diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 2812ca3a..e9f4b700 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -191,18 +191,20 @@ class FmhaKernel:
 
         # Print mQ shape at trace time to understand mode structure
         print(f"D2 DEBUG: mQ shape={cute.shape(mQ)}, mK shape={cute.shape(mK)}")
+        # mQ has 4 modes: (n_h, T, hd, 1). Tiler covers (T, hd) = modes 1,2.
+        # Rest modes are (n_h, 1) = modes 0,3. Coordinate is (head_idx, 0).
 
         # Q: if num_ctas > 1, mQ has a head dimension. local_tile indexes into it.
         # K/V: shared (MQA), always coordinate 0.
         # For single-CTA (num_ctas=1), head_cta_idx=0 and coordinates are the same as before.
         if const_expr(self.num_ctas > 1):
-            gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(head_cta_idx,None,None))
+            gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(head_cta_idx,0))
         else:
             gQ = cute.local_tile(mQ,cute.slice_(self.qk_mma_tiler,(None,0,None)),(None,None,None))
         gK = cute.local_tile(mK,cute.slice_(self.qk_mma_tiler,(0,None,None)),(None,None,None))
         gV = cute.local_tile(mV,cute.slice_(self.pv_mma_tiler,(0,None,None)),(None,None,None))
         if const_expr(self.num_ctas > 1):
-            gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(head_cta_idx,None,None))
+            gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(head_cta_idx,0))
         else:
             gC = cute.local_tile(mC,cute.slice_(self.pv_mma_tiler,(None,None,0)),(None,None,None))
         n_kv_tiles = cute.size(gK, mode=[3])