|
|
@@ -51,8 +51,12 @@ args = [
|
|
|
"--enable-multimodal",
|
|
|
"--moe-runner-backend", "marlin",
|
|
|
"--moe-a2a-backend", "none",
|
|
|
- # Memory / scheduling (128 GB unified per node)
|
|
|
- "--mem-fraction-static", "0.85",
|
|
|
+ # Memory / scheduling (128 GB unified per node). Config B (2026-09-27):
|
|
|
+ # 0.85 → 0.88 — MOPD's +6 GB weights/rank pushed the pool under
|
|
|
+ # context-length (306,752 < 327,680: single 307K+ requests got refused
|
|
|
+ # at admission). The +4 GB should bring the pool back above the ceiling;
|
|
|
+ # measured on RL sibling at 0.85: avail ~16 GB/node after static alloc.
|
|
|
+ "--mem-fraction-static", "0.88",
|
|
|
# Model is natively 1M; the measured pool at GMU 0.85 is 340,736 tokens
|
|
|
# (2x TP2: 83 GB weights/rank leaves ~25 GB/rank; 39/48 SWA layers keep it
|
|
|
# large in tokens). 320K fits a single long request with headroom; a
|
