Problems / LLM serving, medium
Coding agents run 20-step tasks, 1 a second, on 8 self-hosted model replicas with prefix caching. An inference router spreads the calls round robin, so each step of a session lands on a different replica from the last.
Unlock it once for 15 credits; everything inside it is free (runs, changes, Submit, results, hints). Sign in with Google to start; new accounts get free credits.
Success ≥ 95% of tasks, p99 ≤ 30 s.
Like a coding problem, a solution has to pass every case, not just the one it was built for. Submit in the lab runs all 5; 1 of them are hidden until then.
1 requests a second for 45 seconds: 100% task (writes).
What each component does in this design, and what calls it. The simulator mimics each one from these settings.
| Component | Its job here | Settings |
|---|---|---|
| Load balancerLoad balancer | Traffic enters here.Calls Agent API.Layer 4 balancers route by IP and port; layer 7 balancers read HTTP and can route by path or header. | – |
| Agent APIApp service | Load balancer forwards to it.Calls Agent orchestrator.Little's law: throughput ≈ concurrency ÷ latency. 200 slots at 50 ms serve about 4,000 req/s. | 3 replicas × 500 slots · timeout 180 s |
| Agent orchestratorAgent orchestrator | Agent API calls it.Calls Inference router, MCP tool servers.An agent loops: think, call a tool, read the result, repeat until done. | 400 slots · 20 steps |
| GPU replica 1Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 2Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 3Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 4Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 5Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 6Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 7Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| GPU replica 8Self-hosted serving pool | Inference router forwards to it.GPU servers batch requests together; throughput rises with batch size until memory runs out. | 2 slots |
| Inference routerInference router (KV-aware) | Agent orchestrator calls it.Calls GPU replica 1, GPU replica 2, GPU replica 3, GPU replica 4, GPU replica 5, GPU replica 6, GPU replica 7, GPU replica 8.Routing a session back to the replica that holds its prefix can raise KV-cache hit rates from about 20% to 75%. | 100000 slots |
| MCP tool serversTool / MCP server | Agent orchestrator calls it.Tools give the model actions such as search or database queries; each call is another round trip. | 2000 slots |