Support kunlunxin backend (#268)

### PR Category
Others

### PR Type
New Features

### Description
Based on vLLM 0.20.2 (empty), adapted for vllm-plugin-FL with support
for the Kunlunxin backend.The Qwen3.6-35B-A3B and Qwen3.6-27B models
have been verified, the test includes high-concurrency input of text and
images.
Test command：
export FLAGCX_PATH=/workspace/FlagCX
export USE_RESHAPE_AND_CACHE_FLASH=1
vllm serve /workspace/models/Qwen3.6-27B \
  --served-model-name qwen3 \
  --tensor-parallel-size 4 \
  --max-model-len 16384\
  --reasoning-parser qwen3 \
  --block-size 128 \
  --gpu-memory-utilization 0.8 \
  --enforce-eager \
  --port 8100

Verified results：
<img width="609" height="285" alt="image"
src="https://github.com/user-attachments/assets/be79815f-a7ce-45bd-a079-15f115d107b6"
/>




### Changes
Add Kunlunxin backend support.

Author: luoyc123

vllm_fl/__init__.py

@@ -94,8 +94,18 @@ def _patch_custom_ops():
     register_op_schemas()
 
 
+def _init_vendor_device():
+    """Vendor-specific device initialization patches."""
+    from vllm_fl.utils import DeviceInfo
+    if DeviceInfo().vendor_name == "kunlunxin":
+        from vllm_fl.dispatch.backends.vendor.kunlunxin.patches.patch_fla_utils import _patch_xpu_get_device
+        _patch_xpu_get_device()
+
+
 def register():
     """Register the FL platform."""
+    _init_vendor_device()
+
     _patch_custom_ops()
     _patch_flash_attn_import()
     _patch_transformers_compat()
@@ -167,7 +177,6 @@ def register_model():
     except Exception as e:
         logger.error(f"Register DeepseekV4 model error: {str(e)}")
 
-
     # Register DeepseekV4 model
     try:
         ModelRegistry.register_model(

vllm_fl/dispatch/backends/vendor/kunlunxin/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2026 Kunlunxin, Inc. All rights reserved.
+
+"""
+Kunlunxin backend for vllm-plugin-FL dispatch.
+"""
+
+from .kunlunxin import KunlunxinBackend
+
+__all__ = ["KunlunxinBackend"]

vllm_fl/dispatch/backends/vendor/kunlunxin/impl/__init__.py

@@ -0,0 +1 @@
+# Kunlunxin implementation modules

vllm_fl/dispatch/backends/vendor/kunlunxin/impl/activation.py

@@ -0,0 +1,30 @@
+# Copyright (c) 2026 Kunlunxin, Inc. All rights reserved.
+
+"""
+Kunlunxin activation operator implementations.
+"""
+
+from __future__ import annotations
+
+import torch
+
+
+def silu_and_mul_kunlunxin(obj, x: torch.Tensor) -> torch.Tensor:
+    """
+    SiLU activation followed by element-wise multiplication.
+
+    Args:
+        obj: The calling obj (for interface consistency)
+        x: Input tensor of shape [..., 2*d]
+
+    Returns:
+        Output tensor of shape [..., d]
+    """
+    import xtorch_ops
+
+    d = x.shape[-1] // 2
+    out = torch.empty(
+        *x.shape[:-1], d, dtype=x.dtype, device=x.device
+    )
+    xtorch_ops.swiglu(x, out)
+    return out