diff --git a/include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp b/include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
index 1543568eff..c5a1a19e2c 100644
--- a/include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
+++ b/include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp
@@ -366,7 +366,7 @@ struct FmhaFwdSplitKVCombineKernel
                 o_acc_dram_naive,
                 make_tuple(
                     number<kNumWarps>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
-                sequence<true, kPadSeqLenQ, kPadHeadDimV>{});
+                sequence<false, kPadSeqLenQ, kPadHeadDimV>{});
 
             const index_t padded_num_splits =
                 o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<0>{}];
@@ -404,13 +404,10 @@ struct FmhaFwdSplitKVCombineKernel
             make_tuple(number<FmhaPipeline::kMaxSplits>{}, number<FmhaPipeline::kM0>{}),
             {0, i_m0});
 
-        const index_t padded_num_splits =
-            integer_divide_ceil(kargs.num_splits, kNumWarps) * kNumWarps;
-
         auto o_acc_dram_window =
             make_tile_window(o_acc_dram,
                              make_tuple(number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
-                             {i_tile_m * padded_num_splits * FmhaPipeline::kM0, i_n1});
+                             {i_tile_m * kargs.num_splits * FmhaPipeline::kM0, i_n1});
 
         // LSE DRAM window
         auto lse_dram_window = [&, i_nhead_ = i_nhead]() {
diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp
index d119a8c074..1527d04104 100644
--- a/include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp
+++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_combine_pipeline.hpp
@@ -282,12 +282,10 @@ struct BlockFmhaFwdSplitKVCombinePipeline
         auto o_acc = make_static_distributed_tensor<OaccDataType>(o_acc_dist);
         clear_tile(o_acc);
 
-        const index_t padded_num_splits = integer_divide_ceil(num_splits, kNumWarps) * kNumWarps;
-
         __builtin_amdgcn_sched_barrier(0);
         block_sync_lds();
         // each warp handles a [KM0, kN1] tile
-        for(index_t split_start = 0; split_start < padded_num_splits; split_start++)
+        for(index_t split_start = 0; split_start < num_splits; split_start++)
         {
             auto o_tile = load_tile(o_acc_dram_window);
             {