[XPU] Fix the bug of LoRA logits on the XPU platform (#24081)

Signed-off-by: chzhang <[email protected]>

Author: chaojun-zhang

vllm/lora/layers.py

@@ -1151,7 +1151,7 @@ def _get_logits(
         lora_logits = lora_logits.mT
         indices_padded = self.punica_wrapper.sampler_indices_padded
 
-        if current_platform.is_tpu():
+        if current_platform.is_tpu() or current_platform.is_xpu():
             indices_padded = indices_padded[:logits.size(0)]
 
         lora_logits = (lora_logits.reshape(

vllm/lora/punica_wrapper/punica_xpu.py

@@ -225,6 +225,13 @@ def add_lora_linear(self,
             add_inputs=True,
             **kwargs)
 
+    @property
+    def sampler_indices_padded(self) -> torch.Tensor:
+        """
+        This property provides access to padded sampler indices.
+        """
+        return self._sampler_indices_padded[:]
+
     def add_lora_logits(self,
                         y: torch.Tensor,
                         x: torch.Tensor,
@@ -259,11 +266,11 @@ def add_lora_logits(self,
             buffer = torch.zeros((x.size(0), r),
                                  dtype=torch.float32,
                                  device=x.device)
-
-        bgmv_shrink(x, lora_a_stacked, buffer, self.sampler_indices, scale)
+        sampler_indices = torch.narrow(self._sampler_indices, 0, 0, x.size(0))
+        bgmv_shrink(x, lora_a_stacked, buffer, sampler_indices, scale)
         bgmv_expand(buffer,
                     lora_b_stacked,
                     y,
-                    self.sampler_indices,
+                    sampler_indices,
                     add_inputs=True)
         return y.view_as(y_org)

vllm/platforms/xpu.py

@@ -91,7 +91,7 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
             cache_config.block_size = 64
 
         # lazy import to avoid circular import
-        from vllm.config import CUDAGraphMode
+        from vllm.config import CompilationLevel, CUDAGraphMode
         compilation_config = vllm_config.compilation_config
         if compilation_config.cudagraph_mode is None or \
                 compilation_config.cudagraph_mode.max_cudagraph_mode() \
@@ -100,6 +100,9 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
                         "cudagraphs. Fallback to cudagraph_mode=NONE")
             compilation_config.cudagraph_mode = CUDAGraphMode.NONE
 
+        if vllm_config.lora_config is not None:
+            compilation_config.level = CompilationLevel.NO_COMPILATION
+
         # check and update parallel config
         parallel_config = vllm_config.parallel_config
         parallel_config.worker_cls = "vllm.v1.worker.xpu_worker.XPUWorker"