|
|
@@ -171,9 +171,11 @@ args = [
|
|
|
# 0.88 (107.0) is the Flash lane's value and is too tight here — the full
|
|
|
# model's weights leave ~24 GiB before the engine's own ~12 GiB.
|
|
|
"--gpu-memory-utilization", "0.85",
|
|
|
- # 4 GiB/rank probe pin (16 GiB aggregate) — deliberately the lightest pool
|
|
|
- # that still serves long contexts; the DCP4/512K sibling pins 8 GiB.
|
|
|
- "--kv-cache-memory-bytes", "4294967296",
|
|
|
+ # 8 GiB/rank pin (measured 2026-09-13: the engine needs 6.58 GiB/rank for
|
|
|
+ # ONE 131072-token request on this model - 51.6 KB/token/rank, fp8_ds_mla
|
|
|
+ # KV - so the earlier 4 GiB probe pin could not hold a single full-length
|
|
|
+ # request; 8 GiB leaves headroom and matches the DCP4/512K sibling).
|
|
|
+ "--kv-cache-memory-bytes", "8589934592",
|
|
|
# PROBE: no DCP (the sibling has DCP4). Bare TP4: each rank holds the full
|
|
|
# KV sequence, 4 GiB pinned per rank is plenty for a 131K probe.
|
|
|
# No --kv-cache-dtype: let the r35 fork pick its native default for
|
