diff --git a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/fp8_rowwise_gemm.hip b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/fp8_rowwise_gemm.hip
index 05c90f139..23381213c 100644
--- a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/fp8_rowwise_gemm.hip
+++ b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/fp8_rowwise_gemm.hip
@@ -67,11 +67,11 @@ static const std::unordered_map<std::tuple<int, int, int>, RowwiseKernel, IntTup
     {{128, 7168, 8192},
      fp8_rowwise_256x64x64x128_32x32_1x1_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3},
     {{1024, 7168, 8192},
-     fp8_rowwise_256x128x128x128_32x32_2x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v5},
+     fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3},
     {{2048, 7168, 8192},
-     fp8_rowwise_256x256x224x128_16x16_8x7_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3},
+     fp8_rowwise_256x256x192x128_32x32_4x3_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3},
     {{4096, 7168, 8192},
-     fp8_rowwise_256x224x256x128_16x16_7x8_8x32x1_8x32x1_1x32x1x8_8x8x1_1x2_intrawave_v3},
+     fp8_rowwise_256x256x192x128_32x32_4x3_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3},
     {{8192, 7168, 8192},
      fp8_rowwise_256x256x256x128_16x16_8x8_8x32x1_8x32x1_1x32x1x8_8x8x1_1x2_intrawave_v3},
     // Support for decode across batch sizes for [8192, 3584]
@@ -84,7 +84,7 @@ static const std::unordered_map<std::tuple<int, int, int>, RowwiseKernel, IntTup
     {{128, 8192, 3584},
      fp8_rowwise_256x64x64x128_32x32_1x1_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3},
     {{1024, 8192, 3584},
-     fp8_rowwise_256x256x224x128_16x16_8x7_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3},
+     fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3},
     {{2048, 8192, 3584},
      fp8_rowwise_256x256x224x128_16x16_8x7_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3},
     {{4096, 8192, 3584},
diff --git a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3.hip b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3.hip
new file mode 100644
index 000000000..9becb97bd
--- /dev/null
+++ b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3.hip
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include "fp8_rowwise_common.h"
+
+at::Tensor
+fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3(
+    at::Tensor XQ,
+    at::Tensor WQ,
+    at::Tensor x_scale,
+    at::Tensor w_scale,
+    at::Tensor Y) {
+  // Check if this input needs to be padded.
+  int M = size_to_dim_(XQ.dim() - 1, XQ.sizes());
+  int N = WQ.size(0);
+  int K = WQ.size(1);
+  bool pad = (K % 128 != 0);
+
+  // Dispatch based on whether padding is needed or not.
+  if (pad) {
+    using DeviceGemmInstance = DeviceGemmHelper<
+        256,
+        256,
+        128,
+        128,
+        32,
+        32,
+        4,
+        2,
+        S<8, 32, 1>,
+        S<8, 32, 1>,
+        S<1, 32, 1, 8>,
+        S<8, 8, 1>,
+        1,
+        1,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v3,
+        ck::tensor_operation::device::GemmSpecialization::KPadding>;
+    // Run kernel instance.
+    return f8f8bf16_rowwise_impl<DeviceGemmInstance>(
+        XQ, WQ, x_scale, w_scale, Y);
+  } else {
+    using DeviceGemmInstance = DeviceGemmHelper<
+        256,
+        256,
+        128,
+        128,
+        32,
+        32,
+        4,
+        2,
+        S<8, 32, 1>,
+        S<8, 32, 1>,
+        S<1, 32, 1, 8>,
+        S<8, 8, 1>,
+        1,
+        1,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v3,
+        ck::tensor_operation::device::GemmSpecialization::Default>;
+    // Run kernel instance.
+    return f8f8bf16_rowwise_impl<DeviceGemmInstance>(
+        XQ, WQ, x_scale, w_scale, Y);
+  }
+}
diff --git a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3.hip b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3.hip
new file mode 100644
index 000000000..87fe51852
--- /dev/null
+++ b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3.hip
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include "fp8_rowwise_common.h"
+
+at::Tensor
+fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3(
+    at::Tensor XQ,
+    at::Tensor WQ,
+    at::Tensor x_scale,
+    at::Tensor w_scale,
+    at::Tensor Y) {
+  // Check if this input needs to be padded.
+  int M = size_to_dim_(XQ.dim() - 1, XQ.sizes());
+  int N = WQ.size(0);
+  int K = WQ.size(1);
+  bool kpad = (K % 128 != 0);
+
+  // Dispatch based on whether padding is needed or not.
+  if (kpad) {
+    using DeviceGemmInstance = DeviceGemmHelper<
+        256,
+        256,
+        96,
+        128,
+        32,
+        32,
+        2,
+        3,
+        S<8, 32, 1>,
+        S<8, 32, 1>,
+        S<1, 64, 1, 4>,
+        S<8, 8, 1>,
+        2,
+        1,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v3,
+        ck::tensor_operation::device::GemmSpecialization::KPadding>;
+    // Run kernel instance.
+    return f8f8bf16_rowwise_impl<DeviceGemmInstance>(
+        XQ, WQ, x_scale, w_scale, Y);
+  } else {
+    using DeviceGemmInstance = DeviceGemmHelper<
+        256,
+        256,
+        96,
+        128,
+        32,
+        32,
+        2,
+        3,
+        S<8, 32, 1>,
+        S<8, 32, 1>,
+        S<1, 64, 1, 4>,
+        S<8, 8, 1>,
+        2,
+        1,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v3,
+        ck::tensor_operation::device::GemmSpecialization::Default>;
+    // Run kernel instance.
+    return f8f8bf16_rowwise_impl<DeviceGemmInstance>(
+        XQ, WQ, x_scale, w_scale, Y);
+  }
+}
diff --git a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_kernel_manifest.h b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_kernel_manifest.h
index ebfdccd00..7fb17dd17 100644
--- a/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_kernel_manifest.h
+++ b/fbgemm_gpu/experimental/gen_ai/src/quantize/ck_extensions/fp8_rowwise/kernels/fp8_rowwise_kernel_manifest.h
@@ -289,3 +289,19 @@ fp8_rowwise_256x192x256x128_16x16_6x8_8x32x1_8x32x1_1x32x1x8_8x8x1_2x2_intrawave
     at::Tensor x_scale,
     at::Tensor w_scale,
     at::Tensor Y);
+
+at::Tensor
+fp8_rowwise_256x256x96x128_32x32_2x3_8x32x1_8x32x1_1x64x1x4_8x8x1_2x1_intrawave_v3(
+    at::Tensor XQ,
+    at::Tensor WQ,
+    at::Tensor x_scale,
+    at::Tensor w_scale,
+    at::Tensor Y);
+
+at::Tensor
+fp8_rowwise_256x256x128x128_32x32_4x2_8x32x1_8x32x1_1x32x1x8_8x8x1_1x1_intrawave_v3(
+    at::Tensor XQ,
+    at::Tensor WQ,
+    at::Tensor x_scale,
+    at::Tensor w_scale,
+    at::Tensor Y);