[Perf] Set split_k to 1 for triton_kernels (#30528)

Signed-off-by: Xin Yang <xyangx@amazon.com>

vllm/model_executor/layers/quantization/utils/mxfp4_utils.py

@@ -57,7 +57,13 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
                 mx_axis=1, num_warps=num_warps
             )
         )
-    if current_platform.is_cuda() and current_platform.is_device_capability(100):
+    if current_platform.is_cuda():
+        if current_platform.is_device_capability(90):
+            constraints = {
+                "split_k": 1,
+            }
+            opt_flags.update_opt_flags_constraints(constraints)
+        elif current_platform.is_device_capability(100):
             constraints = {
                 "is_persistent": True,
                 "epilogue_subtile": 1,