D1: Fix kvh scoping - define before loops, consume V via pipeline

dsv4/kernels/attention/fmha.py

@@ -247,6 +247,7 @@ class FmhaKernel:
             tmem.wait_for_alloc()
             if const_expr(self.n_k_sub_tiles > 1):
                 # K sub-tiling path (hd=512): pipeline sync for k_sub iterations
+                kvh = kvc.wait_and_advance()  # initialize kvh before loops
                 for kt in range(self.n_kv_tiles):
                     for k_sub in range(self.n_k_sub_tiles):
                         qh = qc.wait_and_advance(); qh.release()
@@ -261,6 +262,8 @@ class FmhaKernel:
                     softmax_done_bar.arrive()
                     softmax_done_bar.arrive_and_wait()
                     pv_mma.set(tcgen05.Field.ACCUMULATE, kt != 0)
+                    # Load V: consume from K/V pipeline
+                    kvh = kvc.wait_and_advance()
                     if not self.use_smem_p:
                         for kb in cutlass.range(cute.size(tOrP0, mode=[2]), unroll_full=True):
                             cute.gemm(pv_mma, tOtO0, tOrP0[(None,None,kb)], tCrV[(None,None,kb,kvh.index)], tOtO0)
@@ -270,6 +273,7 @@ class FmhaKernel:
                             cute.gemm(pv_mma, tOtO0, tCrP[(None,None,kb,0)], tCrV[(None,None,kb,kvh.index)], tOtO0)
                             pv_mma.set(tcgen05.Field.ACCUMULATE, True)
                     cute.arch.fence_view_async_tmem_store()
+                    kvh.release()
                 final_o_bar.arrive()
             else:
                 # Original pipeline path (hd≤256)