diff --git a/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh b/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh
index 7d66627b..5f74472d 100644
--- a/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh
+++ b/dsv4/kernels/attention/fmha_6warp_tma_multirow.cuh
@@ -77,7 +77,8 @@ fmha_6warp_tma_multirow_kernel(FmhaTmaMultiRowParams params) {
     const float scale = params.scale;
 
     const bf16_t* __restrict__ q_head = params.q + head_idx * params.q_head_stride + batch_idx * params.q_batch_stride;
-    const bf16_t* __restrict__ v_head = params.v + head_idx * params.v_head_stride + batch_idx * params.v_batch_stride;
+    // v_head kept for reference; V loaded via TMA
+    // const bf16_t* __restrict__ v_head = params.v + head_idx * params.v_head_stride + batch_idx * params.v_batch_stride;
     bf16_t* __restrict__ o_head = params.o + head_idx * params.o_head_stride + batch_idx * params.o_batch_stride;
     float* __restrict__ lse_head = params.lse ? params.lse + head_idx * params.lse_head_stride + batch_idx * params.lse_batch_stride : nullptr;