D1.3: Initialize tOrP0 before conditional for CuTeDSL scoping

dsv4/kernels/attention/fmha.py

@@ -181,10 +181,11 @@ class FmhaKernel:
         # For SMEM-P, offset is 0 (P not in TMEM).
         # Must be defined unconditionally (CuTeDSL scoping).
         _p0_bf16_offset = max(self.tmem_p0_offset, 0) * (32 // 16)  # Python int
+        # Must define tOrP0 before conditional (CuTeDSL scoping rule).
+        # Initialize to tOrP, then override with offset for TMEM-P.
+        tOrP0 = tOrP
         if _p0_bf16_offset > 0:
             tOrP0 = cute.make_tensor(tOrP.iterator + _p0_bf16_offset, tOrP.layout)
-        else:
-            tOrP0 = tOrP
 
         tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))
         pipeline.pipeline_init_wait(cluster_shape_mn=cl_vmnk)