Kimi-K3 on layerwise fused export: 3 defect fixes, multimodal support, single-B200 NVFP4 - #2218

Draft

Fridah-nv wants to merge 6 commits into

fridah/layerwise-fused-exportfrom

fridah/k3-layerwise-fused-export

Commits on Aug 21, 2026

fix(export): find decoder layers through arbitrarily nested wrappers

Fridah-nv
and
claude
committed
1899b98
feat(recipes): experts-only NVFP4 layerwise fused export for offloaded models

Fridah-nv
and
claude
committed
969c0a8
fix(hf_ptq): pin offloaded weights that are read outside their own forward

Fridah-nv
and
claude
committed
0b369dd
docs(recipes): document the offloaded layerwise export recipe
Fridah-nv
committed
d4f0d50
feat(export): support multimodal models in per-layer export
Fridah-nv
committed
f9c44f7
fix(recipes): stop the offload recipe pinning resume state to /tmp
Fridah-nv
committed
9066997