VLM: change default mm-attention backend from triton_attn to fa4 (on blackwell) (#21595)

python/sglang/srt/layers/attention/vision.py

@@ -888,7 +888,9 @@ class VisionAttention(nn.Module):
         Priority: server args override > constructor arg > platform default.
 
         Platform defaults:
-        - CUDA: "triton_attn"
+        - CUDA (Hopper SM90): "fa3"
+        - CUDA (Blackwell SM100): "fa4"
+        - CUDA (other): "triton_attn"
         - Non-CUDA: "sdpa"
         """
         override_backend = get_global_server_args().mm_attention_backend
@@ -900,6 +902,8 @@ class VisionAttention(nn.Module):
             major, minor = get_device_capability()
             if major == 9:
                 backend = "fa3"
+            elif major == 10:
+                backend = "fa4"
             else:
                 backend = "triton_attn"
         elif _is_hip: