b806b15688
Upstream source: xllm/core/kernels/cuda/moe/moe_topk_softmax_kernels.cuh (855 lines) Adapted from vllm v0.7.3 / TensorRT-LLM v0.7.1. Files added from upstream: - moe_topk_softmax_kernels.cuh — CUB-based fused topk+softmax kernel - moe_topk_sigmoid_kernels.cuh — sigmoid variant (dependency) - device_utils.cuh — CUB warp utilities - corex_moe_topk_softmax.cu — pybind11 wrapper (.so entry point) - build_corex_moe_topk_softmax.sh — corex clang++ build script Files modified: - qwen3_5.py — import corex_moe_topk_softmax, dispatch in _pure_pytorch_experts - patch_ops.sh — compile .so during docker build Dispatch: if BI100_MOE_COREX_TOPK_SOFTMAX=1 and .so loaded, use fused kernel. Fallback: torch.topk + torch.softmax (existing behavior, no breakage). Replaces 2 PyTorch ops with 1 fused CUB kernel for MoE routing.