Add NVTX range markers to track XAttention performance: - GPU-only: xattn_estimate, xattn_bsa_compute - Offload: xattn_estimate_gemm, xattn_estimate_softmax, xattn_estimate_find_blocks, xattn_compute_historical, xattn_compute_current, xattn_compute_merge These markers enable detailed nsys profiling to identify performance bottlenecks in estimate vs compute phases. Generated with [Claude Code](https://claude.ai/code) via [Happy](https://happy.engineering) Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: Happy <yesreply@happy.engineering>
32 KiB
32 KiB