diff --git a/include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp b/include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp
index ba327ee511..9d47061b82 100644
--- a/include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp
+++ b/include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp
@@ -809,36 +809,72 @@ struct FmhaBatchPrefillWithPagedKVCacheKernel
             }
         }();
         const auto k_dram = [&]() {
+            // const auto k_dram_naive = make_naive_tensor_view<address_space_enum::global>(
+            //     k_ptr,
+            //     make_tuple(kargs.num_total_pages * kargs.page_block_size, kargs.hdim_q),
+            //     make_tuple(kargs.stride_k, 1),
+            //     number<FmhaPipeline::kAlignmentK>{},
+            //     number<1>{});
+            // constexpr bool kPadSeqLenK_ = kUseAsyncCopy ? kPadSeqLenK : true;
+            // return pad_tensor_view(
+            //     k_dram_naive,
+            //     make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
+            //     sequence<kPadSeqLenK_, kPadHeadDimQ>{});
+
             const auto k_dram_naive = make_naive_tensor_view<address_space_enum::global>(
                 k_ptr,
-                make_tuple(kargs.num_total_pages * kargs.page_block_size, kargs.hdim_q),
-                make_tuple(kargs.stride_k, 1),
+                make_tuple(kargs.seqlen_k / 16, kargs.hdim_q / 8, 16, 8),
+                make_tuple(kargs.hdim_q * 16, 16 * 8, 8, 1),
                 number<FmhaPipeline::kAlignmentK>{},
                 number<1>{});
 
-            constexpr bool kPadSeqLenK_ = kUseAsyncCopy ? kPadSeqLenK : true;
+	        const auto k_dram_transposed = transform_tensor_view(
+	            k_dram_naive,
+	            make_tuple(make_merge_transform(ck_tile::make_tuple(kargs.seqlen_k /16, 16)),
+	                       make_merge_transform(ck_tile::make_tuple(kargs.hdim_q / 8, 8))),
+	            ck_tile::make_tuple(ck_tile::sequence<0, 2>{}, ck_tile::sequence<1, 3>{}),
+	            ck_tile::make_tuple(ck_tile::sequence<0>{}, ck_tile::sequence<1>{}));
+
+
+            constexpr bool kPadSeqLenK_ = kUseAsyncCopy ? kPadSeqLenK : false;
             return pad_tensor_view(
-                k_dram_naive,
+                k_dram_transposed,
                 make_tuple(number<FmhaPipeline::kN0>{}, number<FmhaPipeline::kK0>{}),
                 sequence<kPadSeqLenK_, kPadHeadDimQ>{});
+
         }();
         const auto v_dram = [&]() {
             if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
             {
+                // const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
+                //     v_ptr,
+                //     make_tuple(kargs.num_total_pages * kargs.page_block_size, kargs.hdim_v),
+                //     make_tuple(kargs.stride_v, 1),
+                //     number<FmhaPipeline::kAlignmentV>{},
+                //     number<1>{});
+                //
+                // const auto v_dram_transposed = transform_tensor_view(
+                //     v_dram_naive,
+                //     make_tuple(
+                //         make_pass_through_transform(kargs.hdim_v),
+                //         make_pass_through_transform(kargs.num_total_pages * kargs.page_block_size)),
+                //     make_tuple(sequence<1>{}, sequence<0>{}),
+                //     make_tuple(sequence<0>{}, sequence<1>{}));
                 const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
                     v_ptr,
-                    make_tuple(kargs.num_total_pages * kargs.page_block_size, kargs.hdim_v),
-                    make_tuple(kargs.stride_v, 1),
+                    make_tuple(kargs.seqlen_k / 16, kargs.hdim_v, 16),
+                    make_tuple(kargs.hdim_v * 16, 16, 1),
+                    // make_tuple(kargs.seqlen_k, kargs.hdim_v),
+                    // make_tuple(kargs.stride_v, 1),
                     number<FmhaPipeline::kAlignmentV>{},
                     number<1>{});
 
-                const auto v_dram_transposed = transform_tensor_view(
-                    v_dram_naive,
-                    make_tuple(
-                        make_pass_through_transform(kargs.hdim_v),
-                        make_pass_through_transform(kargs.num_total_pages * kargs.page_block_size)),
-                    make_tuple(sequence<1>{}, sequence<0>{}),
-                    make_tuple(sequence<0>{}, sequence<1>{}));
+				const auto v_dram_transposed =
+					transform_tensor_view(v_dram_naive,
+										  make_tuple(make_pass_through_transform(kargs.hdim_v),
+								                     make_merge_transform(ck_tile::make_tuple(kargs.seqlen_k /16, 16))),
+										  make_tuple(sequence<1>{}, sequence<0, 2>{}),
+										  make_tuple(sequence<0>{}, sequence<1>{}));
 
                 constexpr bool kPadSeqLenK_ = kUseAsyncCopy ? kPadSeqLenK : true;
                 return pad_tensor_view(
@@ -848,16 +884,37 @@ struct FmhaBatchPrefillWithPagedKVCacheKernel
             }
             else
             {
+                // const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
+                //     v_ptr,
+                //     make_tuple(kargs.hdim_v, kargs.num_total_pages * kargs.page_block_size),
+                //     make_tuple(kargs.stride_v, 1),
+                //     number<FmhaPipeline::kAlignmentV>{},
+                //     number<1>{});
+                //
+                // constexpr bool kPadHeadDimV_ = kUseAsyncCopy ? kPadHeadDimV : false;
+                // return pad_tensor_view(
+                //     v_dram_naive,
+                //     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
+                //     sequence<kPadHeadDimV_, kPadSeqLenK>{});
                 const auto v_dram_naive = make_naive_tensor_view<address_space_enum::global>(
                     v_ptr,
-                    make_tuple(kargs.hdim_v, kargs.num_total_pages * kargs.page_block_size),
-                    make_tuple(kargs.stride_v, 1),
+                    // make_tuple(kargs.hdim_v, kargs.seqlen_k),
+                    // make_tuple(kargs.stride_v, 1),
+                    make_tuple(kargs.seqlen_k / 16, kargs.hdim_v, 16),
+                    make_tuple(kargs.hdim_v * 16, 16, 1),
                     number<FmhaPipeline::kAlignmentV>{},
                     number<1>{});
 
+				const auto v_dram_transposed =
+					transform_tensor_view(v_dram_naive,
+										  make_tuple(make_pass_through_transform(kargs.hdim_v),
+								                     make_merge_transform(ck_tile::make_tuple(kargs.seqlen_k /16, 16))),
+										  make_tuple(sequence<1>{}, sequence<0, 2>{}),
+										  make_tuple(sequence<0>{}, sequence<1>{}));
+
                 constexpr bool kPadHeadDimV_ = kUseAsyncCopy ? kPadHeadDimV : false;
                 return pad_tensor_view(
-                    v_dram_naive,
+                    v_dram_transposed,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
                     sequence<kPadHeadDimV_, kPadSeqLenK>{});
             }