diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 5cf57188..a0e81c94 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -558,7 +558,7 @@ class FmhaKernel:
                         pv_done_bar.arrive_and_wait()  # Wait for PV[kt-1]
                         # Rescale O: load, multiply by acc_scale, store back to TMEM.
                         # CUTLASS pattern: both copies use same tOtO_i (composition-tiled).
-                        rescale_factor = acc_scale
+                        rescale_factor = Float32(0.5)  # DEBUG: force known value
                         if const_expr(self.debug_noop_rescale):
                             rescale_factor = Float32(1.0)
                         n_slices = self.head_dim // corr_tile_size