jan
jan copied to clipboard
feat: TPS Report generator (tokens per second)
Problem Would be nice to have a history/analytics of tokens/sec performance across all runs in jan over time. Should be called TPS Report because everyone loves office space.
Can you elaborate?
What is the main goal here? For users to see how fast certain models are? Or to find out about something else?
We can incorporate more metrics in our latest System Monitor designs:
Goals:
- Compare models vs other models on same prompt/assistant instruction
- Compare prompts vs other prompts
- Compare assistant instructions
- Jan internal: track any performance improvement/degradation with new versions of Jan/nitro
@0xSage I presume this would be dependent on the implementation of the latest system monitoring design?
It would be a great pun if you actually called it a "TPS Report" 🙏
Moving this epic to Notion til it is confirmed: https://www.notion.so/jan-ai/System-monitor-Observability-5005bb241cdb4f94a524d20ca049b93a?pvs=4