Fix CuTeDSL variable scoping: define tOrP0 and tCrP in both branches

dsv4/kernels/attention/fmha.py

@@ -252,6 +252,8 @@ class FmhaKernel:
         tOtO0 = cute.make_tensor(tOtO.iterator + self.tmem_o0_offset, tOtO.layout)
 
         # ── TMEM-P path: PV A-operand from TMEM ──
+        # Define both paths' variables before the branch (CuTeDSL requires variables
+        # to exist before dynamic control flow, even when the branch is compile-time)
         if not use_smem_p:
             tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
             tOrP_base = pv_thr.make_fragment_A(tP)
@@ -260,6 +262,19 @@ class FmhaKernel:
                 tOrP.iterator + self.qk_acc_dtype.width // self.q_dtype.width * self.tmem_p0_offset,
                 tOrP.layout,
             )
+        else:
+            # SMEM-P: PV reads from SMEM, but define tOrP0 as unused dummy
+            tP = cute.make_tensor(tStS.iterator, p_tmem_s.outer)
+            tOrP_base = pv_thr.make_fragment_A(tP)
+            tOrP = tOrP_base[(None, None, None, 0)]
+            tOrP0 = cute.make_tensor(tOrP.iterator, tOrP.layout)
+
+        # ── SMEM-P path: PV A-operand from SMEM ──
+        if use_smem_p:
+            tCrP = pv_mma.make_fragment_A(sP)
+        else:
+            # TMEM-P: PV reads from TMEM, but define tCrP as unused dummy
+            tCrP = pv_mma.make_fragment_A(sP)
 
         tCtO_fake = pv_mma.make_fragment_C(cute.append(pv_as, self.num_acc_stage))