bench export¶
Render an envelope as a markdown block, CSV, or a Slack/Discord-friendly fenced snippet. Useful when an envelope needs to land in a PR comment, a spreadsheet, or a chat channel without bespoke formatting.
Synopsis¶
Only local envelope paths are accepted. Use bench fetch to download a remote envelope first.
Example: PR-comment markdown¶
Expected output (truncated):
## InferenceBench result — `llm.inference.chatbot-short`
- **Model**: `meta-llama/Llama-3.1-8B-Instruct` (revision `main`)
- **Engine**: `vllm v0.21.0`
- **Hardware**: `NVIDIA H100 80GB HBM3` x `1`
- **Signed**: `cosign-dev` — verify with `bench verify 60be8efd6d21.json`
| metric | value |
|---|---|
| joules_per_token | 0.6997 |
| ok_rate | 1 |
| ttft_p50_ms | 41.69 |
| throughput_tok_per_s | 1384 |
Flags¶
| Flag | Default | Description |
|---|---|---|
--format |
markdown |
One of markdown, csv, slack. |
--out |
stdout | Write to a file instead of stdout. |
--metric |
(all) | Repeatable; restricts the output to the listed metric keys. |
Format notes¶
markdown— header block (suite, model, engine, hardware, dataset hash, signing method) plus a sortedmetric / valuetable.csv—# key=valuecomment header rows followed by ametric,valueCSV body.slack— fenced code block (triple-backtick) with paired p50/p99 latencies, percentages forok_rate/compliance_rate, and averify:footer line.
See also¶
- bench summary — directory-level overview
- bench publish — publish the canonical envelope rather than a rendered copy