diff --git a/src/mldebug/layer_info.py b/src/mldebug/layer_info.py
index 12ff9fc..7135189 100644
--- a/src/mldebug/layer_info.py
+++ b/src/mldebug/layer_info.py
@@ -13,6 +13,7 @@
 from pathlib import Path
 
 from mldebug.aie_overlay import Overlay
+from mldebug.arch.device_configs import AIE_DEV_PHX, AIE_DEV_STX
 from mldebug.mladf_report import MladfReport
 from mldebug.work_dir import _parse_flexml_layer_id
 from mldebug.work_dir import WorkDir
@@ -26,17 +27,38 @@
   "mllib_graphs::resize_adf_wrapper",
   # This has many sublayers and needs to be better understood
   "mllib_graphs::mha_type1::mha_adf_wrapper",
-  # Causes failure. TODO: investigate
-  "superkernel_eltunary",
   # Padding preamble; halting on it desyncs PC/iteration stepping on HW
   "buffer_pad_innermost",
-  "superkernel_conv_eltbinary",
   # TODO: investigate why this is causing a failure (ref: AIESW-43867)
   "mllib_graphs::topk_adf_wrapper",
   "mllib_graphs::transpose4d_adf_wrapper<signed char>",
   "mllib_graphs::transpose4d_adf_wrapper<bfloat16>",
 ]
 
+# Kernels that work on PHX/STX but are unsupported on other devices.
+# Causes failure on other devices. TODO: investigate
+phx_stx_supported_kernels = [
+  "superkernel_eltunary",
+  "superkernel_conv_eltbinary",
+]
+
+# Set once by the CLI before any LayerInfo is built
+_device = None
+
+
+def set_device(device):
+  """Record the base device used to filter device-specific unsupported kernels."""
+  global _device  # pylint: disable=global-statement
+  _device = device
+
+
+def _is_unsupported_kernel(name):
+  """True if kernel name matches an unsupported superkernel for the current device."""
+  kernels = unsupported_superkernels
+  if _device not in (AIE_DEV_PHX, AIE_DEV_STX):
+    kernels = kernels + phx_stx_supported_kernels
+  return any(k in name for k in kernels)
+
 
 def _strip_template(name):
   """Strip C++ template parameters for compiler-agnostic name comparison.
@@ -321,7 +343,7 @@ def __init__(
 
     # 1. Layers without any kernel should be skipped
     # 2. Unsupported superkernel should be skipped
-    if info.get("is_concat") or not kname or any(k in kname for k in unsupported_superkernels):
+    if info.get("is_concat") or not kname or _is_unsupported_kernel(kname):
       LOGGER.verbose_print(
         f"[WARNING] unsupported kernel {kname} at Layer {self.layer_order} will be skipped."
       )
@@ -336,7 +358,7 @@ def __init__(
         if (
           not stamp.name
           or stamp.elf_name == -1
-          or any(k in stamp.name for k in unsupported_superkernels)
+          or _is_unsupported_kernel(stamp.name)
         ):
           LOGGER.verbose_print(
             f"[WARNING] unsupported kernel {stamp.name} at Layer {self.layer_order} will be skipped."
diff --git a/src/mldebug/mldebug_cli.py b/src/mldebug/mldebug_cli.py
index 8ffaeea..c256c0e 100644
--- a/src/mldebug/mldebug_cli.py
+++ b/src/mldebug/mldebug_cli.py
@@ -23,6 +23,7 @@
   AIE_DEV_TEL,
 )
 from mldebug.client_debug import ClientDebug
+from mldebug import layer_info
 from mldebug.input_parser import (
   check_hw_context,
   check_registry_keys,
@@ -44,8 +45,6 @@ def _apply_unsupported_kernels_from_args(args):
 
   This must happen before LayerInfo creates Layer objects (ClientDebug -> LayerInfo).
   """
-  from mldebug import layer_info  # pylint: disable=import-outside-toplevel
-
   values = args.unsupported_kernels
   if not values:
     return
@@ -190,6 +189,7 @@ def launch_debug(args, output_dir):
   if args.backend == "xrt":
     context_id, pid = check_hw_context(args)
   # Top debug handle
+  layer_info.set_device(args.device)
   _apply_unsupported_kernels_from_args(args)
   _apply_unsupported_layers_from_args(args)
   handle = ClientDebug(args, context_id, pid, output_dir)