onnxruntime

635c85ab - [CUDA] Add portable QMoEOffload backend with doorbell decode, adaptive promotion, and DMA prefill

[CUDA] Add portable QMoEOffload backend with doorbell decode, adaptive promotion, and DMA prefill - Add QMoEOffload v4 planned execution backend (qmoe_offload.{h,cc}) supporting static GPU resident slabs, hybrid doorbell decode, adaptive background promotion, bounded FileExpertStore host cache, and wave-streaming prefill. - Add v4 offload unit tests in test_qmoe_offload_v4.py and documentation in docs/contrib_ops/cuda/moe_qmoe.md. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>