Run evals off the serving path
Difficulty 4 of 5
Each prompt change runs 800 eval cases. Each case needs one candidate answer and one judge score. The judge allows 9 calls/s. Chat p99 must stay under 6.5 s, and a run must end within 2.5 min. The GPU host gpu-b is free for the candidate.
Two write paths · 4 of 16 served features differ from training
Goals and limits
- Completion time
- under 150 s
- p99 latency (Chat)
- under 6.5 s
- Error rate (*)
- under 1%
- Monthly cost
- under 3,900 USD/mo
- Parts you can add
- Message broker, LLM server
- Part limit
- 3
- Hosts
- 2 × 1 vCPU · 80 GB
- Traffic
- 8 req/s
Concepts
How players did
Player results load from the server.