vLLM-XPU roadmap - 2026 Q1

[jikunshang]

Create this issue to track vLLM-XPU roadmap, feature development plan. will try to move to vLLM main repo.

• kernel migration. [RFC]: XPU kernel migration to vllm-xpu-kernels vllm#33214
• torch accelerator API replacement. [RFC]: Replace torch.cuda API with torch.accelerator for better hardware compatiblity. vllm#30679
• (WIP)xpu graph functionality support [Feature][XPU]: XPU graph support vllm#26970
• MLA support [XPU] support MLA model on Intel GPU vllm#37143
• sparse MLA support @wuxun-zhang
• WoQ compressed tensor support on BMG (Wint4A16/Wfp8A16, gemm/moe_gemm)
• xpu CI pipeline optimization [RFC][XPU]: Enable Intel XPU CI for vLLM vllm#37305
• xpu dockerfile refine. [XPU]Replace pip in docker.xpu with uv pip vllm#31112

[jikunshang]

cc @xinyu-intel @xuechendi @hshen14 @rogerxfeng8

[hshen14]

We may also add WoQ MXFP4 in compressed tensor item. I think @thuang6 @yiliu30 can contribute to the general compressed tensor support.

For new model architecture/ops e.g., linear attention, mHC, do we have plan in upcoming Qs?