diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 61869bf6..87783e49 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -408,6 +408,12 @@ class FmhaKernel:
                     # SMEM-P: Already wrote P values to SMEM in softmax loop
                     # Just need fence and barrier
                     print(f"[SMEM-P CUTLASS] P values already written to SMEM, proceeding to fence")
+                    
+                    # DEBUG: Compute offset for known coordinate to verify mapping
+                    test_coord = ((0,0), 0, (0,0), 0)
+                    test_offset = cute.crd2idx(test_coord, sP.layout)
+                    print(f"[SMEM-P DEBUG] test_coord {test_coord} -> offset {test_offset}")
+                    
                     cute.arch.fence_proxy("async.shared", space="cta")
                 
                 # Barrier for both TMEM-P and SMEM-P paths