close

commit 791d63cd81
Author: royjr <royjr96@gmail.com>
Date:   Sat Jul 25 03:03:28 2026 -0400

    Update host.py

commit f166bfcdc5
Author: royjr <royjr96@gmail.com>
Date:   Sat Jul 25 02:53:13 2026 -0400

    Update process_config.py

commit 418c8154be
Author: royjr <royjr96@gmail.com>
Date:   Sat Jul 25 02:41:08 2026 -0400

    wgpu

openpilot/cereal/log.capnp

@@ -132,7 +132,6 @@ struct OnroadEvent @0xc4fa6047f024e718 {
     userBookmark @95;
     excessiveActuation @96;
     audioFeedback @97;
-
     soundsUnavailableDEPRECATED @47;
   }
 }

openpilot/common/params_keys.h

@@ -134,6 +134,9 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"UsbGpuPresent", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
     {"UsbGpuCompiled", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
     {"Version", {PERSISTENT, STRING}},
+    {"WgpuEnabled", {CLEAR_ON_MANAGER_START | DEVELOPMENT_ONLY, BOOL}},
+    {"WgpuModelName", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION | DEVELOPMENT_ONLY, STRING}},
+    {"WgpuReady", {CLEAR_ON_MANAGER_START | DEVELOPMENT_ONLY, BOOL}},
 
     // --- sunnypilot params --- //
     {"ApiCache_DriveStats", {PERSISTENT, JSON}},

openpilot/selfdrive/modeld/SConscript

@@ -37,6 +37,9 @@ available = probe_devices()
 if 'CUDA' in available:
   tg_backend = 'CUDA'
   tg_flags = f'DEV={tg_backend}'
+elif 'METAL' in available:
+  tg_backend = 'METAL'
+  tg_flags = f'DEV={tg_backend} FLOAT16=1'
 elif 'QCOM' in available:
   tg_backend = 'QCOM'
   tg_flags = f'DEV={tg_backend} IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
@@ -55,6 +58,7 @@ tg_devices = { # which device to put jit inputs to at runtime
 }
 
 USBGPU = usbgpu_present() # or release # TODO always build big model on release
+WGPU = os.getenv('WGPU') == '1'
 if USBGPU:
   usbgpu_tg_flags = f'DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV={tg_backend} FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0'
   # the USB+AMD GPU takes an exclusive flock; serialize all targets that touch it
@@ -84,13 +88,13 @@ compile_modeld_script = [
 model_w, model_h = MEDMODEL_INPUT_SIZE
 frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
 
-for usbgpu in [False, True] if USBGPU else [False]:
+for usbgpu in [False, True] if USBGPU or WGPU else [False]:
   target_pkl_path = File(modeld_pkl_path(usbgpu)).abspath
   # BIG_INTO_SMALL=1 builds the default target from the big model, e.g. to test it without a USB GPU
-  file_prefix, cmd_flags = ('big_', usbgpu_tg_flags) if usbgpu else ('big_' if os.getenv('BIG_INTO_SMALL') else '', tg_flags)
+  file_prefix, cmd_flags = ('big_', usbgpu_tg_flags if USBGPU else tg_flags) if usbgpu else ('big_' if os.getenv('BIG_INTO_SMALL') else '', tg_flags)
   driving_onnx_deps = get_existing_chunks(File(f"models/{file_prefix}driving_supercombo.onnx").abspath)
   camera_res_args = ' '.join(f'{cw}x{ch}' for cw, ch in CAMERA_CONFIGS)
-  cmd = (f'{cmd_flags} {mac_brew_string} python3 {modeld_dir}/compile_modeld.py '
+  cmd = (f'{cmd_flags} {mac_brew_string} {sys.executable} {modeld_dir}/compile_modeld.py '
         f'--model-size {model_w}x{model_h} '
         f'--camera-resolutions {camera_res_args} '
         f'--onnx {File(f"models/{file_prefix}driving_supercombo.onnx").abspath} '
@@ -104,7 +108,7 @@ for usbgpu in [False, True] if USBGPU else [False]:
     tinygrad_files + compile_modeld_script + driving_onnx_deps + [Value(chunk_targets), chunker_file],
     [cmd, Action(do_chunk, "  [CHUNK] $TARGET")],
   )
-  if usbgpu:
+  if usbgpu and USBGPU:
     lenv.SideEffect(usbgpu_lock, node)
 
 # get model metadata

openpilot/selfdrive/modeld/modeld.py

@@ -26,6 +26,7 @@ from openpilot.common.file_chunker import open_file_chunked, get_manifest_path
 from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
 from openpilot.selfdrive.modeld.helpers import usbgpu_present, modeld_pkl_path, get_tg_input_devices, load_oob
 from openpilot.selfdrive.modeld.usbgpu_link import wait_usbgpu_link
+from openpilot.tools.wgpu.zmq import WGPU_CAR_PARAMS, ZmqPubMaster, ZmqSubMaster, ZmqSubSocket
 
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
@@ -76,26 +77,44 @@ class FrameMeta:
       self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
 
 
+def copy_nv12_to_venus(buf: VisionBuf, dst: np.ndarray, nv12: tuple[int, int, int, int]) -> None:
+  stride, y_height, uv_height, _ = nv12
+  if buf.stride < buf.width:
+    raise ValueError(f"invalid VisionIPC stride {buf.stride} for width {buf.width}")
+
+  src = np.frombuffer(buf.data, dtype=np.uint8)
+  src_size = buf.uv_offset + buf.stride * (buf.height // 2)
+  if src.size < src_size:
+    raise ValueError(f"VisionIPC buffer has {src.size} bytes, expected at least {src_size}")
+
+  dst[:stride * y_height].reshape(y_height, stride)[:buf.height, :buf.width] = \
+    src[:buf.stride * buf.height].reshape(buf.height, buf.stride)[:, :buf.width]
+  dst[stride * y_height:stride * (y_height + uv_height)].reshape(uv_height, stride)[:buf.height // 2, :buf.width] = \
+    src[buf.uv_offset:src_size].reshape(buf.height // 2, buf.stride)[:, :buf.width]
+
+
 class ModelState(ModelStateBase):
   prev_desire: np.ndarray  # for tracking the rising edge of the pulse
 
-  def __init__(self, cam_w: int, cam_h: int, usbgpu: bool):
+  def __init__(self, cam_w: int, cam_h: int, usbgpu: bool, big_model: bool = False):
     ModelStateBase.__init__(self)
     self.LAT_SMOOTH_SECONDS = LAT_SMOOTH_SECONDS
     input_devices = get_tg_input_devices(PROCESS_NAME, usbgpu)
     self.WARP_DEV, self.QUEUE_DEV = input_devices['WARP_DEV'], input_devices['QUEUE_DEV']
-    jits = load_oob(open_file_chunked(modeld_pkl_path(usbgpu)))
+    jits = load_oob(open_file_chunked(modeld_pkl_path(usbgpu or big_model)))
     metadata = jits['metadata']
     self.input_shapes = metadata['input_shapes']
     self.vision_input_names = [k for k in self.input_shapes if 'img' in k]
     self.output_slices = metadata['output_slices']
 
     self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
+    self.copy_vision_buffers = self.WARP_DEV.split(":")[0] == "METAL"
 
     self.frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
     self.input_queues, self.npy = make_input_queues(self.input_shapes, self.frame_skip, device=self.QUEUE_DEV)
     self.full_frames: dict[str, Tensor] = {}
     self._blob_cache: dict[tuple[str, int], Tensor] = {}
+    self._vision_staging: dict[str, np.ndarray] = {}
     self.parser = Parser()
     self.frame_buf_params = {k: get_nv12_info(cam_w, cam_h) for k in ('img', 'big_img')}
     self.run_policy = jits['run_policy']
@@ -108,13 +127,22 @@ class ModelState(ModelStateBase):
   def run(self, bufs: dict[str, VisionBuf], transforms: dict[str, np.ndarray],
           inputs: dict[str, np.ndarray]) -> dict[str, np.ndarray] | None:
     for key in bufs.keys():
-      ptr = np.frombuffer(bufs[key].data, dtype=np.uint8).ctypes.data
-      yuv_size = self.frame_buf_params[key][3]
-      # There is a ringbuffer of imgs, just cache tensors pointing to all of them
-      cache_key = (key, ptr)
-      if cache_key not in self._blob_cache:
-        self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8', device=self.WARP_DEV)
-      self.full_frames[key] = self._blob_cache[cache_key]
+      nv12 = self.frame_buf_params[key]
+      yuv_size = nv12[3]
+      if self.copy_vision_buffers:
+        # VisionIPC supplies a CPU pointer. Metal's external_ptr expects an MTLBuffer object,
+        # so wrap its NV12 pixels in the Venus layout expected by the compiled warp, then copy.
+        if key not in self._vision_staging:
+          self._vision_staging[key] = np.zeros(yuv_size, dtype=np.uint8)
+        copy_nv12_to_venus(bufs[key], self._vision_staging[key], nv12)
+        self.full_frames[key] = Tensor(self._vision_staging[key], device=self.WARP_DEV).realize()
+      else:
+        # There is a ringbuffer of imgs, just cache tensors pointing to all of them.
+        frame = np.frombuffer(bufs[key].data, dtype=np.uint8, count=yuv_size)
+        cache_key = (key, frame.ctypes.data)
+        if cache_key not in self._blob_cache:
+          self._blob_cache[cache_key] = Tensor.from_blob(frame.ctypes.data, (yuv_size,), dtype='uint8', device=self.WARP_DEV)
+        self.full_frames[key] = self._blob_cache[cache_key]
 
     # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
     inputs['desire_pulse'][0] = 0
@@ -139,18 +167,32 @@ class ModelState(ModelStateBase):
     return outputs_dict
 
 
-def main(demo=False):
+def main(demo=False, remote_addr: str | None = None, big_model: bool = False):
   cloudlog.warning("modeld init")
 
   _present = usbgpu_present()
   _compiled = os.path.isfile(get_manifest_path(modeld_pkl_path(usbgpu=True)))
   USBGPU = _present and _compiled
+  if big_model and not _compiled:
+    raise FileNotFoundError(f"big model is not compiled: {modeld_pkl_path(usbgpu=True)}")
   params = Params()
   params.put_bool("UsbGpuPresent", _present)
   params.put_bool("UsbGpuCompiled", _compiled)
 
   config_realtime_process(7, 54)
 
+  remote_CP = None
+  if remote_addr is not None:
+    # Do not attach to VisionIPC until all startup prerequisites are available.
+    # Otherwise its notification queue grows while waiting for the infrequent
+    # bridged carParams message and reconnect starts tens of seconds behind.
+    cloudlog.warning("waiting for remote carParams")
+    car_params_socket = ZmqSubSocket(WGPU_CAR_PARAMS, remote_addr, conflate=True)
+    raw_car_params = car_params_socket.receive()
+    assert raw_car_params is not None
+    remote_CP = messaging.log_from_bytes(raw_car_params, car.CarParams)
+    cloudlog.info("modeld got remote CarParams: %s", remote_CP.brand)
+
   # visionipc clients
   while True:
     available_streams = VisionIpcClient.available_streams("camerad", block=False)
@@ -178,12 +220,14 @@ def main(demo=False):
     wait_usbgpu_link()
   st = time.monotonic()
   cloudlog.warning("loading model")
-  model = ModelState(vipc_client_main.width, vipc_client_main.height, USBGPU)
+  model = ModelState(vipc_client_main.width, vipc_client_main.height, USBGPU, big_model=big_model)
   cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
 
   # messaging
-  pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"])
-  sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"])
+  output_services = ["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"]
+  pm = ZmqPubMaster(output_services) if remote_addr is not None else PubMaster(output_services)
+  services = ["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"]
+  sm = ZmqSubMaster(services, remote_addr) if remote_addr is not None else SubMaster(services)
 
   publish_state = PublishState()
   params = Params()
@@ -203,6 +247,9 @@ def main(demo=False):
 
   if demo:
     CP = get_demo_car_params()
+  elif remote_addr is not None:
+    assert remote_CP is not None
+    CP = remote_CP
   else:
     CP = messaging.log_from_bytes(params.get("CarParams", block=True), car.CarParams)
   cloudlog.info("modeld got CarParams: %s", CP.brand)
@@ -332,7 +379,9 @@ if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser()
     parser.add_argument('--demo', action='store_true', help='A boolean for demo mode.')
+    parser.add_argument('--remote', metavar='ADDRESS', help='Run against a remote device over the cereal ZMQ bridge.')
+    parser.add_argument('--big-model', action='store_true', help='Use the locally compiled big driving model.')
     args = parser.parse_args()
-    main(demo=args.demo)
+    main(demo=args.demo, remote_addr=args.remote, big_model=args.big_model)
   except KeyboardInterrupt:
     cloudlog.warning("got SIGINT")

openpilot/selfdrive/selfdrived/events.py

@@ -722,7 +722,6 @@ EVENTS: dict[int, dict[str, Alert | AlertCallbackType]] = {
     ET.NO_ENTRY: NoEntryAlert("Driving Model Lagging"),
     ET.PERMANENT: modeld_lagging_alert,
   },
-
   # Besides predicting the path, lane lines and lead car data the model also
   # predicts the current velocity and rotation speed of the car. If the model is
   # very uncertain about the current velocity while the car is moving, this

openpilot/selfdrive/selfdrived/selfdrived.py

@@ -111,6 +111,7 @@ class SelfdriveD(CruiseHelper):
     self.is_metric = self.params.get_bool("IsMetric")
     self.is_ldw_enabled = self.params.get_bool("IsLdwEnabled")
     self.disengage_on_accelerator = self.params.get_bool("DisengageOnAccelerator")
+    self.wgpu_enabled = self.params.get_bool("WgpuEnabled")
 
     car_recognized = self.CP.brand != 'mock'
 
@@ -398,12 +399,12 @@ class SelfdriveD(CruiseHelper):
     has_disable_events = self.events.contains(ET.NO_ENTRY) and (self.events.contains(ET.SOFT_DISABLE) or self.events.contains(ET.IMMEDIATE_DISABLE))
     no_system_errors = (not has_disable_events) or (len(self.events) == num_events)
     if not self.sm.all_checks() and no_system_errors:
-      if not self.sm.all_alive():
-        self.events.add(EventName.commIssue)
-      elif not self.sm.all_freq_ok():
-        self.events.add(EventName.commIssueAvgFreq)
-      else:
-        self.events.add(EventName.commIssue)
+      # if not self.sm.all_alive():
+      #   self.events.add(EventName.commIssue)
+      # elif not self.sm.all_freq_ok():
+      #   self.events.add(EventName.commIssueAvgFreq)
+      # else:
+      #   self.events.add(EventName.commIssue)
 
       logs = {
         'invalid': [s for s, valid in self.sm.valid.items() if not valid],
@@ -416,13 +417,13 @@ class SelfdriveD(CruiseHelper):
     else:
       self.logged_comm_issue = None
 
-    if not self.CP.notCar:
-      if not self.sm['livePose'].posenetOK:
-        self.events.add(EventName.posenetInvalid)
-      if not self.sm['livePose'].inputsOK:
-        self.events.add(EventName.locationdTemporaryError)
-      if not self.sm['liveParameters'].valid and cal_status == log.LiveCalibrationData.Status.calibrated and not TESTING_CLOSET and (not SIMULATION or REPLAY):
-        self.events.add(EventName.paramsdTemporaryError)
+    # if not self.CP.notCar:
+    #   if not self.sm['livePose'].posenetOK:
+    #     self.events.add(EventName.posenetInvalid)
+    #   if not self.sm['livePose'].inputsOK:
+    #     self.events.add(EventName.locationdTemporaryError)
+    #   if not self.sm['liveParameters'].valid and cal_status == log.LiveCalibrationData.Status.calibrated and not TESTING_CLOSET and (not SIMULATION or REPLAY):
+    #     self.events.add(EventName.paramsdTemporaryError)
 
     # conservative HW alert. if the data or frequency are off, locationd will throw an error
     if any((self.sm.frame - self.sm.recv_frame[s])*DT_CTRL > 10. for s in self.sensor_packets):
@@ -467,9 +468,9 @@ class SelfdriveD(CruiseHelper):
     self.distance_traveled += abs(CS.vEgo) * DT_CTRL
 
     # TODO: fix simulator
-    if not SIMULATION or REPLAY:
-      if self.sm['modelV2'].frameDropPerc > 1:
-        self.events.add(EventName.modeldLagging)
+    # if not SIMULATION or REPLAY:
+    #   if self.sm['modelV2'].frameDropPerc > 1 and not self.wgpu_enabled:
+    #     self.events.add(EventName.modeldLagging)
 
     # mute canBusMissing event if in Park, as it sometimes may trigger a false alarm with MADS in Paused state
     if CS.gearShifter == car.CarState.GearShifter.park and self.mads.enabled:
@@ -625,6 +626,7 @@ class SelfdriveD(CruiseHelper):
       self.is_metric = self.params.get_bool("IsMetric")
       self.is_ldw_enabled = self.params.get_bool("IsLdwEnabled")
       self.disengage_on_accelerator = self.params.get_bool("DisengageOnAccelerator")
+      self.wgpu_enabled = self.params.get_bool("WgpuEnabled")
       self.experimental_mode = self.params.get_bool("ExperimentalMode") and self.CP.openpilotLongitudinalControl
       self.personality = self.params.get("LongitudinalPersonality", return_default=True)
 

openpilot/selfdrive/ui/mici/layouts/home.py

@@ -248,8 +248,10 @@ class MiciHomeLayout(Widget):
 
     # ***** Center-aligned bottom section icons *****
     self._experimental_icon.set_visible(ui_state.experimental_mode)
-    self._egpu_icon.set_visible(ui_state.usbgpu and ui_state.usbgpu_compiled)
-    self._egpu_icon_gray.set_visible(ui_state.usbgpu and not ui_state.usbgpu_compiled)
+    wgpu_running = ui_state.wgpu_enabled and ui_state.sm.alive["modelV2"] and ui_state.sm.valid["modelV2"]
+    self._egpu_icon.set_visible((ui_state.usbgpu and ui_state.usbgpu_compiled) or wgpu_running)
+    self._egpu_icon_gray.set_visible((ui_state.usbgpu and not ui_state.usbgpu_compiled) or
+                                     (ui_state.wgpu_ready and not wgpu_running))
     self._mic_icon.set_visible(ui_state.recording_audio)
     self._body_icon.set_visible(bool(ui_state.is_body))
 

openpilot/selfdrive/ui/mici/onroad/augmented_road_view.py

@@ -1,3 +1,5 @@
+import time
+
 import numpy as np
 import pyray as rl
 from openpilot.cereal import log
@@ -160,6 +162,20 @@ class AugmentedRoadView(CameraView):
                                        text_color=rl.Color(255, 255, 255, int(255 * 0.9)),
                                        alignment=rl.GuiTextAlignment.TEXT_ALIGN_CENTER,
                                        alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE)
+    self._model_status_key_labels = [
+      UnifiedLabel("", 21, FontWeight.ROMAN, text_color=rl.Color(210, 210, 210, 220),
+                   alignment=rl.GuiTextAlignment.TEXT_ALIGN_LEFT,
+                   alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE,
+                   wrap_text=False)
+      for _ in range(6)
+    ]
+    self._model_status_value_labels = [
+      UnifiedLabel("", 22, FontWeight.SEMI_BOLD,
+                   alignment=rl.GuiTextAlignment.TEXT_ALIGN_RIGHT,
+                   alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE,
+                   wrap_text=False)
+      for _ in range(6)
+    ]
 
     self._fade_texture = gui_app.texture("icons_mici/onroad/onroad_fade.png")
 
@@ -230,6 +246,8 @@ class AugmentedRoadView(CameraView):
     self._driver_state_renderer.set_position(self._rect.x + 16, self._rect.y + 10)
     self._driver_state_renderer.render()
 
+    self._render_model_status()
+
     self._hud_renderer.set_can_draw_top_icons(alert_to_render is None)
     self._hud_renderer.set_wheel_critical_icon(alert_to_render is not None and not not_animating_out and
                                                alert_to_render.visual_alert == car.CarControl.HUDControl.VisualAlert.steerRequired)
@@ -248,6 +266,58 @@ class AugmentedRoadView(CameraView):
 
     self._bookmark_icon.render(self.rect)
 
+  def _render_model_status(self):
+    model = ui_state.sm["modelV2"]
+    if ui_state.sm.seen["modelV2"] and model.timestampEof:
+      model_age_ms = max(0., (time.monotonic_ns() - model.timestampEof) / 1e6)
+      execution_ms = max(0., model.modelExecutionTime * 1e3)
+      io_queue_ms = max(0., model_age_ms - execution_ms)
+      frame_drop = model.frameDropPerc
+      age_text = f"{model_age_ms:.0f} ms"
+      execution_text = f"{execution_ms:.0f} ms"
+      io_queue_text = f"{io_queue_ms:.0f} ms"
+      frame_drop_text = f"{frame_drop:.1f}%"
+    else:
+      model_age_ms = float("inf")
+      age_text = execution_text = io_queue_text = "-- ms"
+      frame_drop_text = "--%"
+
+    if ui_state.wgpu_enabled:
+      source, model_name = "WGPU", ui_state.wgpu_model_name
+    else:
+      source, model_name = "LOCAL", "DEVICE"
+
+    if model_age_ms < 200:
+      color = rl.Color(100, 255, 120, 230)
+    elif model_age_ms < 500:
+      color = rl.Color(255, 210, 80, 230)
+    else:
+      color = rl.Color(255, 120, 80, 230)
+
+    panel_w, panel_h = 250, 178
+    status_rect = rl.Rectangle(self._content_rect.x + self._content_rect.width - panel_w - 18,
+                               self._content_rect.y + 18, panel_w, panel_h)
+    rl.draw_rectangle_rounded(status_rect, 0.14, 8, rl.Color(0, 0, 0, 175))
+
+    rows = (
+      ("SOURCE", source),
+      ("MODEL", model_name),
+      ("AGE", age_text),
+      ("EXEC", execution_text),
+      ("IO/QUEUE", io_queue_text),
+      ("DROPPED", frame_drop_text),
+    )
+    row_h = 26
+    for i, ((key, value), key_label, value_label) in enumerate(
+      zip(rows, self._model_status_key_labels, self._model_status_value_labels, strict=True)
+    ):
+      row_rect = rl.Rectangle(status_rect.x + 14, status_rect.y + 11 + i * row_h, status_rect.width - 28, row_h)
+      key_label.set_text(key)
+      key_label.render(rl.Rectangle(row_rect.x, row_rect.y, 105, row_rect.height))
+      value_label.set_text(value)
+      value_label.set_text_color(color)
+      value_label.render(rl.Rectangle(row_rect.x + 108, row_rect.y, row_rect.width - 108, row_rect.height))
+
   def _switch_stream_if_needed(self, sm):
     if sm['selfdriveState'].experimentalMode and WIDE_CAM in self.available_streams:
       v_ego = sm['carState'].vEgo

openpilot/selfdrive/ui/ui_state.py

@@ -82,6 +82,9 @@ class UIState(UIStateSP):
     self.experimental_mode: bool = self.params.get_bool("ExperimentalMode")
     self.usbgpu: bool = self.params.get_bool("UsbGpuPresent")
     self.usbgpu_compiled: bool = self.params.get_bool("UsbGpuCompiled")
+    self.wgpu_enabled: bool = self.params.get_bool("WgpuEnabled")
+    self.wgpu_model_name: str = self.params.get("WgpuModelName") or "UNKNOWN"
+    self.wgpu_ready: bool = self.params.get_bool("WgpuReady")
     self.started: bool = False
     self.ignition: bool = False
     self.recording_audio: bool = False
@@ -213,6 +216,9 @@ class UIState(UIStateSP):
     self.experimental_mode = self.params.get_bool("ExperimentalMode")
     self.usbgpu = self.params.get_bool("UsbGpuPresent")
     self.usbgpu_compiled = self.params.get_bool("UsbGpuCompiled")
+    self.wgpu_enabled = self.params.get_bool("WgpuEnabled")
+    self.wgpu_model_name = self.params.get("WgpuModelName") or "UNKNOWN"
+    self.wgpu_ready = self.params.get_bool("WgpuReady")
 
     UIStateSP.update_params(self)
 

openpilot/system/manager/process_config.py

@@ -95,6 +95,9 @@ def is_stock_model(started, params, CP: car.CarParams) -> bool:
   """Check if the active model runner is stock."""
   return bool(get_active_model_runner(params, not started) == custom.ModelManagerSP.Runner.stock)
 
+def not_wgpu(started: bool, params: Params, CP: car.CarParams) -> bool:
+  return not params.get_bool("WgpuEnabled")
+
 def mapd_ready(started: bool, params: Params, CP: car.CarParams) -> bool:
   return bool(os.path.exists(Paths.mapd_root()))
 
@@ -128,7 +131,7 @@ procs = [
   PythonProcess("micd", "openpilot.system.micd", iscar),
   PythonProcess("timed", "openpilot.system.timed", always_run, enabled=not PC),
 
-  PythonProcess("modeld", "openpilot.selfdrive.modeld.modeld", and_(only_onroad, is_stock_model)),
+  PythonProcess("modeld", "openpilot.selfdrive.modeld.modeld", and_(and_(only_onroad, is_stock_model), not_wgpu)),
   PythonProcess("dmonitoringmodeld", "openpilot.selfdrive.modeld.dmonitoringmodeld", driverview, enabled=(WEBCAM or not PC)),
 
   PythonProcess("sensord", "openpilot.system.sensord.sensord", only_onroad, enabled=not PC),
@@ -177,7 +180,7 @@ procs = [
 procs += [
   # Models
   PythonProcess("models_manager", "openpilot.sunnypilot.models.manager", only_offroad),
-  NativeProcess("modeld_tinygrad", "openpilot/sunnypilot/modeld_v2", ["./modeld"], and_(only_onroad, is_tinygrad_model)),
+  NativeProcess("modeld_tinygrad", "openpilot/sunnypilot/modeld_v2", ["./modeld"], and_(and_(only_onroad, is_tinygrad_model), not_wgpu)),
 
   # Backup
   PythonProcess("backup_manager", "openpilot.sunnypilot.sunnylink.backups.manager", and_(only_offroad, sunnylink_ready_shim)),

openpilot/tools/camerastream/compressed_vipc.py

@@ -1,5 +1,4 @@
 #!/usr/bin/env python3
-import os
 import argparse
 import multiprocessing
 import time
@@ -10,6 +9,7 @@ from collections import deque
 import openpilot.cereal.messaging as messaging
 from msgq.visionipc import VisionIpcServer, VisionStreamType
 from openpilot.tools.camerastream.ffmpeg_decoder import Decoder, FFmpegError
+from openpilot.tools.wgpu.zmq import ZmqSubMaster, ZmqSubSocket
 
 V4L2_BUF_FLAG_KEYFRAME = 8
 
@@ -30,10 +30,7 @@ def decoder(addr, vipc_server, vst, W, H, debug=False):
 
   codec = Decoder("hevc")
 
-  os.environ["ZMQ"] = "1"
-  messaging.reset_context()
-  sock = messaging.sub_sock(sock_name, None, addr=addr, conflate=False)
-  cnt = 0
+  sock = ZmqSubSocket(sock_name, addr)
   last_idx = -1
   seen_iframe = False
 
@@ -46,8 +43,9 @@ def decoder(addr, vipc_server, vst, W, H, debug=False):
     time_q.clear()
 
   while 1:
-    msgs = messaging.drain_sock(sock, wait_for_one=True)
-    for evt in msgs:
+    msgs = sock.drain(wait_for_one=True)
+    for raw in msgs:
+      evt = messaging.log_from_bytes(raw)
       evta = getattr(evt, evt.which())
       if last_idx != -1 and evta.idx.encodeId != (last_idx + 1):
         if debug:
@@ -94,8 +92,9 @@ def decoder(addr, vipc_server, vst, W, H, debug=False):
         continue
 
       frame_start_time = time_q.popleft()
-      vipc_server.send(vst, img_yuv.data, cnt, int(frame_start_time*1e9), int(time.monotonic()*1e9))
-      cnt += 1
+      # Preserve the device camera metadata so remote model outputs line up with
+      # the rest of the device's cereal timeline.
+      vipc_server.send(vst, img_yuv.data, evta.idx.frameId, evta.idx.timestampSof, evta.idx.timestampEof)
 
       pc_latency = (time.monotonic()-frame_start_time)*1000
       if debug:
@@ -105,25 +104,30 @@ def decoder(addr, vipc_server, vst, W, H, debug=False):
 
 class CompressedVipc:
   def __init__(self, addr, vision_streams, server_name, debug=False):
-    print("getting frame sizes")
-    os.environ["ZMQ"] = "1"
-    messaging.reset_context()
-    sm = messaging.SubMaster([ENCODE_SOCKETS[s] for s in vision_streams], addr=addr)
+    print("waiting for remote camera stream metadata", flush=True)
+    sm = ZmqSubMaster([ENCODE_SOCKETS[s] for s in vision_streams], addr)
     while min(sm.recv_frame.values()) == 0:
       sm.update(100)
-    os.environ.pop("ZMQ")
-    messaging.reset_context()
+
+    stream_dimensions = {
+      vst: (sm[ENCODE_SOCKETS[vst]].width, sm[ENCODE_SOCKETS[vst]].height)
+      for vst in vision_streams
+    }
+    # The metadata subscribers are setup-only. Leaving them connected creates a
+    # second unread camera subscription whose TCP queues grow for the entire run.
+    sm.close()
 
     self.vipc_server = VisionIpcServer(server_name)
     for vst in vision_streams:
-      ed = sm[ENCODE_SOCKETS[vst]]
-      self.vipc_server.create_buffers(vst, 4, ed.width, ed.height)
+      width, height = stream_dimensions[vst]
+      self.vipc_server.create_buffers(vst, 4, width, height)
     self.vipc_server.start_listener()
 
     self.procs = []
+    process_context = multiprocessing.get_context("fork")
     for vst in vision_streams:
-      ed = sm[ENCODE_SOCKETS[vst]]
-      p = multiprocessing.Process(target=decoder, args=(addr, self.vipc_server, vst, ed.width, ed.height, debug))
+      width, height = stream_dimensions[vst]
+      p = process_context.Process(target=decoder, args=(addr, self.vipc_server, vst, width, height, debug))
       p.start()
       self.procs.append(p)
 

openpilot/tools/camerastream/ffmpeg_decoder.py

@@ -46,10 +46,23 @@ def _bind(fn, restype, *argtypes):
   return fn
 
 
+def _library_path(name: str, major: int) -> str:
+  candidates = (
+    f"lib{name}.so.{major}",
+    f"lib{name}.{major}.dylib",
+    f"lib{name}.dylib",
+  )
+  for candidate in candidates:
+    path = os.path.join(ffmpeg.LIB_DIR, candidate)
+    if os.path.isfile(path):
+      return path
+  raise FileNotFoundError(f"FFmpeg library not found in {ffmpeg.LIB_DIR}: {', '.join(candidates)}")
+
+
 def _load_libraries():
-  avutil = ctypes.CDLL(os.path.join(ffmpeg.LIB_DIR, "libavutil.so.59"), mode=ctypes.RTLD_GLOBAL)
-  avcodec = ctypes.CDLL(os.path.join(ffmpeg.LIB_DIR, "libavcodec.so.61"), mode=ctypes.RTLD_GLOBAL)
-  swscale = ctypes.CDLL(os.path.join(ffmpeg.LIB_DIR, "libswscale.so.8"), mode=ctypes.RTLD_GLOBAL)
+  avutil = ctypes.CDLL(_library_path("avutil", 59), mode=ctypes.RTLD_GLOBAL)
+  avcodec = ctypes.CDLL(_library_path("avcodec", 61), mode=ctypes.RTLD_GLOBAL)
+  swscale = ctypes.CDLL(_library_path("swscale", 8), mode=ctypes.RTLD_GLOBAL)
 
   c_int, c_char_p, c_void_p, c_size_t = ctypes.c_int, ctypes.c_char_p, ctypes.c_void_p, ctypes.c_size_t
   c_uint8_p = ctypes.POINTER(ctypes.c_uint8)

openpilot/tools/wgpu/README.md

@@ -0,0 +1,74 @@
+# Wireless modeld proof of concept
+
+This runs driving `modeld` on a laptop and returns its cereal outputs to a comma
+device over the existing Wi-Fi network. It reuses the existing HEVC camera
+stream, VisionIPC decoder, and cereal ZMQ bridge.
+
+This is for controlled bench testing only. Wi-Fi has no deterministic latency
+or availability guarantee. The device-side helper switches only after receiving
+a fresh remote model and after manager has stopped the local model publisher.
+It restores local `modeld` if the remote model is missing for one second.
+
+## Build
+
+Use the same commit on the laptop and comma device. Build the cereal bridge on
+the device:
+
+```sh
+scons -u openpilot/cereal/messaging/bridge
+```
+
+Build and test the normal model on the laptop first:
+
+```sh
+PATH="$PWD/.venv/bin:$PATH" scons -u
+```
+
+To compile the big external-GPU model for the laptop's local tinygrad backend:
+
+```sh
+PATH="$PWD/.venv/bin:$PATH" WGPU=1 scons -u \
+  openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl.chunkmanifest
+```
+
+On macOS, the build selects tinygrad's Metal backend when it is available.
+On an 8-GPU-core M5 MacBook Air, the small model's compiled policy pass measured
+about 6–11 ms, while the big model measured about 79–81 ms. The latter already
+misses the 50 ms model cadence before network and codec latency, so start with
+the small model on that class of laptop.
+
+## Run
+
+Find the laptop's LAN IP address that the comma device can reach. The helper can
+start while onroad: it forwards camera/state while local `modeld` remains active,
+then performs an exclusive publisher handoff after the laptop produces a fresh
+valid model:
+
+```sh
+cd /data/openpilot
+python3 -m openpilot.tools.wgpu.device LAPTOP_IP
+```
+
+Keep that terminal open. On the laptop, run:
+
+```sh
+cd /path/to/openpilot
+python3 -m openpilot.tools.wgpu.host COMMA_IP
+```
+
+Add `--big-model` after `COMMA_IP` to use the locally compiled big model.
+
+The device helper sends cached `carParams` over a dedicated startup channel, so
+the host does not attach to the camera streams and accumulate stale frames while
+waiting for the periodic state bridge. Stop either side with Ctrl+C. A host
+disconnect automatically stops remote publication and restores local `modeld`
+after a one-second timeout; the device helper then stays running and waits for
+the next host session. Stop the device helper itself with Ctrl+C before changing
+branches or rebooting.
+
+While WGPU is active, model lag does not create an engagement-blocking alert.
+The fixed-column diagnostics panel at the top-right of the mici onroad UI shows
+the active source (`LOCAL` or `WGPU`), remote model size (`SMALL` or `BIG`),
+total model-frame age, execution time, remaining I/O and queue time, and dropped
+frames. The home GPU icon is gray while the device helper is ready and waiting,
+and green while a valid remote model is active.

openpilot/tools/wgpu/device.py

@@ -0,0 +1,152 @@
+#!/usr/bin/env python3
+import argparse
+import signal
+import subprocess
+import time
+from pathlib import Path
+
+import openpilot.cereal.messaging as messaging
+from openpilot.common.params import Params
+from openpilot.tools.wgpu.zmq import WGPU_CAR_PARAMS, ZmqPubMaster, ZmqSubSocket
+
+
+MODEL_OUTPUTS = "modelV2,drivingModelData,cameraOdometry,modelDataV2SP"
+MODEL_PROCESSES = {"modeld", "modeld_tinygrad"}
+REMOTE_MODEL_TIMEOUT = 1.0
+WGPU_STATUS = "wgpuStatus"
+ROOT = Path(__file__).resolve().parents[3]
+BRIDGE = ROOT / "openpilot/cereal/messaging/bridge"
+
+
+def stop_process(proc: subprocess.Popen) -> None:
+  proc.terminate()
+  try:
+    proc.wait(timeout=5)
+  except subprocess.TimeoutExpired:
+    proc.kill()
+    proc.wait()
+
+
+def handle_sigterm(*_) -> None:
+  raise KeyboardInterrupt
+
+
+def receive_model(sock: ZmqSubSocket) -> tuple[bool, float]:
+  raw = sock.receive(non_blocking=True)
+  if raw is None:
+    return False, float("inf")
+  event = messaging.log_from_bytes(raw)
+  if event.which() != "modelV2" or not event.valid:
+    return True, float("inf")
+  model_age = (time.monotonic_ns() - event.modelV2.timestampEof) / 1e9
+  return True, model_age
+
+
+def receive_model_name(sock: ZmqSubSocket) -> str | None:
+  raw = sock.receive(non_blocking=True)
+  if raw is None:
+    return None
+  model_name = raw.decode(errors="replace").upper()
+  return model_name if model_name in ("SMALL", "BIG") else None
+
+
+def wait_for_local_modeld_stop(timeout: float = 10.) -> None:
+  sm = messaging.SubMaster(["managerState"])
+  deadline = time.monotonic() + timeout
+  while time.monotonic() < deadline:
+    sm.update(100)
+    if sm.seen["managerState"]:
+      running = {p.name for p in sm["managerState"].processes if p.running}
+      if not running.intersection(MODEL_PROCESSES):
+        return
+  raise RuntimeError("timed out waiting for local modeld publisher to stop")
+
+
+def main() -> None:
+  parser = argparse.ArgumentParser(description="Route modeld traffic between this device and a wireless host.")
+  parser.add_argument("host", help="Laptop IP address reachable from this device")
+  args = parser.parse_args()
+
+  if not BRIDGE.is_file():
+    raise FileNotFoundError(f"build the cereal bridge first: {BRIDGE}")
+
+  params = Params()
+  forward: subprocess.Popen | None = None
+  reverse: subprocess.Popen | None = None
+  wgpu_enabled = False
+  try:
+    forward = subprocess.Popen([str(BRIDGE)])
+    params.put_bool("WgpuReady", True, block=True)
+    car_params_pub = ZmqPubMaster([WGPU_CAR_PARAMS])
+    remote_model = ZmqSubSocket("modelV2", args.host, conflate=True)
+    remote_status = ZmqSubSocket(WGPU_STATUS, args.host, conflate=True)
+    car_params = params.get("CarParams") or params.get("CarParamsPersistent")
+
+    while True:
+      # Keep local modeld publishing while a host connects and warms up.
+      print(f"forwarding camera/state to {args.host}; waiting for a fresh remote model")
+      model_name = None
+      while True:
+        if car_params is None:
+          car_params = params.get("CarParams") or params.get("CarParamsPersistent")
+        if car_params is not None:
+          # This dedicated conflated startup channel lets remote modeld obtain CP
+          # before it connects to VisionIPC and accumulates stale frame metadata.
+          car_params_pub.send_raw(WGPU_CAR_PARAMS, car_params)
+        received, model_age = receive_model(remote_model)
+        model_name = receive_model_name(remote_status) or model_name
+        if received and 0 <= model_age < REMOTE_MODEL_TIMEOUT and model_name is not None:
+          break
+        if forward.poll() is not None:
+          raise RuntimeError(f"forward bridge exited with status {forward.returncode}")
+        time.sleep(0.05)
+
+      # Stop the local publisher before attaching the reverse bridge. msgq permits
+      # only one publisher for each model service.
+      params.put("WgpuModelName", model_name, block=True)
+      params.put_bool("WgpuEnabled", True, block=True)
+      wgpu_enabled = True
+      wait_for_local_modeld_stop()
+      reverse = subprocess.Popen([str(BRIDGE), args.host, MODEL_OUTPUTS])
+      print("wgpu active; Ctrl+C or loss of the remote model restores local modeld")
+
+      last_remote_model = time.monotonic()
+      while True:
+        received, _ = receive_model(remote_model)
+        if received:
+          last_remote_model = time.monotonic()
+        if forward.poll() is not None:
+          raise RuntimeError(f"forward bridge exited with status {forward.returncode}")
+        if reverse.poll() is not None:
+          print(f"reverse bridge exited with status {reverse.returncode}; restoring local modeld")
+          break
+        if time.monotonic() - last_remote_model > REMOTE_MODEL_TIMEOUT:
+          print("remote model timed out; restoring local modeld")
+          break
+        time.sleep(0.05)
+
+      stop_process(reverse)
+      reverse = None
+      params.put_bool("WgpuEnabled", False, block=True)
+      wgpu_enabled = False
+      params.remove("WgpuModelName")
+      print("wgpu disabled; local modeld restored; ready for the next host run")
+  finally:
+    # Stop remote publication before allowing the local publisher to restart.
+    if reverse is not None and reverse.poll() is None:
+      stop_process(reverse)
+    if wgpu_enabled:
+      params.put_bool("WgpuEnabled", False, block=True)
+    params.put_bool("WgpuReady", False, block=True)
+    params.remove("WgpuModelName")
+    if forward is not None and forward.poll() is None:
+      stop_process(forward)
+    print("wgpu disabled; local modeld restored")
+
+
+if __name__ == "__main__":
+  signal.signal(signal.SIGTERM, handle_sigterm)
+  try:
+    main()
+  except KeyboardInterrupt:
+    pass

openpilot/tools/wgpu/host.py

@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+import argparse
+import os
+import signal
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+from openpilot.tools.wgpu.zmq import ZmqPubMaster
+
+
+ROOT = Path(__file__).resolve().parents[3]
+CAMERASTREAM = ROOT / "openpilot/tools/camerastream/compressed_vipc.py"
+WGPU_STATUS = "wgpuStatus"
+
+
+def stop_process(proc: subprocess.Popen) -> None:
+  try:
+    os.killpg(proc.pid, signal.SIGTERM)
+  except ProcessLookupError:
+    proc.wait()
+    return
+  try:
+    proc.wait(timeout=5)
+  except subprocess.TimeoutExpired:
+    os.killpg(proc.pid, signal.SIGKILL)
+    proc.wait()
+
+
+def main() -> None:
+  parser = argparse.ArgumentParser(description="Run modeld on this host for a remote comma device.")
+  parser.add_argument("device", help="comma device hostname or IP address")
+  parser.add_argument("--big-model", action="store_true", help="use the locally compiled big driving model")
+  args = parser.parse_args()
+
+  camera = subprocess.Popen([sys.executable, str(CAMERASTREAM), args.device, "--cams", "0,2"], start_new_session=True)
+  model_args = [sys.executable, "-m", "openpilot.selfdrive.modeld.modeld", "--remote", args.device]
+  if args.big_model:
+    model_args.append("--big-model")
+  model = subprocess.Popen(model_args, cwd=ROOT, start_new_session=True)
+  status = ZmqPubMaster([WGPU_STATUS])
+  model_name = b"BIG" if args.big_model else b"SMALL"
+
+  procs = {"camera bridge": camera, "modeld": model}
+  try:
+    while all(proc.poll() is None for proc in procs.values()):
+      status.send_raw(WGPU_STATUS, model_name)
+      time.sleep(0.25)
+    failed_name, failed = next((name, proc) for name, proc in procs.items() if proc.poll() is not None)
+    raise RuntimeError(f"wgpu {failed_name} exited with status {failed.returncode}")
+  finally:
+    for proc in procs.values():
+      if proc.poll() is None:
+        stop_process(proc)
+    print("wgpu host stopped")
+
+
+if __name__ == "__main__":
+  try:
+    main()
+  except KeyboardInterrupt:
+    pass

openpilot/tools/wgpu/zmq.py

@@ -0,0 +1,105 @@
+import time
+
+import zmq
+
+import openpilot.cereal.messaging as messaging
+
+
+WGPU_CAR_PARAMS = "wgpuCarParams"
+
+
+def service_port(endpoint: str) -> int:
+  # Keep this in sync with cereal/messaging/bridge_zmq.cc.
+  value = 0xcbf29ce484222325
+  for char in endpoint.encode():
+    value ^= char
+    value = (value * 0x100000001b3) & 0xffffffffffffffff
+  return 8023 + (value % (65535 - 8023))
+
+
+class ZmqSubSocket:
+  def __init__(self, endpoint: str, address: str, conflate: bool = False):
+    self.context = zmq.Context()
+    self.socket = self.context.socket(zmq.SUB)
+    self.socket.setsockopt(zmq.SUBSCRIBE, b"")
+    self.socket.setsockopt(zmq.RECONNECT_IVL_MAX, 500)
+    if conflate:
+      self.socket.setsockopt(zmq.CONFLATE, 1)
+    self.socket.connect(f"tcp://{address}:{service_port(endpoint)}")
+
+  def receive(self, non_blocking: bool = False) -> bytes | None:
+    try:
+      return self.socket.recv(flags=zmq.NOBLOCK if non_blocking else 0)
+    except zmq.Again:
+      return None
+
+  def drain(self, wait_for_one: bool = False) -> list[bytes]:
+    messages = []
+    if wait_for_one:
+      message = self.receive()
+      if message is not None:
+        messages.append(message)
+    while (message := self.receive(non_blocking=True)) is not None:
+      messages.append(message)
+    return messages
+
+  def close(self) -> None:
+    self.socket.close(linger=0)
+    self.context.term()
+
+
+class ZmqSubMaster:
+  def __init__(self, services: list[str], address: str):
+    self.services = services
+    self.sockets = {service: ZmqSubSocket(service, address, conflate=True) for service in services}
+    self.poller = zmq.Poller()
+    self.socket_to_service = {}
+    for service, sub in self.sockets.items():
+      self.poller.register(sub.socket, zmq.POLLIN)
+      self.socket_to_service[sub.socket] = service
+
+    self.data = {service: getattr(messaging.new_message(service).as_reader(), service) for service in services}
+    self.seen = dict.fromkeys(services, False)
+    self.updated = dict.fromkeys(services, False)
+    self.recv_frame = dict.fromkeys(services, 0)
+    self.frame = -1
+
+  def __getitem__(self, service: str):
+    return self.data[service]
+
+  def update(self, timeout: int = 100) -> None:
+    self.frame += 1
+    self.updated = dict.fromkeys(self.services, False)
+    for socket, _ in self.poller.poll(timeout):
+      service = self.socket_to_service[socket]
+      raw = self.sockets[service].receive(non_blocking=True)
+      if raw is None:
+        continue
+      event = messaging.log_from_bytes(raw)
+      self.data[service] = getattr(event, service)
+      self.seen[service] = True
+      self.updated[service] = True
+      self.recv_frame[service] = self.frame
+
+  def close(self) -> None:
+    for sub in self.sockets.values():
+      self.poller.unregister(sub.socket)
+      sub.close()
+
+
+class ZmqPubMaster:
+  def __init__(self, services: list[str]):
+    context = zmq.Context.instance()
+    self.sockets = {}
+    for service in services:
+      socket = context.socket(zmq.PUB)
+      socket.bind(f"tcp://*:{service_port(service)}")
+      self.sockets[service] = socket
+    # Give already-running bridge subscribers time to finish their handshake.
+    time.sleep(0.1)
+
+  def send(self, service: str, message) -> None:
+    self.sockets[service].send(message.to_bytes(), flags=zmq.NOBLOCK)
+
+  def send_raw(self, service: str, data: bytes) -> None:
+    self.sockets[service].send(data, flags=zmq.NOBLOCK)