Qwen4-27B's 4-way HC shift?
I just saw a PR in vllm-xpu-kernels today, revealing a significant architectural change in Qwen4-27B, and now that page shows 404 not found? It looks like a 4-way parallel residual stream to improve information flow and representation density within a fixed 27B parameter budget.