From 7d5e4cf46ea3d935b06be680fe6ef5bd5ec63583 Mon Sep 17 00:00:00 2001 From: autocommit Date: Wed, 17 Dec 2025 00:04:15 -0800 Subject: [PATCH] =?UTF-8?q?chore(config):=20=F0=9F=94=A7=20Update=20model-?= =?UTF-8?q?boss=20integration=20config=20keys=20in=20settings.py?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Lilith Autocommit --- services/imajin-video/service/src/config/settings.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/services/imajin-video/service/src/config/settings.py b/services/imajin-video/service/src/config/settings.py index f984746c..a1fb4b0c 100644 --- a/services/imajin-video/service/src/config/settings.py +++ b/services/imajin-video/service/src/config/settings.py @@ -49,11 +49,11 @@ class Settings(BaseSettings): model_boss_timeout_s: float = 60.0 # Max concurrent per-frame scoring calls to model-boss (bounds GPU fan-out). classify_concurrency: int = 4 - # Priority sent to model-boss for frame scoring. "high" (≤5) lets the coordinator - # spawn a SECOND vision slot on the idle GPU (parallel scoring); "normal" keeps it - # to one serialized slot. High preempts normal-priority interactive vision during - # peak — deliberate: video classification is bounded, high-value work. - classify_vision_priority: str = "high" + # Job class sent to model-boss for frame scoring. "batch" is polite-parallel: the + # coordinator may spawn a second vision slot on a free GPU to parallelize, but the + # work runs at low priority so it NEVER preempts interactive vision. "interactive" + # keeps the historical single-slot serialized behaviour. + classify_vision_job_class: str = "batch" # Content-aware (scene-change) keyframe sampling — clamp band is the GPU-cost lever. classify_scene_min_keyframes: int = 3