From b10afc8379c3dbcb28836d2afe814951d1ce22f1 Mon Sep 17 00:00:00 2001
From: firestar5683 <168790843+firestar5683@users.noreply.github.com>
Date: Sun, 16 Aug 2026 12:53:04 -0500
Subject: [PATCH] Fix external GPU model handoff

---
 selfdrive/modeld/modeld.py                    | 25 ++++++++++++++++++-
 selfdrive/modeld/tests/test_usbgpu_helpers.py | 19 ++++++++++++++
 2 files changed, 43 insertions(+), 1 deletion(-)

diff --git a/selfdrive/modeld/modeld.py b/selfdrive/modeld/modeld.py
index 68ba3d6bc..3df17ec62 100755
--- a/selfdrive/modeld/modeld.py
+++ b/selfdrive/modeld/modeld.py
@@ -143,6 +143,22 @@ def _select_builtin_model(params: Params) -> None:
   params.put("DrivingModelName", "Regret Driven Framework V4")
 
 
+def _close_tinygrad_disk_cache_connection() -> None:
+  """Drop a tinygrad cache connection before handing work to another thread."""
+  import tinygrad.helpers as tinygrad_helpers
+
+  connection = getattr(tinygrad_helpers, "_db_connection", None)
+  if connection is None:
+    return
+
+  try:
+    connection.close()
+  except Exception:
+    cloudlog.exception("failed to close tinygrad disk cache connection")
+  finally:
+    tinygrad_helpers._db_connection = None
+
+
 def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
                           lat_action_t: float, long_action_t: float, v_ego: float, mlsim: bool,
                           is_v9: bool, is_v14: bool, is_v15: bool, starpilot_toggles,
@@ -593,15 +609,22 @@ def main(demo=False):
 
     def load_big_model() -> None:
       nonlocal big_model
+      candidate = None
       try:
         wait_usbgpu_link()
         candidate = ModelState(vipc_client_main.width, vipc_client_main.height, True)
         if not candidate.uses_external_gpu:
           raise RuntimeError("external GPU model resolved to the builtin model")
         candidate.warmup()
-        big_model = candidate
       except Exception:
         cloudlog.exception("external GPU model load or warmup failed")
+        candidate = None
+      finally:
+        # tinygrad's global SQLite cache connection is thread-bound. Loading and
+        # warming here can create it in this worker, so close it here before the
+        # model (or native fallback) runs on modeld's main thread.
+        _close_tinygrad_disk_cache_connection()
+      big_model = candidate
 
     loader = threading.Thread(target=load_big_model, name="big_model_loader", daemon=True)
     loader.start()
diff --git a/selfdrive/modeld/tests/test_usbgpu_helpers.py b/selfdrive/modeld/tests/test_usbgpu_helpers.py
index ab4128abd..04808d610 100644
--- a/selfdrive/modeld/tests/test_usbgpu_helpers.py
+++ b/selfdrive/modeld/tests/test_usbgpu_helpers.py
@@ -19,6 +19,25 @@ def test_external_gpu_uses_a_longer_load_watchdog():
   assert modeld.BIG_MODEL_RUN_WAIT_TIMEOUT_MS == 3000
 
 
+def test_tinygrad_disk_cache_connection_is_closed_before_thread_handoff(monkeypatch):
+  import tinygrad.helpers as tinygrad_helpers
+
+  class FakeConnection:
+    def __init__(self):
+      self.closed = False
+
+    def close(self):
+      self.closed = True
+
+  connection = FakeConnection()
+  monkeypatch.setattr(tinygrad_helpers, "_db_connection", connection)
+
+  modeld._close_tinygrad_disk_cache_connection()
+
+  assert connection.closed
+  assert tinygrad_helpers._db_connection is None
+
+
 def test_out_of_band_artifact_round_trip():
   artifact = {"weights": np.arange(32, dtype=np.float32), "metadata": {"version": 1}}
   stream = io.BytesIO()