feat(moe): add heterogeneous NVMe expert streaming tier - #560

Draft

davide221 wants to merge 21 commits into

mainfrom

codex/nvme-moe-stream

Commits on Jul 30, 2026

feat(moe): add heterogeneous NVMe expert streaming tier
davide221
committed
e0eb457
perf(moe): generalize streamed experts for single-device Strix
davide221
committed
2a11e85
perf(moe): reuse streamed expert host buffers
davide221
committed
49b2780
feat(kimi): add native K3 target backend
davide221
committed
ad29ab1
fix(moe): pad streamed quant tensors on device
davide221
committed
5c51598
test(moe): cover padded MXFP4 stream slots
davide221
committed
cc1146d
ci(moe): run MXFP4 stream test on GPU runners
davide221
committed
bc4eb28

Commits on Jul 31, 2026

feat(kimi): stream routed experts from split GGUF
davide221
committed
1b25c9a
feat(moe): add reusable expert GPU ownership
davide221
committed
c42a254
feat(moe): overlap SSD experts across GPU owners
davide221
committed
9796c05
fix(moe): harden NVMe streaming for production
davide221
committed
49a898c
feat(server): add explicit MoE storage policy
davide221
committed
c8ad62a
bench(kimi): add SSD deployment comparison harness
davide221
committed
9abec86

Commits on Aug 1, 2026

feat(moe): add expert-major package and profile cache
davide221
committed
e0a0b4e

Commits on Aug 2, 2026

perf(moe): fuse resident decode expert graphs
davide221
committed
3450da3

Commits on Aug 4, 2026

perf(moe): avoid UMA cache stalls and overlap cache hits
davide221
committed
2404599
fix(kimi): suppress Windows min max macros
davide221
committed
f34eba6