[BugFix][XPU] fix lora ops bgmv_expand size not match (#39989)

Signed-off-by: Ma, Liangliang <liangliang.ma@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/lora/ops/xpu_ops/lora_ops.py

@@ -27,8 +27,41 @@ def bgmv_expand(
     lora_indices_tensor: torch.Tensor,
     add_inputs: bool = True,
 ) -> None:
+    weight_out_dim = lora_b_weights.size(-2)
+    output_dim = output_tensor.size(1)
+
+    if weight_out_dim == output_dim:
         torch.ops._xpu_C.bgmv_expand(
-        output_tensor, inputs, lora_b_weights, lora_indices_tensor, add_inputs
+            output_tensor,
+            inputs,
+            lora_b_weights,
+            lora_indices_tensor,
+            add_inputs,
+        )
+    elif weight_out_dim < output_dim:
+        # LoRA weight output dim can be smaller than the output tensor
+        # (e.g. vocab_size vs padded logits). Use expand_slice to write
+        # only the matching portion, mirroring torch_ops common_len logic.
+        torch.ops._xpu_C.bgmv_expand_slice(
+            output_tensor,
+            inputs,
+            lora_b_weights,
+            lora_indices_tensor,
+            0,
+            weight_out_dim,
+            add_inputs,
+        )
+    else:
+        # Weight output dim larger than output tensor: truncate weights.
+        lora_b_weights = lora_b_weights[..., :output_dim, :].contiguous()
+        torch.ops._xpu_C.bgmv_expand_slice(
+            output_tensor,
+            inputs,
+            lora_b_weights,
+            lora_indices_tensor,
+            0,
+            output_dim,
+            add_inputs,
         )