D1.3: Fix LSE tensor layout for weakly congruent store

dsv4/kernels/attention/fmha.py

@@ -117,7 +117,7 @@ class FmhaKernel:
         # CuTeDSL doesn't support None parameters in @cute.kernel.
         # For normalize=True, mLSE is unused (dead-code-eliminated by compiler).
         if const_expr(lse is None):
-            lse = cute.make_tensor(c.iterator, cute.make_layout((1, 1, 1), stride=(1, 1, 1)))
+            lse = cute.make_tensor(c.iterator, cute.make_layout((1,), stride=(0,)))
         self._kernel(qk_mma,pv_mma,tma_q,mQ,tma_k,mK,tma_v,mV,tma_c,mC,self.cluster_layout_vmnk,self.q_smem_s,self.k_smem_s,self.v_smem_s,self.p_tmem_s,self.p_smem_s,self.c_smem_s,self.epi_tile,lse).launch(grid=(1,1,1),block=[self.threads_per_cta,1,1],stream=stream)
 
     @cute.kernel
@@ -469,7 +469,7 @@ class FmhaKernel:
                 if sfw_idx == 0:
                     _ln2 = Float32(0.6931471805599453)  # ln(2)
                     lse_val = cute.math.log(row_sum, fastmath=True) + _row_max_safe * _ln2
-                    mLSE[0, 0, 0] = lse_val
+                    mLSE[0] = lse_val
 
             tmem.relinquish_alloc_permit()
             tmem.free(tmem_ptr)