feat: add tiered CUDA acceleration for routed experts - #16

Merged

JustVugg merged 7 commits into

JustVugg:mainfrom

ZacharyZcR:feat/cuda-hybrid-backend

Jul 10, 2026

Commits on Jul 10, 2026

feat: add experimental CUDA backend for resident tensors
ZacharyZcR
committed
51ee4bf
feat: promote pinned experts to a bounded VRAM tier
ZacharyZcR
committed
89b0340
feat: preload the GPU expert tier at startup
ZacharyZcR
committed
a176042
fix: harden CUDA backend failure handling
ZacharyZcR
committed
f087590
feat: add deterministic multi-GPU tensor placement
ZacharyZcR
committed
e308f96
test: add deterministic CUDA benchmark fixture
ZacharyZcR
committed
5b7040f
perf: make routed experts the default CUDA path
ZacharyZcR
committed
025fbb8