diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index f319a43f..565f66a8 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -394,7 +394,8 @@ class FmhaKernel:
                     gP_flat = cute.make_tensor(tCgP.iterator, cute.coalesce(tCgP.layout))
                     for idx in cutlass.range(cute.size(rP_flat), vectorize=True):
                         gP_flat[idx] = rP_flat[idx]
-                    cute.arch.fence_proxy("async.global", space="cta")
+                    # Ensure gP writes are visible before TMA reads
+                    cute.arch.fence_proxy("async", space="cta")
                     cute.arch.fence_proxy("async.shared", space="cta")
                 if kt > 0:
                     for i in range(n_corr_tiles):