diff --git a/dsv4/kernels/router/nvfp4_fused_router_kernel.py b/dsv4/kernels/router/nvfp4_fused_router_kernel.py
index 1635c0b5..a44720f1 100644
--- a/dsv4/kernels/router/nvfp4_fused_router_kernel.py
+++ b/dsv4/kernels/router/nvfp4_fused_router_kernel.py
@@ -852,15 +852,20 @@ def run_nvfp4_fused_router(
         gsb=gsb_val,
     )
 
-    # Add e_bias (selection bias) and run top-k
-    # The kernel writes sqrt(softplus(logits)) in FP32
-    # activation_topk expects raw logits, so we pass the activated scores
-    # and tell it to skip the activation step
+    # Apply sqrt(softplus) activation in PyTorch (CuTeDSL MLIR crashes on exp+log+sqrt)
+    # softplus(x) = max(x, 0) + log(1 + exp(-|x|))
+    abs_x = activated_scores.abs()
+    pos = activated_scores.clamp(min=0.0)
+    exp_neg = torch.exp(-abs_x)
+    sp = pos + torch.log1p(exp_neg)
+    activated = torch.sqrt(sp)
+
+    # Top-k + renorm on activated scores
     from dsv4.kernels.router._activation_topk import run_fused_activation_topk_pre_activated
     out_weights = torch.empty(N, top_k, dtype=torch.float32, device=device)
     out_ids = torch.empty(N, top_k, dtype=torch.int32, device=device)
     run_fused_activation_topk_pre_activated(
-        activated_scores, e_bias, routed_scaling_factor, top_k,
+        activated, e_bias, routed_scaling_factor, top_k,
         out_weights, out_ids,
     )