[BUGFIX] Fix dp size > 1 for qwen3 vl model (#17624)

Co-authored-by: yizhang2077 <1109276519@qq.com>

python/sglang/srt/layers/linear.py

@@ -21,7 +21,10 @@ from sglang.srt.distributed import (
 from sglang.srt.distributed.device_communicators.pynccl_allocator import (
     use_symmetric_memory,
 )
-from sglang.srt.layers.dp_attention import is_allocation_symmetric
+from sglang.srt.layers.dp_attention import (
+    get_attention_tp_group,
+    is_allocation_symmetric,
+)
 from sglang.srt.layers.parameter import (
     BasevLLMParameter,
     BlockQuantScaleParameter,
@@ -1263,6 +1266,7 @@ class RowParallelLinear(LinearBase):
         tp_rank: Optional[int] = None,
         tp_size: Optional[int] = None,
         use_presharded_weights: bool = False,
+        use_dp_attention_reduce: bool = False,
     ):
         quant_config = None if _disable_hip_linear_quant else quant_config
         super().__init__(
@@ -1271,6 +1275,7 @@ class RowParallelLinear(LinearBase):
 
         self.input_is_parallel = input_is_parallel
         self.reduce_results = reduce_results
+        self.use_dp_attention_reduce = use_dp_attention_reduce
 
         # Divide the weight matrix along the last dimension.
         if tp_rank is None:
@@ -1419,7 +1424,10 @@ class RowParallelLinear(LinearBase):
             output_parallel = self.quant_method.apply(self, input_parallel, bias=bias_)
 
         if self.reduce_results and self.tp_size > 1 and not skip_all_reduce:
-            output = tensor_model_parallel_all_reduce(output_parallel)
+            if self.use_dp_attention_reduce:
+                output = get_attention_tp_group().all_reduce(output_parallel)
+            else:
+                output = tensor_model_parallel_all_reduce(output_parallel)
         else:
             output = output_parallel