diff --git a/aiter/ops/gemm_op_a8w8.py b/aiter/ops/gemm_op_a8w8.py
index 92832d8256..cbf913dfe8 100644
--- a/aiter/ops/gemm_op_a8w8.py
+++ b/aiter/ops/gemm_op_a8w8.py
@@ -988,6 +988,7 @@ def gemm_a8w8_blockscale_bpreshuffle(
             w_scale,
             dtype=dtype,
             config=_fallback_cfg,
+            is_x_scale_tranposed=x_scale.stride(0) != 1,
         )
     config = get_CKGEMM_config(
         m, n, k, AITER_CONFIGS.AITER_CONFIG_GEMM_A8W8_BLOCKSCALE_BPRESHUFFLE_FILE
@@ -1015,6 +1016,7 @@ def gemm_a8w8_blockscale_bpreshuffle(
             w_scale,
             dtype=dtype,
             backend=backend,
+            is_x_scale_tranposed=x_scale.stride(0) != 1,
         )
     if config is not None:
         libtype = config["libtype"]
diff --git a/aiter/ops/triton/gemm/basic/gemm_a8w8_blockscale.py b/aiter/ops/triton/gemm/basic/gemm_a8w8_blockscale.py
index fead8a5ba5..7870b4791e 100644
--- a/aiter/ops/triton/gemm/basic/gemm_a8w8_blockscale.py
+++ b/aiter/ops/triton/gemm/basic/gemm_a8w8_blockscale.py
@@ -6,6 +6,7 @@
 
 import torch
 import triton
+from packaging.version import Version
 
 from aiter.ops.triton._triton_kernels.common.splitk_reduce import (
     _gemm_splitk_reduce_kernel,
@@ -25,6 +26,7 @@
 
 _LOGGER = AiterTritonLogger()
 _FORCE_GFX1250_EX = os.environ.get("AITER_FORCE_GFX1250_EX", "0") == "1"
+_TRITON_VERSION = Version(triton.__version__)
 
 _GLUON_SUPPORTED_ARCHS = ("gfx1250",)
 
@@ -305,6 +307,16 @@ def gemm_a8w8_blockscale_preshuffle(
     if config is None:
         config, _ = _get_config(M, N, K, True, backend=backend)
 
+    # Triton 3.6 fails TritonAMDGPUConvertToBufferOps for gfx950 preshuffle
+    # configs with three pipeline stages. Keep the tuned tile and split-K.
+    if (
+        backend == "triton"
+        and get_arch() == "gfx950"
+        and _TRITON_VERSION < Version("3.7.0")
+        and config.get("num_stages", 1) > 2
+    ):
+        config["num_stages"] = 2
+
     kernel_type_from_config = config.pop("kernel_type", None)
     if kernel_type_from_config is not None:
         kernel_type = kernel_type_from_config
diff --git a/op_tests/test_gemm_a8w8_blockscale.py b/op_tests/test_gemm_a8w8_blockscale.py
index c99ce8f5ec..9281279bba 100644
--- a/op_tests/test_gemm_a8w8_blockscale.py
+++ b/op_tests/test_gemm_a8w8_blockscale.py
@@ -116,6 +116,17 @@ def test_gemm(dtype, m, n, k, ck_preshuffle=True, use_flydsl=False):
     b, avg_b = run_func(x, gemm_weight, gemm_x_scale, w_scale, dtype)
 
     err_ck = checkAllclose(a, b, msg="ck", catastrophic_check=True)
+    if ck_preshuffle:
+        x_scale_strided = x_scale.transpose(0, 1).contiguous().transpose(0, 1)
+        b_strided = aiter.gemm_a8w8_blockscale_bpreshuffle(
+            x, gemm_weight, x_scale_strided, w_scale, dtype
+        )
+        checkAllclose(
+            a,
+            b_strided,
+            msg="ck strided x_scale",
+            catastrophic_check=True,
+        )
     ret["ck us"] = avg_b
     ret["ck TFLOPS"] = m * n * k * 2 / avg_b / 1e6
     ret["ck TB/s"] = (x.nbytes + weight.nbytes) / avg_b / 1e6