diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 5c5c045a..5f6fdc98 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -108,11 +108,13 @@ class FmhaKernel:
     def __call__(self, q, k, v, c, stream, lse=None):
         self.q_dtype = q.element_type; self.o_dtype = c.element_type; self.c_dtype = self.o_dtype
         # For multi-CTA, q has shape (n_h, T, hd, 1). Layout depends on inner (T, hd) dims only.
-        # Create a view of q without the head dimension for layout computation.
+        # LayoutEnum.from_tensor needs 2D/3D tensors. For 4D Q, we extract the
+        # inner dimensions by slicing away the head dim.
         if const_expr(self.num_ctas > 1):
-            q_inner = cute.local_tile(q, cute.make_layout(1, stride=1), (0,))  # select head 0's view
+            # q shape: (n_h, T, hd, 1). Slice head dim to get (T, hd, 1) view.
+            q_inner = q[0]  # select head 0 → (T, hd, 1)
             self.a_major = LayoutEnum.from_tensor(q_inner).mma_major_mode()
-            c_inner = cute.local_tile(c, cute.make_layout(1, stride=1), (0,))
+            c_inner = c[0]  # select head 0 → (T, hd, 1)
             self.c_layout = LayoutEnum.from_tensor(c_inner)
         else:
             self.a_major = LayoutEnum.from_tensor(q).mma_major_mode()