onnxruntime

c4b9b5a9 - [CUDA] Optimize INT4 QMoE offload prefill/decode and S=8 doorbell execution

[CUDA] Optimize INT4 QMoE offload prefill/decode and S=8 doorbell execution - Initialize CutlassMoeFCRunner<T, cutlass::uint4b_t, T> (64x128x64) for int4_bank_mode_ prefill with 4,096-route sub-batching. - Raise INT bank default_stream_rows to 2048 and default_wave to 16, defer wave lease release with BankLeaseGuard, and overlap run_cpu_tail_experts() with GPU waves. - Fuse parallel chunked FileExpertStore miss loading (BeginLoadExpert / LoadExpertChunk / FinishLoadExpert) into RunColdExperts's single ParallelFor fused_task while keeping FileExpertStore byte-preserving (scales_are_block_major = false on CPU). - Enable SwapHostCandidateForVictim for int4_bank_mode_. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>