Feature: Baidu KunLun XPU Accelerator Support

[micytao]

Summary

Add support for Baidu KunLun XPU accelerators as a new compute backend in vLLM Playground, alongside the existing NVIDIA (CUDA), AMD (ROCm), and Google TPU options.

Motivation

Baidu KunLun XPUs are increasingly used for LLM inference in China-based deployments. vLLM already supports KunLun devices through the vllm_kunlun plugin, but vLLM Playground currently has no UI or container management support for this hardware.

Scope

UI Changes

• Add "Baidu KunLun (XPU)" option to the accelerator selector dropdown
• Enforce Container mode when KunLun is selected (subprocess/remote not supported)
• XPU-specific help text and GPU device label ("XPU Device")
• XPU device monitoring via xpu_smi in the GPU Status panel

Container Management

• KunLun-specific container image (ccr-2nf6531g-pub.cnc.bj.baidubce.com/hac-aiacc/aiak-inference-llm)
• Device passthrough for all 8 XPU devices (/dev/xpu0–/dev/xpu7 + /dev/xpuctrl)
• KunLun-specific container flags: --net=host, --cap-add=SYS_PTRACE, --security-opt seccomp=unconfined, --tmpfs /dev/shm (instead of --ipc=host)
• Podman runtime support (not Docker)
• bash -lc entrypoint wrapping to ensure conda environment activation inside the container

Environment Variables

• VLLM_TARGET_DEVICE=kunlun
• VLLM_HOST_IP (auto-resolved)
• XPU-specific tuning flags (XPU_USE_MOE_SORTED_THRES, XFT_USE_FAST_SWIGLU, XMLIR_CUDNN_ENABLED, etc.)
• XPU_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES for device selection

vLLM Server Configuration

• Launch via python -m vllm.entrypoints.openai.api_server (not vllm serve)
• Auto-force dtype=float16 when dtype is auto (KunLun XPU kernels don't support bfloat16)
• Port passed directly via --port (no -p mapping due to --net=host)
• Local model path mounted at original path (not /models)
• Auto-adjust tensor_parallel_size to match selected device count

XPU Monitoring

• Parse xpu_smi output for device metrics (temperature, memory usage, utilization)
• Fallback detection of KunLun XPU in /api/gpu-capabilities

Known Limitations

• Device selection (XPU_VISIBLE_DEVICES) interacts with vLLM's model-inspection subprocess; CUDA_VISIBLE_DEVICES is used alongside to control C-level device remapping
• setup_env.sh referenced in the container's .bashrc may not exist; not fatal but logs a warning
• tensor_parallel_size=8 may hit system ulimit nproc limits

Branch

feat/kunlun-support (based on v0.1.7)

[xyDong0223]

Huge thanks to micytao and the vLLM-Kunlun team for their collaboration! As a vLLM-Kunlun developer, I'm excited to bring vLLM-playground to Baidu's Kunlun chip. Kunlun branch verification coming soon!