[XPU] Support Triton path for LoRA operations on XPU (#28511)

Signed-off-by: Fanli Lin <fanli.lin@intel.com>

vllm/lora/ops/triton_ops/lora_expand_op.py

@@ -48,6 +48,7 @@ def _lora_expand_kernel(
     SLICE_NUM: tl.constexpr,
     SAME_STRIDE: tl.constexpr,
     USE_GDC: tl.constexpr,
+    launch_pdl: tl.constexpr,
 ):
     cta_n_num = tl.cdiv(N, BLOCK_N)
     cta_m_num = tl.cdiv(M, BLOCK_M)

vllm/lora/ops/triton_ops/lora_shrink_op.py

@@ -46,6 +46,7 @@ def _lora_shrink_kernel(
     GROUP_SIZE_M: tl.constexpr,
     SLICE_NUM: tl.constexpr,
     USE_GDC: tl.constexpr,
+    launch_pdl: tl.constexpr,
 ):
     cta_n_num = tl.cdiv(N, BLOCK_N)
     cta_m_num = tl.cdiv(M, BLOCK_M)

vllm/platforms/xpu.py

@@ -101,7 +101,11 @@ class XPUPlatform(Platform):
 
     @classmethod
     def get_punica_wrapper(cls) -> str:
+        xpu_use_triton_kernel = os.getenv("XPU_USE_TRITON_KERNEL", "0") == "1"
+        if not xpu_use_triton_kernel:
             return "vllm.lora.punica_wrapper.punica_xpu.PunicaWrapperXPU"
+        else:
+            return "vllm.lora.punica_wrapper.punica_gpu.PunicaWrapperGPU"
 
     @classmethod
     def get_device_total_memory(cls, device_id: int = 0) -> int: