# Ornith P4 — Concurrency / Slot Geometry

Run: `/home/lab/bench/ornith/p4/p4-20260917T234638Z`

Workload: 6 simultaneous coding/reasoning requests, thinking_budget_tokens=2048, max_tokens=4096, 3 cold rounds per geometry.

| Geometry | slots | ctx/slot | requests | strict pass | batch wall med s | req med s | TTFT med ms | aggregate tok/s | req gen t/s | max busy | avg GPU % | peak VRAM MiB |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| G3-256K | 3 | 262144 | 18 | 14/18 (77.78%) | 96.56 | 72.44 | 27748.13 | 113.43 | 44.00 | 3 | 67.67 | 17066 |
| G4-192K | 4 | 196608 | 18 | 15/18 (83.33%) | 111.68 | 69.28 | 9634.85 | 95.97 | 32.45 | 4 | 67.64 | 16858 |
| G4-160K | 4 | 163840 | 18 | 16/18 (88.89%) | 100.25 | 70.22 | 9735.70 | 109.86 | 32.62 | 4 | 68.21 | 15068 |

## Geometry definitions

- `G3-256K`: 3 × 262144 = 786432 total context.
- `G4-192K`: 4 × 196608 = 786432 total context.
- `G4-160K`: 4 × 163840 = 655360 total context.

## Notes

- Every round starts from a fresh llama-server, so no prompt-cache carryover between rounds.
- All profiles see the exact same prompts for a given round.
- `aggregate tok/s` is total decoded tokens from the six requests divided by batch wall time.
- This P4 isolates llama.cpp scheduling/slot geometry; it is not an OpenCode integration benchmark.

## Baseline restored

- 3 × 256K
- KV Q4/Q4
- FlashAttention ON
- DFlash OFF
- cache-prompt ON
- hard output 4096

