s/master/run_gpt_oss.sh \
  | bash -s -- --work-dir "$WORK_DIR"

This single command runs the full pipeline:

1. Downloads the GPT-OSS dataset and model (~1.5 hours at ~24 MB/s)
2. Clones and pins the NVIDIA MLPerf partner repo
3. Builds the Docker image
4. Runs 4 benchmarks (Server/Offline x Performance/Accuracy)
5. Runs audit tests for Server and Offline

Flags

Flag	Description
--work-dir <path>	(required) Working directory for repo, data, and models
--skip-download	Skip data/model download (use on re-runs)
--skip-benchmark	Skip all 4 benchmark runs
--skip-audit	Skip audit tests
--audit-scenarios=<list>	Comma-separated audit scenarios (default: Server,Offline)

Examples

Re-run without re-downloading data:

curl -fsSL https://raw.githubusercontent.com/LambdaLabsML/mlperf_v60_inf_automated_b300s/master/run_gpt_oss.sh \
  | bash -s -- --work-dir "$WORK_DIR" --skip-download

Run only audit tests for the Server scenario:

curl -fsSL https://raw.githubusercontent.com/LambdaLabsML/mlperf_v60_inf_automated_b300s/master/run_gpt_oss.sh \
  | bash -s -- --work-dir "$WORK_DIR" --skip-download --skip-benchmark --audit-scenarios=Server