sync: latest measured numbers
Browse files
LADDER.md
CHANGED
|
@@ -55,3 +55,19 @@ endpoints + rank). This is the #1 next-frontier item; expected the largest struc
|
|
| 55 |
| SERVE | FINAL + ctx 131072 / max-total 262144 | 29.3 (25.2–31.8) | 22.7 (21.9–23.5) | **PRODUCTION CONFIG** — 2× context window at statistical parity; Hermes cut over to this endpoint 2026-08-28 |
|
| 56 |
|
| 57 |
Campaign complete. Closeout data in RESULTS.md; morning package in MORNING.md.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
| SERVE | FINAL + ctx 131072 / max-total 262144 | 29.3 (25.2–31.8) | 22.7 (21.9–23.5) | **PRODUCTION CONFIG** — 2× context window at statistical parity; Hermes cut over to this endpoint 2026-08-28 |
|
| 56 |
|
| 57 |
Campaign complete. Closeout data in RESULTS.md; morning package in MORNING.md.
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
## Concurrency curves (2026-08-28, agent-style prompts, 400 tok, aggregate tok/s)
|
| 61 |
+
| c | DFlash (max_req 8) | no-spec (max_req 12) |
|
| 62 |
+
|---|---|---|
|
| 63 |
+
| 1 | **37.3** | 14.5 |
|
| 64 |
+
| 2 | **49.5** | 27.1 |
|
| 65 |
+
| 4 | **51.7** (12.9/stream) | 36.8 |
|
| 66 |
+
| 8 | 47.3 | **55.1** |
|
| 67 |
+
| 12 | 48.8 | **55.0** |
|
| 68 |
+
|
| 69 |
+
**Correction:** our earlier "DFlash verify saturates at c1" was an artifact of
|
| 70 |
+
`max_running_requests=2` — with headroom, DFlash scales and dominates to ~c8.
|
| 71 |
+
~55 aggregate is the machine's bandwidth plateau either way. **Production config
|
| 72 |
+
changed to DFlash + max_running 8** (best latency at every c<8, ~94% of the fleet
|
| 73 |
+
ceiling at c12).
|