[JJ] Add DeepSeek V4 Vision and qualified LMCache profiles - #634

Closed

voipmonitor wants to merge 10 commits into

local-inference-lab:dev/jovian-judgementfrom

voipmonitor:port/jj-ds4-vision-20260903

Commits on Sep 5, 2026

[New model][Multimodal] Add DeepSeek-V4-Flash-Vision-Exp support (#54566)

Isotr0py
authored and
voipmonitor
committed
f252e52
test: make DeepSeek V4 Vision CUDA coverage self-contained
voipmonitor
committed
a03bff0
fix: stream DeepSeek V4 Vision checkpoint loading
voipmonitor
committed
0330b86
fix(ds4-vision): enforce multimodal runtime contracts
voipmonitor
committed
a05dd65
fix(ds4): reserve vision LMCache runtime memory
voipmonitor
committed
e6cd713
fix(ds4): qualify vision LMCache 900k memory budget
voipmonitor
committed
dc468c4
fix(ds4): reserve direct LMCache runtime memory
voipmonitor
committed
026f62e
fix(ds4): qualify engine-driven LMCache memory
voipmonitor
committed
e08327f
fix(ds4): preserve automatic model length sentinel
voipmonitor
committed
6a4840b

Commits on Sep 6, 2026

fix(ds4): default model loading to InstantTensor
voipmonitor
committed
3807886