onnxruntime: initialize CUDA provider before predictions (#24330)

* initialize CUDA runtime before predictions

* dmond pauses modeld execution...

* done in other PR

selfdrive/modeld/runners/onnx_runner.py

@@ -25,6 +25,11 @@ def write(d):
 def run_loop(m):
   ishapes = [[1]+ii.shape[1:] for ii in m.get_inputs()]
   keys = [x.name for x in m.get_inputs()]
+
+  # run once to initialize CUDA provider
+  if "CUDAExecutionProvider" in m.get_providers():
+    m.run(None, dict(zip(keys, [np.zeros(shp, dtype=np.float32) for shp in ishapes])))
+
   print("ready to run onnx model", keys, ishapes, file=sys.stderr)
   while 1:
     inputs = []

selfdrive/test/process_replay/regen.py

@@ -224,6 +224,11 @@ def regen_segment(lr, frs=None, outdir=FAKEDATA):
   }
 
   try:
+    # TODO: make first run of onnxruntime CUDA provider fast
+    managed_processes["modeld"].start()
+    managed_processes["dmonitoringmodeld"].start()
+    time.sleep(5)
+
     # start procs up
     ignore = list(fake_daemons.keys()) + ['ui', 'manage_athenad', 'uploader']
     ensure_running(managed_processes.values(), started=True, not_run=ignore)