diff --git a/openpilot/selfdrive/modeld/SConscript b/openpilot/selfdrive/modeld/SConscript
index d9af91af4aa5c7..601fbd9a1fb980 100644
--- a/openpilot/selfdrive/modeld/SConscript
+++ b/openpilot/selfdrive/modeld/SConscript
@@ -28,7 +28,7 @@ def estimate_pickle_max_size(onnx_size):
 
 if arch == 'larch64':
   tg_backend = 'QCOM'
-  tg_flags = f'DEV={tg_backend} IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
+  tg_flags = f'DEV={tg_backend}:IR3 IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
 else:
   tg_backend = 'CPU'
   tg_flags = f'DEV=CPU' if arch == 'Darwin' else 'DEV=CPU:LLVM'
diff --git a/openpilot/selfdrive/modeld/modeld.py b/openpilot/selfdrive/modeld/modeld.py
index f1a406b64ab1e6..9828429db9d19f 100755
--- a/openpilot/selfdrive/modeld/modeld.py
+++ b/openpilot/selfdrive/modeld/modeld.py
@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 import os
 os.environ['GMMU'] = '0' # for usbgpu fast loading, noop for qcom
+os.environ['DEV'] = 'QCOM:IR3' # tmp: we shouldn't be compiling anything at runtime
 from tinygrad.tensor import Tensor
 import threading
 import time
diff --git a/tinygrad_repo b/tinygrad_repo
index ef37830d138838..550225f603c4f7 160000
--- a/tinygrad_repo
+++ b/tinygrad_repo
@@ -1 +1 @@
-Subproject commit ef37830d138838933d6d70d0024de9f130cb2308
+Subproject commit 550225f603c4f7cee18dc43deccc45cba09af527