diff --git a/__tmp_kernel_none_f0_c0_r0_g0.cu b/__tmp_kernel_none_f0_c0_r0_g0.cu
index 4d7cd525a32..ea10ce2335c 100644
--- a/__tmp_kernel_none_f0_c0_r0_g0.cu
+++ b/__tmp_kernel_none_f0_c0_r0_g0.cu
@@ -11344,6 +11344,7 @@ __global__ void __cluster_dims__(2, 1, 1) nvfuser_none_f0_c0_r0_g0(Tensor<__half
   }
   __syncthreads();
   if ((((nvfuser_index_t)threadIdx.y) == 2)) {
+    asm volatile("{setmaxnreg.dec.sync.aligned.u32 56; \n\t}");
     if ((Hopper::electSync(4294967295U) && b16)) {
       #pragma unroll 4
       for(nvfuser_index_t i24 = 0; i24 < i3; ++i24) {
@@ -11368,11 +11369,13 @@ __global__ void __cluster_dims__(2, 1, 1) nvfuser_none_f0_c0_r0_g0(Tensor<__half
         }
       }
     }
+    return;
   } else {
     #pragma unroll
     for(nvfuser_index_t i31 = 0; i31 < 4; ++i31) {
       mbarrier::arrive(toSmem((&T8[(i31 + 4LL)])));
     }
+    asm volatile("{setmaxnreg.inc.sync.aligned.u32 224; \n\t}");
     #pragma unroll 4
     for(nvfuser_index_t i32 = 0; i32 < i3; ++i32) {
       nvfuser_index_t i33;
diff --git a/report3.nsys-rep b/report3.nsys-rep
new file mode 100644
index 00000000000..1ef9a36451c
Binary files /dev/null and b/report3.nsys-rep differ
diff --git a/report3.sqlite b/report3.sqlite
new file mode 100644
index 00000000000..dff2c9517bc
Binary files /dev/null and b/report3.sqlite differ