Fix MindSpeed NPU LoRA projections and Qwen3.8 PLE host gather - #209

Draft

addsubmuldiv wants to merge 6 commits into

modelscope:mainfrom

addsubmuldiv:qwen3_8_flash_next_npu

Commits on Sep 24, 2026

fix: restrict PLE pinned-host triton fast path to CUDA (NPU lacks UVA)
addsubmuldiv
committed
398479a
fix: LoRA dispatch and construction for MindSpeed-replaced native parallel linears
addsubmuldiv
committed
27bb52b
fix: preserve replicated NPU LoRA output across tensor parallel ranks
addsubmuldiv
committed
e5038ea
style: format replicated LoRA handling
addsubmuldiv
committed
64ab83c
docs: shorten NPU LoRA compatibility comments
addsubmuldiv
committed
478b1d6
fix: handle native row-parallel LoRA targets
addsubmuldiv
committed
700e31a