TL;DR: llama.cpp build b11721 collapses a four-kernel CUDA sequence for Qwen4exp's gated post-processing into one fused kernel, cutting launch overhead and memory traffic for self-hosted inference.
Summary: llama.cpp release b11721 ports the Vulkan HC_POST_GATE fusion (PR #29520) to the CUDA backend (PR #30602). The change fuses the SCALE -> SIGMOID -> SCALE -> hc_post chain: hc_combine now builds hc_post weights as 2sigmoid(inject/hc), and dsv4_hc_post_f32 takes a gated template flag to compute scale_outsigmoid(scale_in*p) directly from the raw weights, so the chain runs as a single kernel instead of four.
Why it matters: Kernel-launch and bandwidth overhead in the gate/post path directly affects throughput for Qwen4exp-style models served locally on NVIDIA GPUs. Rebuild or pull b11721 and benchmark token throughput against the previous build to confirm the gain on your hardware.
Source: github_releases