From d2bc893fdc78f2d6cdc7b1c0b65efe275cfc3621 Mon Sep 17 00:00:00 2001
From: anurag <anurag@amd.com>
Date: Wed, 29 Jul 2026 09:55:58 -0600
Subject: [PATCH 1/4] add subdevices

Signed-off-by: anurag <anurag@amd.com>
---
 mldebuglib.py                         |   4 +-
 src/mldebug/arch/__init__.py          |  26 ++++-
 src/mldebug/arch/aie2p_defs.py        |  16 ++-
 src/mldebug/arch/aie2ps_defs.py       |  16 ++-
 src/mldebug/arch/device_configs.py    | 146 ++++++++++++++++++++++++++
 src/mldebug/arch/loader.py            |  25 +++--
 src/mldebug/backend/core_dump_impl.py |  52 ++-------
 src/mldebug/backend/factory.py        |   9 +-
 src/mldebug/client_debug.py           |   1 +
 src/mldebug/input_parser.py           | 113 +++++++++++++++-----
 src/mldebug/mldebug_cli.py            |  11 +-
 11 files changed, 320 insertions(+), 99 deletions(-)
 create mode 100644 src/mldebug/arch/device_configs.py

diff --git a/mldebuglib.py b/mldebuglib.py
index b7374e8..23e90d5 100644
--- a/mldebuglib.py
+++ b/mldebuglib.py
@@ -10,7 +10,7 @@
 from types import SimpleNamespace
 
 from mldebug.aie_status import AIEStatus as _AIES
-from mldebug.arch import AIE_DEV_PHX, AIE_DEV_STX
+from mldebug.arch import AIE_DEV_STX
 from mldebug.arch import load_aie_arch as _load_aie
 from mldebug.aie_overlay import Overlay as _OL
 from mldebug.input_parser import RunFlags as _RunFlags
@@ -64,7 +64,7 @@ def __init__(self, device=AIE_DEV_STX, overlay="4x4", ctxid=None, pid=None, back
       raise ValueError(f"Unsupported backend '{backend}'. Use BACKEND_XRT or BACKEND_TEST.")
     self.backend = backend
     self.aie_iface = _load_aie(device)
-    self.aie_iface.init(device == AIE_DEV_PHX)
+    self.aie_iface.init(device)
 
     if backend == BACKEND_XRT:
       args = SimpleNamespace(device=device, aie_iface=self.aie_iface, l3=False)
diff --git a/src/mldebug/arch/__init__.py b/src/mldebug/arch/__init__.py
index 5fde850..a741d10 100644
--- a/src/mldebug/arch/__init__.py
+++ b/src/mldebug/arch/__init__.py
@@ -5,4 +5,28 @@
 Top level AIE Arch Module
 """
 
-from .loader import load_aie_arch, AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3
+from .device_configs import (
+  AIE_DEV_NPU3,
+  AIE_DEV_PHX,
+  AIE_DEV_STX,
+  AIE_DEV_T10,
+  AIE_DEV_T50,
+  AIE_DEV_TEL,
+  DEVICE_CONFIGS,
+  get_base_device,
+  resolve_variant,
+)
+from .loader import load_aie_arch
+
+__all__ = [
+  "AIE_DEV_NPU3",
+  "AIE_DEV_PHX",
+  "AIE_DEV_STX",
+  "AIE_DEV_T10",
+  "AIE_DEV_T50",
+  "AIE_DEV_TEL",
+  "DEVICE_CONFIGS",
+  "get_base_device",
+  "load_aie_arch",
+  "resolve_variant",
+]
diff --git a/src/mldebug/arch/aie2p_defs.py b/src/mldebug/arch/aie2p_defs.py
index b5e860a..437c3f0 100644
--- a/src/mldebug/arch/aie2p_defs.py
+++ b/src/mldebug/arch/aie2p_defs.py
@@ -7,6 +7,8 @@
 
 import json
 
+from .device_configs import AIE_DEV_PHX, DEVICE_CONFIGS
+
 AIE_TILE_T = "aie_tile"
 SHIM_TILE_T = "shim_tile"
 MEM_TILE_T = "mem_tile"
@@ -126,11 +128,19 @@ def _create_bds(tile_t, registers):
 }
 
 
-def init(is_aie2):
+def init(device=None):
   """
-  Inititalize aie2/2p specific
+  Initialize aie2/2p specific state from the central registry.
+
+  `device` is the resolved sub-device name (e.g. 'phx' or 'stx'). It selects
+  AIE_TILE_ROW_OFFSET/MEM_TILE_SZ; AIE2 (phx) also relocates CORE_PC.
   """
-  if is_aie2:
+  global AIE_TILE_ROW_OFFSET, MEM_TILE_SZ
+  cfg = DEVICE_CONFIGS.get(device)
+  if cfg:
+    AIE_TILE_ROW_OFFSET = cfg["core_row_start"]
+    MEM_TILE_SZ = cfg["mem_tile_sz"]
+  if device == AIE_DEV_PHX:
     Core_registers["CORE_PC"] = 0x31100
 
 
diff --git a/src/mldebug/arch/aie2ps_defs.py b/src/mldebug/arch/aie2ps_defs.py
index 8b2d38e..5da306d 100644
--- a/src/mldebug/arch/aie2ps_defs.py
+++ b/src/mldebug/arch/aie2ps_defs.py
@@ -7,6 +7,8 @@
 
 import json
 
+from .device_configs import DEVICE_CONFIGS
+
 AIE_TILE_T = "aie_tile"
 SHIM_TILE_T = "shim_tile"
 MEM_TILE_T = "mem_tile"
@@ -128,11 +130,19 @@ def _create_bds(tile_t, registers):
 }
 
 
-def init(_):
+def init(device=None):
   """
-  consistent interface with aie2p
+  Apply device/variant geometry from the central registry.
+
+  `device` is the resolved sub-device name (e.g. 'telluride' or 't50');
+  it selects AIE_TILE_ROW_OFFSET (= core_row_start) and MEM_TILE_SZ so
+  same-hwGen variants can differ in geometry.
   """
-  return
+  global AIE_TILE_ROW_OFFSET, MEM_TILE_SZ
+  cfg = DEVICE_CONFIGS.get(device)
+  if cfg:
+    AIE_TILE_ROW_OFFSET = cfg["core_row_start"]
+    MEM_TILE_SZ = cfg["mem_tile_sz"]
 
 
 _create_bds(AIE_TILE_T, Core_registers)
diff --git a/src/mldebug/arch/device_configs.py b/src/mldebug/arch/device_configs.py
new file mode 100644
index 0000000..47cdeea
--- /dev/null
+++ b/src/mldebug/arch/device_configs.py
@@ -0,0 +1,146 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright (C) 2024-2026 Advanced Micro Devices, Inc. All rights reserved.
+
+"""
+Central device geometry registry -- single source of truth for per-device
+and per-variant AIE geometry.
+
+Keyed by device/variant name. Fields:
+  arch      -- which defs module the loader should import (aie2p/aie2ps)
+  base      -- base device name understood by the C++ binding and xrt-smi;
+               variants (e.g. t50) map back to their base (telluride)
+  hwGen     -- hardware generation (matches the core-dump header hwGen)
+  core_row_start -- first core-tile row; also the AIE tile row offset
+  numrows/numcols -- full-device geometry, used to disambiguate variants
+                     that share a hwGen
+
+Variants may share a hwGen (telluride/t50) and are disambiguated by
+(numrows, numcols).
+"""
+
+MB = 1024 * 1024
+
+# Device / variant name constants. Defined here (the lowest-level arch module)
+# so they can key DEVICE_CONFIGS; re-exported by loader for the rest of the code.
+AIE_DEV_PHX = "phx"
+AIE_DEV_STX = "stx"
+AIE_DEV_TEL = "telluride"
+AIE_DEV_NPU3 = "npu3"
+AIE_DEV_T50 = "t50"
+AIE_DEV_T10 = "t10"
+
+# Arch (defs-module) names, resolved to modules by the loader.
+ARCH_AIE2P = "aie2p"
+ARCH_AIE2PS = "aie2ps"
+
+DEVICE_CONFIGS = {
+  AIE_DEV_PHX: {
+    "arch": ARCH_AIE2P,
+    "base": AIE_DEV_PHX,
+    "hwGen": 2,
+    "baseAddr": 0x0,
+    "core_row_start": 2,
+    "mem_row_start": 1,
+    "memtile_rows": 1,
+    "numrows": 6,
+    "numcols": 4,
+    "shim_tile_block_size": MB,
+    "mem_tile_block_size": MB,
+    "core_tile_block_size": MB,
+    "mem_tile_sz": 0x80000,
+  },
+  AIE_DEV_STX: {
+    "arch": ARCH_AIE2P,
+    "base": AIE_DEV_STX,
+    "hwGen": 4,
+    "baseAddr": 0x0,
+    "core_row_start": 2,
+    "mem_row_start": 1,
+    "memtile_rows": 1,
+    "numrows": 6,
+    "numcols": 8,
+    "shim_tile_block_size": MB,
+    "mem_tile_block_size": MB,
+    "core_tile_block_size": MB,
+    "mem_tile_sz": 0x80000,
+  },
+  AIE_DEV_TEL: {
+    "arch": ARCH_AIE2PS,
+    "base": AIE_DEV_TEL,
+    "hwGen": 5,
+    "baseAddr": 0x0,
+    "core_row_start": 3,
+    "mem_row_start": 1,
+    "memtile_rows": 2,
+    "numrows": 7,
+    "numcols": 36,
+    "shim_tile_block_size": MB,
+    "mem_tile_block_size": MB,
+    "core_tile_block_size": MB,
+    "mem_tile_sz": 0x80000,
+  },
+  # aie2ps-based variants sharing telluride's hwGen; disambiguated by geometry.
+  AIE_DEV_T50: {
+    "arch": ARCH_AIE2PS,
+    "base": AIE_DEV_TEL,
+    "hwGen": 5,
+    "baseAddr": 0x0,
+    "core_row_start": 2,
+    "mem_row_start": 1,
+    "memtile_rows": 1,
+    "numrows": 6,
+    "numcols": 24,
+    "shim_tile_block_size": MB,
+    "mem_tile_block_size": MB,
+    "core_tile_block_size": MB,
+    "mem_tile_sz": 0x80000,
+  },
+  AIE_DEV_T10: {
+    "arch": ARCH_AIE2PS,
+    "base": AIE_DEV_TEL,
+    "hwGen": 5,
+    "baseAddr": 0x0,
+    "core_row_start": 2,
+    "mem_row_start": 1,
+    "memtile_rows": 1,
+    "numrows": 4,
+    "numcols": 8,
+    "shim_tile_block_size": MB,
+    "mem_tile_block_size": MB,
+    "core_tile_block_size": MB,
+    "mem_tile_sz": 0x80000,
+  },
+}
+
+
+def get_base_device(name):
+  """
+  Return the base device name for a device/variant. Unknown names map to
+  themselves so callers can pass through user-specified devices unchanged.
+  """
+  cfg = DEVICE_CONFIGS.get(name)
+  return cfg["base"] if cfg else name
+
+
+def resolve_variant(hw_gen, num_rows=None, num_cols=None):
+  """
+  Resolve a device/variant name from a hardware generation and (optional)
+  full-device geometry.
+
+  Matches hwGen first. When several variants share a hwGen, disambiguate by
+  (num_rows, num_cols). Falls back to the first (base) candidate when the
+  geometry is unavailable or does not match a variant. Returns None if no
+  device matches the hwGen.
+  """
+  candidates = [n for n, c in DEVICE_CONFIGS.items() if c["hwGen"] == hw_gen]
+  if not candidates:
+    return None
+  if len(candidates) == 1:
+    return candidates[0]
+  if num_rows is not None and num_cols is not None:
+    for n in candidates:
+      c = DEVICE_CONFIGS[n]
+      if c["numrows"] == num_rows and c["numcols"] == num_cols:
+        return n
+  # Base device is listed first among same-hwGen candidates.
+  return candidates[0]
diff --git a/src/mldebug/arch/loader.py b/src/mldebug/arch/loader.py
index be7738b..3dadfaa 100644
--- a/src/mldebug/arch/loader.py
+++ b/src/mldebug/arch/loader.py
@@ -7,19 +7,24 @@
 
 import importlib
 
-AIE_DEV_PHX = "phx"
-AIE_DEV_STX = "stx"
-AIE_DEV_TEL = "telluride"
-AIE_DEV_NPU3 = "npu3"
+from .device_configs import AIE_DEV_NPU3, ARCH_AIE2P, ARCH_AIE2PS, DEVICE_CONFIGS
+
+# Maps a config's `arch` field to its defs module.
+_ARCH_MODULES = {
+  ARCH_AIE2P: ".aie2p_defs",
+  ARCH_AIE2PS: ".aie2ps_defs",
+}
 
 
 def load_aie_arch(device):
   """
-  return specific aie arch module based on name
+  return specific aie arch module based on device/variant name
   """
-  mod = ".aie2p_defs"
-  if device == AIE_DEV_TEL:
-    mod = ".aie2ps_defs"
-  elif device == AIE_DEV_NPU3:
-    mod = ".npu3_defs"
+  # npu3 is not yet in the geometry registry; keep it as a special case.
+  if device == AIE_DEV_NPU3:
+    return importlib.import_module(".npu3_defs", package="mldebug.arch")
+
+  cfg = DEVICE_CONFIGS.get(device)
+  arch = cfg["arch"] if cfg else ARCH_AIE2P
+  mod = _ARCH_MODULES.get(arch, ".aie2p_defs")
   return importlib.import_module(mod, package="mldebug.arch")
diff --git a/src/mldebug/backend/core_dump_impl.py b/src/mldebug/backend/core_dump_impl.py
index aeebb10..ed409ea 100644
--- a/src/mldebug/backend/core_dump_impl.py
+++ b/src/mldebug/backend/core_dump_impl.py
@@ -7,8 +7,10 @@
 
 import struct
 from pathlib import Path
+
+from mldebug.arch import DEVICE_CONFIGS, resolve_variant
 from mldebug.utils import print_tile_grid
-from mldebug.arch import AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3
+
 from .backend_interface import BackendInterface
 
 try:
@@ -18,46 +20,6 @@
 except ImportError:
   HAS_XRT_BACKEND = False
 
-# Device architecture metadata (from C++ CoreDumpDataAccessBackend)
-DEVICE_CONFIGS = {
-  AIE_DEV_PHX: {
-    "hwGen": 2,
-    "baseAddr": 0x0,
-    "core_row_start": 2,
-    "mem_row_start": 1,
-    "memtile_rows": 1,
-    "numrows": 6,
-    "numcols": 4,
-    "shim_tile_block_size": 1024 * 1024,  # 1MB
-    "mem_tile_block_size": 1024 * 1024,  # 1MB
-    "core_tile_block_size": 1024 * 1024,  # 1MB
-  },
-  AIE_DEV_STX: {
-    "hwGen": 4,
-    "baseAddr": 0x0,
-    "core_row_start": 2,
-    "mem_row_start": 1,
-    "memtile_rows": 1,
-    "numrows": 6,
-    "numcols": 8,
-    "shim_tile_block_size": 1024 * 1024,
-    "mem_tile_block_size": 1024 * 1024,
-    "core_tile_block_size": 1024 * 1024,
-  },
-  AIE_DEV_TEL: {
-    "hwGen": 5,
-    "baseAddr": 0x0,
-    "core_row_start": 3,
-    "mem_row_start": 1,
-    "memtile_rows": 2,
-    "numrows": 7,
-    "numcols": 36,
-    "shim_tile_block_size": 1024 * 1024,
-    "mem_tile_block_size": 1024 * 1024,
-    "core_tile_block_size": 1024 * 1024,
-  },
-}
-
 
 class CoreDumpFallbackReader:
   """
@@ -155,11 +117,9 @@ def peek_device(filename):
       "<BBBBBB", header[8:14]
     )
 
-    detected = None
-    for name, cfg in DEVICE_CONFIGS.items():
-      if cfg["hwGen"] == hw_gen:
-        detected = name
-        break
+    # Match hwGen, disambiguating same-hwGen variants (e.g. telluride/t50)
+    # by the header's total rows/cols.
+    detected = resolve_variant(hw_gen, total_rows, total_cols)
 
     print("[INFO] Core dump header:")
     print(f"  Magic: {magic.decode('ascii', errors='ignore').rstrip(chr(0))}")
diff --git a/src/mldebug/backend/factory.py b/src/mldebug/backend/factory.py
index b94f437..dc52326 100644
--- a/src/mldebug/backend/factory.py
+++ b/src/mldebug/backend/factory.py
@@ -22,15 +22,17 @@ class BackendConfig:
   Superset of parameters needed by all backend constructors.
 
   Each backend extracts only the fields it needs:
-    XRT:       tiles, ctx_id, pid, device
+    XRT:       tiles, ctx_id, pid, device (base device for the binding/xrt-smi)
     Test:      tiles, design_info, args
-    CoreDump:  tiles, ctx_id, pid, device, core_dump_file, no_header
+    CoreDump:  tiles, ctx_id, pid, sub_device (variant geometry),
+               core_dump_file, no_header
   """
 
   tiles: list = field(default_factory=list)
   ctx_id: int = 0
   pid: int = 0
   device: str = ""
+  sub_device: str = ""
   design_info: Any = None
   args: Any = None
   core_dump_file: str = None
@@ -69,12 +71,13 @@ def create_backend(backend_type, config):
     return test_mod.TestImpl(config.tiles, config.design_info, config.args)
 
   # core_dump (default)
+  # Uses the variant name so same-hwGen variants read the correct geometry.
   core_dump_mod = importlib.import_module("mldebug.backend.core_dump_impl")
   return core_dump_mod.CoreDumpImpl(
     config.tiles,
     config.ctx_id,
     config.pid,
-    config.device,
+    config.sub_device or config.device,
     core_dump_file=config.core_dump_file,
     no_header=config.no_header,
   )
diff --git a/src/mldebug/client_debug.py b/src/mldebug/client_debug.py
index 85cfbc7..ed79832 100644
--- a/src/mldebug/client_debug.py
+++ b/src/mldebug/client_debug.py
@@ -79,6 +79,7 @@ def __init__(self, args, ctx_id, pid, output_dir):
         ctx_id=ctx_id,
         pid=pid,
         device=args.device,
+        sub_device=getattr(args, "sub_device", None) or args.device,
         design_info=self.design_info,
         args=args,
         core_dump_file=getattr(args, "core_dump", None),
diff --git a/src/mldebug/input_parser.py b/src/mldebug/input_parser.py
index 83cf96b..f94d612 100644
--- a/src/mldebug/input_parser.py
+++ b/src/mldebug/input_parser.py
@@ -15,7 +15,14 @@
 import subprocess
 import re
 
-from mldebug.arch import load_aie_arch, AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL
+from mldebug.arch import (
+  load_aie_arch,
+  resolve_variant,
+  get_base_device,
+  AIE_DEV_PHX,
+  AIE_DEV_STX,
+  AIE_DEV_TEL,
+)
 from mldebug.backend.core_dump_impl import CoreDumpFallbackReader
 from mldebug.backend.factory import BackendConfig, create_backend
 from mldebug.utils import LOGGER, cleanup_and_exit, input_with_timeout, is_aarch64, is_windows
@@ -102,9 +109,13 @@ def create_run_flags(args, subgraph_path: str, fsp: str, fsp_execution_order: li
     args.interactive = True
 
   # AIE interface for aie2p and aie2 are shared
-  # We need to differentiate between them for a few items
-  args.aie_iface = load_aie_arch(args.device)
-  args.aie_iface.init(args.device == AIE_DEV_PHX)
+  # We need to differentiate between them for a few items.
+  # sub_device carries any same-hwGen variant (e.g. t50); it drives the arch
+  # module and geometry, while args.device stays the base device for backends.
+  sub_device = getattr(args, "sub_device", None) or args.device
+  args.sub_device = sub_device
+  args.aie_iface = load_aie_arch(sub_device)
+  args.aie_iface.init(sub_device)
 
   # Finally Create run flags
   if isinstance(args.run_flags, RunFlags):
@@ -222,9 +233,58 @@ def check_registry_keys(args, npu3=False) -> None:
     LOGGER.log("\nRegistry settings check passed. No modifications were necessary.")
 
 
+def _detect_vaiml_variant(aie_dir):
+  """
+  Detect the device/variant for a VAIML design.
+
+  Primary source is aie_trace_config.json's driver_config (hw_gen plus full
+  device geometry), which lets us distinguish same-hwGen variants (e.g.
+  telluride vs t50). Falls back to the HW_GEN define in aie_control.cpp when
+  the trace config is missing. Both files live in <work>/ps/c_rts/.
+
+  Returns a device/variant name, or None when nothing could be detected
+  (caller keeps the platform default).
+  """
+  c_rts = f"{aie_dir}/ps/c_rts"
+
+  # Primary: aie_trace_config.json driver_config
+  trace_cfg = f"{c_rts}/aie_trace_config.json"
+  try:
+    with open(trace_cfg, encoding="utf-8") as f:
+      driver = json.load(f)["aie_metadata"]["driver_config"]
+    variant = resolve_variant(
+      int(driver["hw_gen"]), int(driver["num_rows"]), int(driver["num_columns"])
+    )
+    if variant:
+      return variant
+  except (FileNotFoundError, KeyError, ValueError, json.JSONDecodeError):
+    pass
+
+  # Fallback: HW_GEN define in aie_control.cpp
+  ctrl_cpp = f"{c_rts}/aie_control.cpp"
+  try:
+    with open(ctrl_cpp, encoding="utf-8") as f:
+      for line in f.read().split("\n"):
+        if "#define HW_GEN" in line:
+          genstr = line.split(" ")[-1]
+          if genstr == "XAIE_DEV_GEN_AIE2PS":
+            return AIE_DEV_TEL
+          if genstr == "XAIE_DEV_GEN_AIE2":
+            return AIE_DEV_PHX
+          break
+  except (FileNotFoundError, KeyError):
+    pass
+  return None
+
+
 def set_device(args) -> None:
   """
-  Detects and sets the device target (phx, stx, or tel) for the current work directory.
+  Detects and sets the device target for the current work directory.
+
+  Sets two fields: ``args.device`` (base device understood by the backends,
+  binding, and xrt-smi) and ``args.sub_device`` (the resolved variant, e.g.
+  't50', used to select the arch module and geometry). For most devices these
+  are identical.
 
   Args:
       args: Argument object that is updated to set the detected device.
@@ -233,36 +293,31 @@ def set_device(args) -> None:
       None
   """
   endmsg = "\n"
-  if not args.device:
+  if args.device:
+    # User-specified device; honor it as the variant and derive the base.
+    args.sub_device = args.device
+    args.device = get_base_device(args.device)
+  else:
     endmsg = " Use -d to specify a diferent device.\n"
+    variant = None
+
     # For core dumps, the device is baked into the file header. Detect it now
     # so the overlay (built before the backend) uses the correct aie_iface.
     if getattr(args, "core_dump", None) and not getattr(args, "no_header", False):
-      cd_dev = CoreDumpFallbackReader.peek_device(args.core_dump)
-      if cd_dev:
-        args.device = cd_dev
-        print(f"[INFO] Using AIE Device: {args.device} (detected from core dump header).")
-        return
+      variant = CoreDumpFallbackReader.peek_device(args.core_dump)
 
-    # if on ARM, default is telluride else STX
-    args.device = AIE_DEV_TEL if is_aarch64() else AIE_DEV_STX
-    genstr = "XAIE_DEV_GEN_AIE2P"
+    if variant is None:
+      variant = _detect_vaiml_variant(args.aie_dir)
 
-    ctrl_cpp = args.aie_dir + "/ps/c_rts/aie_control.cpp"
-    try:
-      with open(ctrl_cpp, encoding="utf-8") as f:
-        data = f.read().split("\n")
-        for line in data:
-          if "#define HW_GEN" in line:
-            genstr = line.split(" ")[-1]
-            break
-        if genstr == "XAIE_DEV_GEN_AIE2PS":
-          args.device = AIE_DEV_TEL
-        if genstr == "XAIE_DEV_GEN_AIE2":
-          args.device = AIE_DEV_PHX
-    except (FileNotFoundError, KeyError):
-      pass
-      # LOGGER.log("[INFO] Unable to detect device automatically.")
+    if variant is None:
+      # if on ARM, default is telluride else STX
+      variant = AIE_DEV_TEL if is_aarch64() else AIE_DEV_STX
+
+    args.sub_device = variant
+    args.device = get_base_device(variant)
+
+  if args.sub_device != args.device:
+    print(f"[INFO] Detected sub-device: {args.sub_device} (base {args.device}).")
   print(f"[INFO] Using AIE Device: {args.device}.", end=endmsg)
 
 
diff --git a/src/mldebug/mldebug_cli.py b/src/mldebug/mldebug_cli.py
index da84f00..f3bcde0 100644
--- a/src/mldebug/mldebug_cli.py
+++ b/src/mldebug/mldebug_cli.py
@@ -14,7 +14,14 @@
 import os
 import time
 
-from mldebug.arch import AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3
+from mldebug.arch import (
+  AIE_DEV_NPU3,
+  AIE_DEV_PHX,
+  AIE_DEV_STX,
+  AIE_DEV_T10,
+  AIE_DEV_T50,
+  AIE_DEV_TEL,
+)
 from mldebug.client_debug import ClientDebug
 from mldebug.input_parser import (
   check_hw_context,
@@ -244,7 +251,7 @@ def app():
     "-d",
     "--device",
     help="Specify device if it can't be detected from aie_dir.",
-    choices=[AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3],
+    choices=[AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3, AIE_DEV_T50, AIE_DEV_T10],
     required=False,
   )
   # Hidden Argument

From 0bd8d74867cff7ae937a1db7ca16bc4c19f7c2c0 Mon Sep 17 00:00:00 2001
From: anurag <anurag@amd.com>
Date: Wed, 29 Jul 2026 10:23:06 -0600
Subject: [PATCH 2/4] fix type

Signed-off-by: anurag <anurag@amd.com>
---
 src/mldebug/arch/__init__.py          | 4 ++--
 src/mldebug/arch/aie2ps_defs.py       | 2 +-
 src/mldebug/arch/device_configs.py    | 8 ++++----
 src/mldebug/backend/core_dump_impl.py | 2 +-
 src/mldebug/input_parser.py           | 6 +++---
 src/mldebug/mldebug_cli.py            | 4 ++--
 6 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/src/mldebug/arch/__init__.py b/src/mldebug/arch/__init__.py
index a741d10..35cbfa2 100644
--- a/src/mldebug/arch/__init__.py
+++ b/src/mldebug/arch/__init__.py
@@ -10,7 +10,7 @@
   AIE_DEV_PHX,
   AIE_DEV_STX,
   AIE_DEV_T10,
-  AIE_DEV_T50,
+  AIE_DEV_T20,
   AIE_DEV_TEL,
   DEVICE_CONFIGS,
   get_base_device,
@@ -23,7 +23,7 @@
   "AIE_DEV_PHX",
   "AIE_DEV_STX",
   "AIE_DEV_T10",
-  "AIE_DEV_T50",
+  "AIE_DEV_T20",
   "AIE_DEV_TEL",
   "DEVICE_CONFIGS",
   "get_base_device",
diff --git a/src/mldebug/arch/aie2ps_defs.py b/src/mldebug/arch/aie2ps_defs.py
index 5da306d..9eafe44 100644
--- a/src/mldebug/arch/aie2ps_defs.py
+++ b/src/mldebug/arch/aie2ps_defs.py
@@ -134,7 +134,7 @@ def init(device=None):
   """
   Apply device/variant geometry from the central registry.
 
-  `device` is the resolved sub-device name (e.g. 'telluride' or 't50');
+  `device` is the resolved sub-device name (e.g. 'telluride' or 't20');
   it selects AIE_TILE_ROW_OFFSET (= core_row_start) and MEM_TILE_SZ so
   same-hwGen variants can differ in geometry.
   """
diff --git a/src/mldebug/arch/device_configs.py b/src/mldebug/arch/device_configs.py
index 47cdeea..022b4ba 100644
--- a/src/mldebug/arch/device_configs.py
+++ b/src/mldebug/arch/device_configs.py
@@ -8,13 +8,13 @@
 Keyed by device/variant name. Fields:
   arch      -- which defs module the loader should import (aie2p/aie2ps)
   base      -- base device name understood by the C++ binding and xrt-smi;
-               variants (e.g. t50) map back to their base (telluride)
+               variants (e.g. t20) map back to their base (telluride)
   hwGen     -- hardware generation (matches the core-dump header hwGen)
   core_row_start -- first core-tile row; also the AIE tile row offset
   numrows/numcols -- full-device geometry, used to disambiguate variants
                      that share a hwGen
 
-Variants may share a hwGen (telluride/t50) and are disambiguated by
+Variants may share a hwGen (telluride/t20) and are disambiguated by
 (numrows, numcols).
 """
 
@@ -26,7 +26,7 @@
 AIE_DEV_STX = "stx"
 AIE_DEV_TEL = "telluride"
 AIE_DEV_NPU3 = "npu3"
-AIE_DEV_T50 = "t50"
+AIE_DEV_T20 = "t20"
 AIE_DEV_T10 = "t10"
 
 # Arch (defs-module) names, resolved to modules by the loader.
@@ -80,7 +80,7 @@
     "mem_tile_sz": 0x80000,
   },
   # aie2ps-based variants sharing telluride's hwGen; disambiguated by geometry.
-  AIE_DEV_T50: {
+  AIE_DEV_T20: {
     "arch": ARCH_AIE2PS,
     "base": AIE_DEV_TEL,
     "hwGen": 5,
diff --git a/src/mldebug/backend/core_dump_impl.py b/src/mldebug/backend/core_dump_impl.py
index ed409ea..0825d20 100644
--- a/src/mldebug/backend/core_dump_impl.py
+++ b/src/mldebug/backend/core_dump_impl.py
@@ -117,7 +117,7 @@ def peek_device(filename):
       "<BBBBBB", header[8:14]
     )
 
-    # Match hwGen, disambiguating same-hwGen variants (e.g. telluride/t50)
+    # Match hwGen, disambiguating same-hwGen variants (e.g. telluride/t20)
     # by the header's total rows/cols.
     detected = resolve_variant(hw_gen, total_rows, total_cols)
 
diff --git a/src/mldebug/input_parser.py b/src/mldebug/input_parser.py
index f94d612..c7c4c90 100644
--- a/src/mldebug/input_parser.py
+++ b/src/mldebug/input_parser.py
@@ -110,7 +110,7 @@ def create_run_flags(args, subgraph_path: str, fsp: str, fsp_execution_order: li
 
   # AIE interface for aie2p and aie2 are shared
   # We need to differentiate between them for a few items.
-  # sub_device carries any same-hwGen variant (e.g. t50); it drives the arch
+  # sub_device carries any same-hwGen variant (e.g. t20); it drives the arch
   # module and geometry, while args.device stays the base device for backends.
   sub_device = getattr(args, "sub_device", None) or args.device
   args.sub_device = sub_device
@@ -239,7 +239,7 @@ def _detect_vaiml_variant(aie_dir):
 
   Primary source is aie_trace_config.json's driver_config (hw_gen plus full
   device geometry), which lets us distinguish same-hwGen variants (e.g.
-  telluride vs t50). Falls back to the HW_GEN define in aie_control.cpp when
+  telluride vs t20). Falls back to the HW_GEN define in aie_control.cpp when
   the trace config is missing. Both files live in <work>/ps/c_rts/.
 
   Returns a device/variant name, or None when nothing could be detected
@@ -283,7 +283,7 @@ def set_device(args) -> None:
 
   Sets two fields: ``args.device`` (base device understood by the backends,
   binding, and xrt-smi) and ``args.sub_device`` (the resolved variant, e.g.
-  't50', used to select the arch module and geometry). For most devices these
+  't20', used to select the arch module and geometry). For most devices these
   are identical.
 
   Args:
diff --git a/src/mldebug/mldebug_cli.py b/src/mldebug/mldebug_cli.py
index f3bcde0..6dd79f4 100644
--- a/src/mldebug/mldebug_cli.py
+++ b/src/mldebug/mldebug_cli.py
@@ -19,7 +19,7 @@
   AIE_DEV_PHX,
   AIE_DEV_STX,
   AIE_DEV_T10,
-  AIE_DEV_T50,
+  AIE_DEV_T20,
   AIE_DEV_TEL,
 )
 from mldebug.client_debug import ClientDebug
@@ -251,7 +251,7 @@ def app():
     "-d",
     "--device",
     help="Specify device if it can't be detected from aie_dir.",
-    choices=[AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3, AIE_DEV_T50, AIE_DEV_T10],
+    choices=[AIE_DEV_PHX, AIE_DEV_STX, AIE_DEV_TEL, AIE_DEV_NPU3, AIE_DEV_T20, AIE_DEV_T10],
     required=False,
   )
   # Hidden Argument

From d9534f1110f2fa953ea5699296fe28bc8196b645 Mon Sep 17 00:00:00 2001
From: anurag <anurag@amd.com>
Date: Wed, 29 Jul 2026 11:19:36 -0600
Subject: [PATCH 3/4] add t20 elfs

Signed-off-by: anurag <anurag@amd.com>
---
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt1x4x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x2x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x4x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt3x4x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt4x4x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt5x4x4.elf | 3 +++
 src/mldebug/bin/initial_halt_elfs/t20/aieHalt6x4x4.elf | 3 +++
 7 files changed, 21 insertions(+)
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt1x4x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x2x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x4x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt3x4x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt4x4x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt5x4x4.elf
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t20/aieHalt6x4x4.elf

diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt1x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt1x4x4.elf
new file mode 100644
index 0000000..741a6c6
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt1x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a171ba0891a912e96616548157d6cd9a10b0798819001ab19ed258e9416c1090
+size 11648
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x2x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x2x4.elf
new file mode 100644
index 0000000..571fa5e
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x2x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d38cdcfacfb53dda49188c3ab847349bf2032cec3e903c2484bd04e5d448ec70
+size 11648
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x4x4.elf
new file mode 100644
index 0000000..5e8fef0
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt2x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abea2adc36bd6a610809fcf320a82d1032ceef33bebec9a56d2f08405c671d3c
+size 14224
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt3x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt3x4x4.elf
new file mode 100644
index 0000000..99bd0c3
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt3x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a70cfbfab21ef0da2ee64902fc406edc44dead493a613545397a4a473fa55057
+size 16816
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt4x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt4x4x4.elf
new file mode 100644
index 0000000..d42aad4
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt4x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa2a7daab9f7363c68dcbcb458b2604704badc8585838fcf0404ae951ed6982c
+size 19408
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt5x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt5x4x4.elf
new file mode 100644
index 0000000..b280b80
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt5x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0178bbccc6949d2c0470e19302afdad07f462ccc9912451a22a811ee600005d
+size 22016
diff --git a/src/mldebug/bin/initial_halt_elfs/t20/aieHalt6x4x4.elf b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt6x4x4.elf
new file mode 100644
index 0000000..9ce0dc6
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t20/aieHalt6x4x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fda1a607c2e72ad6fbb9cab0ef7831d1644982e5b57ea7f655e30be3fb04a260
+size 24624

From 063ea02bfdfdd7464345b7ef089fb9ff45ed9eaf Mon Sep 17 00:00:00 2001
From: anurag <anurag@amd.com>
Date: Wed, 29 Jul 2026 11:33:50 -0600
Subject: [PATCH 4/4] add t10 elfs

Signed-off-by: anurag <anurag@amd.com>
---
 src/mldebug/bin/initial_halt_elfs/t10/aieHalt2x2x4.elf | 3 +++
 1 file changed, 3 insertions(+)
 create mode 100644 src/mldebug/bin/initial_halt_elfs/t10/aieHalt2x2x4.elf

diff --git a/src/mldebug/bin/initial_halt_elfs/t10/aieHalt2x2x4.elf b/src/mldebug/bin/initial_halt_elfs/t10/aieHalt2x2x4.elf
new file mode 100644
index 0000000..571fa5e
--- /dev/null
+++ b/src/mldebug/bin/initial_halt_elfs/t10/aieHalt2x2x4.elf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d38cdcfacfb53dda49188c3ab847349bf2032cec3e903c2484bd04e5d448ec70
+size 11648